WO2020248423A1 - 一种神经网络的量化参数确定方法及相关产品 - Google Patents

一种神经网络的量化参数确定方法及相关产品 Download PDF

Info

Publication number
WO2020248423A1
WO2020248423A1 PCT/CN2019/106754 CN2019106754W WO2020248423A1 WO 2020248423 A1 WO2020248423 A1 WO 2020248423A1 CN 2019106754 W CN2019106754 W CN 2019106754W WO 2020248423 A1 WO2020248423 A1 WO 2020248423A1
Authority
WO
WIPO (PCT)
Prior art keywords
data
quantization
time point
bit width
value
Prior art date
Application number
PCT/CN2019/106754
Other languages
English (en)
French (fr)
Inventor
刘少礼
孟小甫
张曦珊
郭家明
Original Assignee
上海寒武纪信息科技有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 上海寒武纪信息科技有限公司 filed Critical 上海寒武纪信息科技有限公司
Priority to EP19786896.1A priority Critical patent/EP3770823A4/en
Priority to US16/622,541 priority patent/US11675676B2/en
Priority to US16/720,113 priority patent/US11676029B2/en
Priority to US16/720,093 priority patent/US11676028B2/en
Publication of WO2020248423A1 publication Critical patent/WO2020248423A1/zh

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/06Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons
    • G06N3/063Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons using electronic means
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F11/00Error detection; Error correction; Monitoring
    • G06F11/07Responding to the occurrence of a fault, e.g. fault tolerance
    • G06F11/14Error detection or correction of the data by redundancy in operation
    • G06F11/1476Error detection or correction of the data by redundancy in operation in neural networks
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/08Learning methods
    • G06N3/084Backpropagation, e.g. using gradient descent
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/04Architecture, e.g. interconnection topology
    • G06N3/045Combinations of networks
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/04Architecture, e.g. interconnection topology
    • G06N3/047Probabilistic or stochastic networks
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/04Architecture, e.g. interconnection topology
    • G06N3/048Activation functions
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/04Architecture, e.g. interconnection topology
    • G06N3/0495Quantised networks; Sparse networks; Compressed networks
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/08Learning methods
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F2201/00Indexing scheme relating to error detection, to error correction, and to monitoring
    • G06F2201/81Threshold
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F2201/00Indexing scheme relating to error detection, to error correction, and to monitoring
    • G06F2201/865Monitoring of software

Definitions

  • the embodiments of the present disclosure relate to a method for determining quantitative parameters of a neural network and related products.
  • Neural network is a mathematical model or calculation model that imitates the structure and function of biological neural networks. Through the training of sample data, the neural network continuously corrects the network weights and thresholds to make the error function drop in the direction of negative gradient and approach the expected output. It is a widely used recognition and classification model, which is mostly used for function approximation, model recognition and classification, data compression and time series forecasting.
  • the data of neural network is usually 32Bit, and the data of the existing neural network occupies more bits. Although the accuracy is ensured, it requires higher storage space and processing bandwidth, which increases the cost.
  • the present disclosure proposes a method for determining quantitative parameters of a neural network and related products.
  • the present disclosure provides a method for determining quantization parameters of a neural network, and the method includes:
  • the statistical results and data bit widths of each type of data to be quantified are used to determine the corresponding quantization parameters; wherein the quantization parameters are used by the artificial intelligence processor to correspondingly quantify the data in the neural network operation process.
  • the present disclosure provides a quantization parameter determination device of a neural network, which includes a memory and a processor.
  • the memory stores a computer program that can run on the processor.
  • the processor executes the computer program, Implement the steps of the method described above.
  • the present disclosure provides a computer-readable storage medium on which a computer program is stored, wherein the computer program is executed by a processor to implement the steps of the foregoing method.
  • the present disclosure provides a quantization parameter determination device of a neural network, the device includes:
  • a statistical result obtaining unit configured to obtain a statistical result of each type of data to be quantified; wherein the data to be quantified includes at least one of the neuron, weight, gradient, and bias of the neural network;
  • the quantization parameter determination unit is used to determine the corresponding quantization parameter by using the statistical results of each type of data to be quantized and the data bit width; wherein the quantization parameter is used for the artificial intelligence processor to correspondingly quantify the data in the neural network operation process.
  • the quantization parameter is used by the artificial intelligence processor to quantify the data in the neural network operation process and convert high-precision data into low-precision fixed-point numbers. It can reduce all the space size of data storage involved in the process of neural network operation. For example: converting float32 to fix8 can reduce model parameters by 4 times. As the data storage space becomes smaller, a smaller space is used for neural network deployment, so that the on-chip memory on the artificial intelligence processor chip can hold more data, reducing the data access to the artificial intelligence processor chip, and improving computing performance.
  • Figure 1 is a schematic diagram of the neural network structure
  • Fig. 2 is a flowchart of a method for determining quantized parameters of a neural network proposed in this application;
  • Figure 3 is a schematic diagram showing a symmetrical fixed-point number
  • Fig. 4 is a schematic diagram showing a fixed-point number that introduces an offset
  • Figure 5a is one of the graphs of the variation range of the weight data of the neural network during the training process
  • Figure 5b is the second graph of the variation range of weight data of the neural network during training
  • Figure 6 is one of the flowcharts of the method for determining the target iteration interval
  • Figure 7 is the second flow chart of the method for determining the target iteration interval
  • Figure 8 is the third flow chart of the method for determining the target iteration interval
  • FIG. 9 is a block diagram of the hardware configuration of a device for determining quantized parameters of a neural network proposed by this application.
  • FIG. 10 is a schematic diagram of the application of the neural network quantitative parameter determination device proposed in this application to the artificial intelligence processor chip;
  • FIG. 11 is a functional block diagram of a device for determining quantitative parameters of a neural network proposed in this application.
  • Fig. 12 is a structural block diagram of a board according to an embodiment of the application.
  • the term “if” can be interpreted as “when” or “once” or “in response to determination” or “in response to detection” depending on the context.
  • the phrase “if determined” or “if detected [described condition or event]” can be interpreted as meaning “once determined” or “response to determination” or “once detected [described condition or event]” depending on the context ]” or “in response to detection of [condition or event described]”.
  • the floating-point number representation in the computer is divided into three fields, which are respectively coded:
  • a single sign bit s directly encodes the sign s.
  • Fixed-point number Consists of three parts: shared exponent (exponent), sign bit (sign), and mantissa (mantissa).
  • the shared exponent means that the exponent is shared within a set of real numbers that need to be quantified;
  • the sign bit marks the positive or negative of the fixed-point number.
  • the mantissa determines the number of significant digits of the fixed-point number, that is, the precision.
  • the numerical calculation method is:
  • any decimal number can be represented by the formula ⁇ j*10 i .
  • binary decimals can also be expressed in this way.
  • the left side of the decimal point is a positive power of 2
  • the right side of the decimal point is counted as a negative power of 2.
  • Overflow In fixed-point arithmetic, the number representation has a certain range. During the operation. If the size of the number exceeds the range that the fixed-point number can represent, it is called "overflow".
  • KL divergence also known as relative entropy (relative entropy), information divergence (information divergence), information gain (information gain).
  • KL divergence is a measure of the asymmetry of the difference between two probability distributions P and Q.
  • KL divergence is a measure of the number of extra bits required to encode the average of samples from P using Q-based coding.
  • P represents the true distribution of data
  • Q represents the theoretical distribution, model distribution, or approximate distribution of P.
  • Data bit width How many bits are used to represent the data.
  • Quantization The process of converting high-precision numbers expressed in the past with 32bit or 64bit into fixed-point numbers that take up less memory space. The process of converting high-precision numbers to fixed-point numbers will cause a certain loss in accuracy.
  • Neural network is a mathematical model that imitates the structure and function of a biological neural network.
  • the neural network is calculated by a large number of neuron connections. Therefore, a neural network is a computational model, which consists of a large number of nodes (or "neurons") connected to each other. Each node represents a specific output function, called an activation function. Each connection between two neurons represents a weighted value of the signal passing through the connection, called a weight, which is equivalent to the memory of a neural network.
  • the output of the neural network varies according to the connection method between neurons, as well as the weight and activation function.
  • the neuron is the basic unit of the neural network.
  • Connection is the connection of a neuron to another layer or another neuron in the same layer, and the connection is accompanied by the weight associated with it.
  • the bias is an additional input to the neuron, it is always 1 and has its own connection weight. This ensures that the neuron will activate even if all inputs are empty (all 0).
  • the neural network In application, if a non-linear function is not applied to the neurons in the neural network, the neural network is just a linear function, then it is not more powerful than a single neuron. If the output result of a neural network is between 0 and 1, for example, in the case of cat-dog identification, the output close to 0 can be regarded as a cat, and the output close to 1 can be regarded as a dog.
  • an activation function is introduced into the neural network, such as the sigmoid activation function.
  • FIG. 1 it is a schematic diagram of the neural network structure.
  • the hidden layer shown in Figure 1 is five layers. Among them, the leftmost layer of the neural network is called the input layer, and the neurons in the input layer are called input neurons.
  • the input layer is the first layer in the neural network, accepting the required input signals (values) and passing them to the next layer. It generally does not operate on the input signal (value), and has no associated weights and biases.
  • the hidden layer contains neurons (nodes).
  • the first hidden layer has 4 neurons (nodes)
  • the second layer has 5 neurons
  • the third layer has 6 neurons
  • the 4th layer has 4 neurons
  • the 5th layer has 3 neurons.
  • the hidden layer passes the calculated value of the neuron to the output layer.
  • the neural network shown in Figure 1 completely connects each neuron in the five hidden layers, that is, each neuron in each hidden layer is connected to each neuron in the next layer. It should be noted that not every hidden layer of neural network is fully connected.
  • FIG. 1 The rightmost layer of the neural network is called the output layer, and the neurons in the output layer are called output neurons.
  • the output layer receives the output from the last hidden layer.
  • the output layer has 3 neurons and 3 output signals y1, y2, and y3.
  • a large amount of sample data (including input and output) is given in advance to train the initial neural network, and after the training is completed, the trained neural network is obtained.
  • the neural network can give a correct output for the future input of the real environment.
  • the loss function is a function that measures the performance of a neural network in performing a specific task.
  • the loss function can be obtained as follows: in the process of training a certain neural network, for each sample data, the output value is passed along the neural network, and then the difference between the output value and the expected value is squared, so The calculated loss function is the distance between the predicted value and the true value, and the purpose of training the neural network is to reduce this distance or the value of the loss function.
  • the loss function can be expressed as:
  • y represents the expected value
  • i is the index of each sample data in the sample data set.
  • the error value between. m is the number of sample data in the sample data set.
  • the loss function In order to calculate the loss function, it is necessary to traverse each sample image in the sample data set to obtain the actual result corresponding to each sample image Then calculate the loss function according to the above definition. If the loss function is relatively large, then the neural network has not been trained well and the weights need to be further adjusted.
  • the weights should be initialized randomly. Obviously, the initialized neural network will not provide a good result. In the training process, suppose you start with a bad neural network. Through training, you can get a network with high accuracy.
  • the training process of the neural network is divided into two stages.
  • the first stage is the forward processing of the signal, from the input layer to the hidden layer, and finally to the output layer.
  • the second stage is to back-propagate the gradient, from the output layer to the hidden layer, and finally to the input layer. According to the gradient, the weight and bias of each layer in the neural network are adjusted in turn.
  • the input value is input to the input layer of the neural network, and the output of the so-called predicted value is obtained from the output layer of the neural network.
  • the input value is provided to the input layer of the neural network, it does nothing.
  • the second hidden layer obtains the predicted intermediate result value from the first hidden layer and performs calculation and activation operations, and then passes the obtained intermediate predicted result value to the next hidden layer. Perform the same operation in the subsequent layers, and finally get the output value in the output layer of the neural network.
  • an output value called the predicted value is obtained.
  • the predicted value is compared with the actual output value to obtain the corresponding error value.
  • Back propagation uses the chain rule of differential calculus.
  • the derivative of the error value corresponding to the last layer of the neural network is first calculated. Call these derivatives gradients, and then use these gradients to calculate the gradient of the penultimate layer in the neural network. Repeat this process until the gradient corresponding to each weight in the neural network is obtained. Finally, each weight value in the neural network is subtracted from the corresponding gradient, and the weight value is updated once to achieve the purpose of reducing the error value.
  • fine-tuning is to load the trained neural network.
  • the fine-tuning process is the same as the training process. It is divided into two stages. The first stage is the forward processing of the signal, and the second stage is the back propagation gradient. The weights of the passed neural network are updated.
  • the difference between training and fine-tuning is that training is to randomly process the initialized neural network and train the neural network from scratch, while fine-tuning is not.
  • the weights in the neural network are updated once using the gradient. This is called an iteration (iteration ).
  • iteration In order to obtain a neural network whose accuracy meets expectations, a very large sample data set is required during the training process. In this case, it is impossible to input the sample data set into the computer at once. Therefore, in order to solve this problem, the sample data set needs to be divided into multiple blocks, and each block is passed to the computer. After each block of the data set is processed forward, the weight of the neural network is updated correspondingly.
  • the data of the neural network is expressed in high-precision data formats, such as floating-point numbers, so in the training or fine-tuning process, the data involved are all high-precision data formats, and then the trained neural network is quantified.
  • the quantification object Take the quantification object as the weight of the entire neural network, and the quantized weights are all 8bit fixed-point numbers. Since there are often millions of connections in a neural network, almost all the space is occupied by the weights of the connections of neurons. . Moreover, these weights are all different floating point numbers.
  • the weights of each layer tend to a certain interval of normal distribution, such as (-3.0,3.0).
  • the interval is linearly divided into 256 quantization intervals within the range of the maximum value and the minimum value, and each quantization interval is represented by an 8bit fixed-point number.
  • byte 0 means -3.0
  • byte 255 means 3.0.
  • byte 128 represents 0.
  • the floating-point arithmetic unit needs to consume more resources to process, so that the power consumption gap between fixed-point arithmetic and floating-point arithmetic is usually orders of magnitude.
  • the chip area and power consumption occupied by floating-point arithmetic are many times larger than fixed-point arithmetic.
  • floating-point arithmetic is usually preferred, mainly because in supervised learning neural networks, only floating-point arithmetic can record and capture the small amount of training. Increment. Then, how to greatly improve the computing power of the chip for training without increasing the chip area and power consumption of the artificial intelligence processor is a problem that needs to be solved urgently.
  • the fixed-point number represented by the low bit width is used for training. According to practical feedback, it is necessary to use a fixed-point number higher than 8bit to process the back propagation gradient, which makes the fixed-point number represented by the low bit width to realize the training process Very complicated. How to replace the fixed-point arithmetic unit with the floating-point arithmetic unit to achieve the speed of fixed-point operation, improve the peak computing power of the artificial intelligence processor chip and meet the accuracy of the floating-point operation required for the operation is the technical problem solved in this manual.
  • one characteristic of the neural network is its high tolerance to input noise. If you consider recognizing objects in photos, the neural network can ignore the main noise and focus on important similarities. This function means that the neural network can use low-precision calculations as a noise source, and can still produce accurate prediction results in a numerical format that contains less information. To do low-precision training or fine-tuning, it is necessary to find a universal data representation that can not only improve the overflow of data, but also better express the data near 0 within the target interval. Therefore, this data indicates the need for adaptability, which can be adjusted along with the training or fine-tuning process.
  • FIG. 2 a flow chart of a method for determining quantized parameters of a neural network proposed in this application.
  • the quantization parameters determined by the technical solution shown in FIG. 2 are used to represent the data to be quantized, so as to confirm the number of fixed points after quantization.
  • the quantified fixed-point number is used for training, fine-tuning or reasoning of the neural network.
  • the method includes:
  • Step 201) Obtain statistical results of each type of data to be quantified; wherein the data to be quantified includes at least one of the neuron, weight, gradient, and bias of the neural network.
  • each layer of the neural network includes four types of data, which are neurons, weights, gradients, and biases.
  • each layer of the neural network includes three types of data, namely neurons, weights and biases. These data are expressed in high-precision data format.
  • This manual uses floating-point numbers as high-precision data as an example. It should be clear that the floating-point number as an example is only a partial list of examples, not an exhaustive list, and those skilled in the art may produce other cases based on the technical solutions of the present application if they understand the essence of the technical solutions.
  • the high-precision data can be a fixed-point number with a large range and a low minimum precision and a high data bit width, which can be converted into a fixed-point number with a low data bit width using this technical solution.
  • the high-precision data can be a fixed-point number with a large range and a low minimum precision and a high data bit width, which can be converted into a fixed-point number with a low data bit width using this technical solution.
  • the data to be quantified includes at least one of the neurons, weights, gradients, and biases of the neural network.
  • the data to be quantified It includes at least one data of neurons, weights, and biases of the neural network.
  • the data to be quantified can be the ownership value of a certain layer in the neural network, or part of the weight value of a certain layer in the neural network.
  • the data to be quantized may also be the ownership value or part of the weight value in the convolutional layer in units of channels, and the channel is all or part of the channels of the convolutional layer. It needs to be emphasized that only the convolutional layer has the concept of channels, and in the convolutional layer, only the weight layer is quantized in the way of channels.
  • the data to be quantified is the neuron and the weight of the target layer in the neural network as an example, and the technical solution is described in detail.
  • the neurons and weights of each layer in the target layer are separately counted to obtain the maximum and minimum values of each type of data to be quantized, and the maximum absolute value of each type of data to be quantized can also be obtained.
  • the target layer as the layer that needs to be quantified in the neural network, can be one layer or multiple layers. With one layer as the unit, the absolute maximum value of each data to be quantized can be confirmed by the maximum and minimum values in each data to be quantized. It is also possible to first obtain the absolute value of each type of data to be quantized, and traverse the results after obtaining the absolute value to obtain the maximum value of the absolute value of each type of data to be quantized.
  • the reason why the absolute maximum value of each data to be quantized is obtained according to the maximum and minimum values of each data to be quantized is that during quantization, under normal circumstances, the target layer of each layer to be quantized The maximum and minimum values corresponding to the data are saved, and there is no need to consume more resources to find the absolute value of the quantized data, and the absolute maximum value can be obtained directly based on the maximum and minimum values corresponding to the saved data to be quantified.
  • Step 202) Determine the corresponding quantization parameter by using the statistical result and the data bit width of each type of data to be quantized; wherein the quantization parameter is used by the artificial intelligence processor to correspondingly quantify the data in the neural network operation process.
  • the quantization parameter can be divided into the following six situations.
  • the following formula (1) can be used to quantize the data to be quantized to obtain the quantized data I x :
  • s is the point position parameter
  • I x is the n-bit binary representation value of the data x after quantization
  • F x is the floating point value of the data x before quantization
  • round is the rounding operation performed by rounding. It should be noted that this is not only limited to the rounding operation, but other rounding operations can also be used, such as: rounding up, rounding down, and rounding to zero. Replace the round operation in formula (1).
  • n-bit fixed-point number can represent the maximum value of floating-point number A is 2 s (2 n-1 -1), then n-bit fixed-point number can represent the maximum value in the number field of the data to be quantized as 2 s (2 n -1 -1), n-bit fixed-point number can indicate that the minimum value in the number field of the data to be quantized is -2 s (2 n-1 -1). It can be seen from equation (1) that when the quantization parameter corresponding to the first case is used to quantize the data to be quantized, the quantization interval is 2 s and the quantization interval is denoted as C.
  • the quantized n-bit binary representation value I x of data x is dequantized to obtain dequantized data
  • the dequantized data The data format of is the same as the data format of the corresponding data F x before quantization, and both are floating point values.
  • the quantization parameter is the first scaling factor f 1 .
  • the following formula (4) can be used to quantize the data to be quantized to obtain the quantized data I x :
  • f 1 is the first scaling factor
  • I x is the n-bit binary representation value after data x is quantized
  • F x is the floating point value before data x is quantized
  • round is the rounding operation performed by rounding. It should be noted that this is not only limited to the rounding operation, but other rounding operations can also be used, such as: rounding up, rounding down, and rounding to zero. Replace the round operation in formula (4). It can be seen from formula (4) that when the quantization parameter corresponding to the second case is used to quantize the data to be quantized, the quantization interval is f 1 and the quantization interval is denoted as C.
  • the point position parameter s is a fixed known value and no longer changes.
  • Set 2 s T, and T is a fixed value.
  • a fixed point number of n bits can be used
  • the maximum value A of floating-point numbers is (2 n-1 -1) ⁇ T. In this case, the maximum value A depends on the data bit width n.
  • the n-bit fixed-point number can indicate that the maximum value in the number field of the data to be quantized is (2 n-1 -1) ⁇ f 1
  • the n-bit fixed-point number can indicate that the minimum value in the number field of the data to be quantized is -(2 n-1 -1) ⁇ f 1
  • 2 s ⁇ f 2 as a whole is used as the first scaling factor f 1 . At this time, it can be regarded as there is no independent point position parameter s.
  • f 2 is the second scaling factor.
  • the n-bit fixed-point number can indicate that the maximum value in the number field of the data to be quantized is (2 n-1 -1) ⁇ f 1
  • the n-bit fixed-point number can indicate that the minimum value in the number field of the data to be quantized is -(2 n-1 -1) ⁇ f 1 .
  • the quantized n-bit binary representation value I x of data x is dequantized to obtain dequantized data
  • the dequantized data The data format of is the same as the data format of the corresponding data F x before quantization, and both are floating point values.
  • the quantization parameter is the point position parameter s and the second scaling factor f 2 .
  • the following formula (6) can be used to quantize the data to be quantized to obtain the quantized data I x :
  • s is the point position parameter
  • f 2 is the second scaling factor
  • I x is the n-bit binary representation value after the quantization of the data x
  • F x is the floating point value before the quantization of the data x
  • round is the rounding operation. It should be noted that this is not only limited to the rounding operation, but other rounding operations can also be used, such as: rounding up, rounding down, and rounding to zero. Replace the round operation in formula (6).
  • the maximum value A in the number field of the data to be quantized that can be represented by a fixed-point number of n bits is 2 s (2 n-1 -1). It can be seen from formula (6) that when the quantization parameter corresponding to the third case is used to quantize the data to be quantized, the quantization interval is 2 s ⁇ f 2 , and the quantization interval is denoted as C.
  • the n-bit fixed-point number can indicate that the maximum value in the number field of the data to be quantized is (2 n-1 -1) ⁇ 2 s ⁇ f 2
  • the n-bit fixed-point number can indicate that the minimum value in the number field of the data to be quantized is -(2 n-1 -1) ⁇ 2 s ⁇ f 2 .
  • the quantized n-bit binary representation value I x of data x is inversely quantized to obtain inverse quantized data
  • the dequantized data The data format of is the same as the data format of the corresponding data F x before quantization, and both are floating point values.
  • a symmetrical fixed-point number represents a schematic diagram.
  • the number field of the data to be quantized shown in FIG. 3 is distributed with "0" as the symmetric center.
  • Z is the maximum absolute value of all floating-point numbers in the number domain of the data to be quantized.
  • A is the maximum value of a floating-point number that can be represented by an n-bit fixed-point number.
  • the conversion of a floating-point number A to a fixed-point number is 2 n-1 -1.
  • A needs to include Z.
  • the floating-point data in the process of neural network operations tends to have a normal distribution in a certain interval, but it does not necessarily satisfy the distribution with "0" as the symmetric center.
  • an offset is introduced in the quantization parameter, as shown in Figure 4.
  • the number field of the data to be quantized is not distributed with "0" as the symmetric center
  • Z min is the minimum value of all floating-point numbers in the number field of the data to be quantized
  • Z max is all floating points in the number field of the data to be quantized.
  • P is the center point between Z min and Z max .
  • the number field of the data to be quantized is shifted as a whole, so that the number field of the data to be quantized after translation is distributed with "0" as the symmetric center.
  • the maximum absolute value in the number field is Z. It can be seen from FIG. 4 that the offset is the horizontal distance from point "0" to point "P”, and this distance is called offset O. among them,
  • the quantization parameter includes the point position parameter and the offset.
  • the following formula (8) can be used to quantize the data to be quantized to obtain quantized data I x :
  • s is the point position parameter
  • O is the offset
  • I x is the n-bit binary representation value after the quantization of the data x
  • F x is the floating point value before the quantization of the data x
  • round is the rounding operation. It should be noted that this is not only limited to the rounding operation, but other rounding operations can also be used, such as: rounding up, rounding down, and rounding to zero. Replace the round operation in formula (8).
  • n-bit fixed-point number can represent the maximum value of floating-point number A is 2 s (2 n-1 -1), then n-bit fixed-point number can represent the maximum value in the number field of the data to be quantized as 2 s (2 n -1 -1)+O, n-bit fixed-point number can indicate that the minimum value in the number field of the data to be quantized is -2 s (2 n-1 -1)+O. It can be seen from equation (8) that when the quantization parameter corresponding to the fourth case is used to quantize the data to be quantized, the quantization interval is 2 s and the quantization interval is denoted as C.
  • the quantized n-bit binary representation value I x of data x is inversely quantized to obtain inverse quantized data
  • the dequantized data The data format of is the same as the data format of the corresponding data F x before quantization, and both are floating point values.
  • f 1 is the first scaling factor
  • O is the offset
  • I x is the n-bit binary representation value after data x is quantized
  • F x is the floating point value before data x is quantized
  • round is the rounding operation performed by rounding.
  • this is not only limited to the rounding operation, but other rounding operations can also be used, such as: rounding up, rounding down, and rounding to zero. Replace the round operation in formula (10).
  • the point position parameter s is a fixed known value and no longer changes.
  • Set 2 s T
  • T is a fixed value.
  • the maximum value A of floating-point numbers that can be represented by n-bit fixed-point numbers is (2 n-1 -1) ⁇ T.
  • the maximum value A depends on the data bit width n.
  • the n-bit fixed-point number can indicate that the maximum value in the number field of the data to be quantized is (2 n-1 -1) ⁇ f 1
  • the n-bit fixed-point number can indicate that the minimum value in the number field of the data to be quantized is -(2 n-1 -1) ⁇ f 1 .
  • the n-bit fixed-point number can indicate that the maximum value in the number field of the data to be quantized is (2 n-1 -1) ⁇ f 1 +O, and the n-bit fixed-point number can indicate that the minimum value in the number field of the data to be quantized is -(2 n -1 -1) ⁇ f 1 +O.
  • the quantized n-bit binary representation value I x of data x is dequantized to obtain dequantized data
  • the dequantized data The data format of is the same as the data format of the corresponding data F x before quantization, and both are floating point values.
  • the quantization parameter includes the point position parameter, the second scaling factor f 2 and the offset O.
  • the following formula (12) can be used to quantize the quantized data to obtain quantized data I x :
  • s is the point position parameter
  • the offset is O
  • f 2 is the second scaling factor
  • I x is the n-bit binary representation value after the quantization of the data x
  • F x is the floating point value before the quantization of the data x
  • round is the rounding operation. It should be noted that this is not only limited to the rounding operation, but other rounding operations can also be used, such as: rounding up, rounding down, and rounding to zero. Replace the round operation in formula (12).
  • the maximum value A in the number field of the data to be quantized that can be represented by a fixed-point number of n bits is 2 s (2 n-1 -1). It can be seen from equation (12) that when the quantization parameter corresponding to the sixth case is used to quantize the data to be quantized, the quantization interval is 2 s ⁇ f 2 , and the quantization interval is denoted as C.
  • Z can be accurately expressed without loss.
  • f 2 1
  • the n-bit fixed-point number can indicate that the maximum value in the number field of the data to be quantized is (2 n-1 -1) ⁇ 2 s ⁇ f 2 +O
  • the n-bit fixed-point number can indicate that the minimum value in the number field of the data to be quantized is- (2 n-1 -1) ⁇ 2 s ⁇ f 2 +O.
  • the quantized n-bit binary representation value I x of data x is dequantized to obtain dequantized data
  • the dequantized data The data format of is the same as the data format of the corresponding data F x before quantization, and both are floating point values.
  • the data bit width n can be manually set. In the range of different iteration times, call the corresponding data bit width n set in advance.
  • This adjustment method of artificially setting the data bit width in advance basically does not meet the requirements of practical applications.
  • the data bit width n is adjusted according to the quantization error diff bit .
  • the quantization error diff bit is compared with the threshold, and the comparison result is obtained.
  • the threshold value includes a first threshold value and a second threshold value, and the first threshold value is greater than the second threshold value
  • the comparison result has three cases, the first case is: the quantization error diff bit is greater than or equal to the first threshold value, in this case, The data bit width is increased.
  • the second case is: the quantization error diff bit is less than or equal to the second threshold. In this case, the data bit width is reduced.
  • the third case is that the quantization error diff bit is between the first threshold and the second threshold. In this case, the data bit width remains unchanged.
  • the first threshold and the second threshold may be empirical values or variable hyperparameters. The conventional hyperparameter optimization methods are suitable for the first threshold and the second threshold, and the hyperparameter optimization scheme will not be repeated here.
  • the data bit width can be adjusted according to a fixed number of bits, or according to the difference between the quantization error and the error threshold, the data bit width can be adjusted according to the variable adjustment step.
  • the actual needs of the neural network operation process adjust the data bit width longer or shorter.
  • the data bit width n of the current convolutional layer is 16, and the data bit width n is adjusted to 12 according to the quantization error diff bit . That is to say, in practical applications, the data bit width n is 12 instead of 16 to meet the accuracy requirements in the neural network operation process, so that the fixed-point operation speed can be greatly increased within the accuracy allowable range. Thereby improving the resource utilization rate of the artificial intelligence processor chip.
  • the quantization error diff bit is determined according to the quantized data and the corresponding data before quantization.
  • the quantization error determination methods there are three quantization error determination methods, all of which are applicable to this technical solution.
  • the first way Determine the quantization error according to formula (14) according to the quantization interval, the number of quantized data, and the corresponding data before quantization.
  • C is the corresponding quantization interval during quantization
  • m is the number of quantized data obtained after quantization
  • F i is the corresponding floating point value to be quantized
  • i is the subscript of the data in the data set to be quantized.
  • the second method Determine the quantization error diff bit according to formula (15) according to the quantized data and the corresponding inverse quantization data.
  • F i is the corresponding floating point value to be quantized
  • i is the subscript of the data in the data set to be quantized.
  • I is the dequantized data corresponding to the floating point value.
  • the third way Determine the quantization error diff bit according to formula (16) according to the quantized data and the corresponding inverse quantization data.
  • F i is the corresponding floating point value to be quantized
  • i is the subscript of the data in the data set to be quantized.
  • I is the dequantized data corresponding to the floating point value.
  • Fig. 5a is one of the graphs of the variation range of the weight data of the neural network during the training process.
  • Figure 5b is the second graph of the variation range of the weight data of the neural network during the training process.
  • the abscissa indicates the number of iterations, and the ordinate indicates the maximum value of the weight after taking the logarithm.
  • the weight data variation amplitude curve shown in FIG. 5a shows the variation of weight data corresponding to different iterations in the same epoch of any convolution layer of the neural network.
  • the conv0 layer corresponds to the weight data change amplitude curve A
  • the conv1 layer corresponds to the weight data change amplitude curve B
  • the conv2 layer corresponds to the weight data change amplitude curve C
  • the conv3 layer corresponds to the weight data change amplitude curve D
  • conv4 The layer corresponds to the weight data variation range curve e. It can be seen from Fig. 5a and Fig. 5b that in the same epoch, in the initial training stage, the weight change range of each iteration is relatively large. In the middle and late stages of training, the change of weights in each iteration will not be too large.
  • the weight data of the corresponding layer of each generation has similarity within a certain iteration interval.
  • the data bit width used in the quantization of the corresponding layer in the previous iteration can be used.
  • the corresponding layer quantization used in the previous iteration is used.
  • the data bit width quantizes the weight data of the corresponding layer of the current generation, or quantizes the weight data of the current layer based on the preset data bit width n of the current layer to obtain the quantized fixed point number.
  • the data bit width n or the current quantization of the corresponding layer in the previous iteration The preset data bit width n of the layer is adjusted, and the adjusted data bit width is applied to the quantization of the weight data of the corresponding layer in the current iteration.
  • the weight data between each layer of the neural network is independent of each other and does not have similarity. Because the weight data does not have similarity, the neuron data between each layer is also independent of each other and does not have similarity. Therefore, in the process of neural network training or fine-tuning, the data bit width of each layer in each iteration of the neural network is only applicable to the corresponding neural network layer.
  • the data bit widths corresponding to the neuron data and the gradient data are also the same, which will not be repeated here.
  • the weight data between each layer of the neural network is independent of each other and does not have similarity. Because the weight data does not have similarity, the neuron data between each layer is also independent of each other and does not have similarity. Therefore, in the neural network inference process, the data bit width of each layer of the neural network is applied to the corresponding layer.
  • the input neuron data each time in the inference process is likely to be different or dissimilar, and because the weight data between each layer of the neural network is independent of each other, then each of the hidden layers of the neural network The input neuron data of the layers are not similar.
  • the data bit width used by the input neuron data of the previous layer is not suitable for the input neuron data of the current layer.
  • the input neuron data of the current layer is quantified using the data bit width used in the quantization of the input neuron data of the previous layer, or based on the current layer
  • the preset data bit width n quantifies the input neuron data of the current layer to obtain the quantized fixed point number.
  • the bit width n or the preset data bit width n of the current layer is adjusted, and the adjusted data bit width is applied to the quantization of the input neuron data of the current layer.
  • the data bit width corresponding to the weight data is also the same, which is not repeated here.
  • the quantization parameter of each layer in each iteration of the neural network is applied to the corresponding data to be quantized in the corresponding layer.
  • the quantization parameters corresponding to the neuron data and the gradient data are also the same, which will not be repeated here.
  • the weight data between each layer of the neural network is independent of each other and does not have similarity. Because the weight data does not have similarity, the neuron data between each layer is also independent of each other and does not have similarity. Therefore, in the neural network inference process, the quantization parameters of each layer of the neural network are applied to the data to be quantified in the corresponding layer.
  • the current layer of the neural network is the convolutional layer.
  • the quantization parameter of the data to be quantized in the current convolutional layer is obtained according to the technical scheme shown in Figure 2. The quantization parameter can only be applied to the current The convolutional layer cannot be applied to other layers of the neural network, even if other layers are convolutional layers.
  • the extension strategy of the data bit width and quantization parameter is determined based on the similarity between the data. If the data is similar, the data bit width and the quantization parameter can be used. For similarity, the data bit width or quantization parameter needs to be adjusted.
  • the measure of similarity between data is usually measured by KL divergence, and it can also be measured by the following formula (17).
  • the above-described methods for confirming quantization errors, adjusting data bit width, data bit width and quantization parameter extension strategies are only examples, not exhaustive, such as: the above The method of confirming the quantization error, the method of adjusting the data bit width, the strategy of extending the data bit width and the quantization parameter are all suitable for the fine-tuning process of the neural network.
  • the above-mentioned KL divergence and the similarity measurement method of formula (17) are just examples, not exhaustive, such as: histogram matching method , Matrix decomposition method, image similarity calculation method based on feature points, proximity measurement standard method, etc.
  • the weight data of the corresponding layer of each iteration has similarity within a certain iteration interval.
  • the solution has better applicability in training or fine-tuning to meet the requirements of artificial intelligence processor chip resources to achieve reasonable applications.
  • a strategy is needed to determine the iteration interval so that within the iteration interval, the corresponding layer of each iteration
  • the data bit width n remains unchanged, beyond the iteration interval, the data bit width n will change, and there is no need to determine whether to adjust the data bit width n generation by generation.
  • the quantization parameter is the same, so as to improve the peak computing power of the artificial intelligence processor chip while meeting the accuracy of floating-point operations required for quantization.
  • the target iteration interval includes at least one weight update iteration, and the same data bit width is used in the quantization process within the same target iteration interval.
  • the step of determining the target iteration interval includes:
  • Step 601) At a predictive time point, determine the change trend value of the position parameter of the corresponding point of the data to be quantified during the weight iteration process; wherein the predictive time point is used to determine whether the data bit width needs to be adjusted.
  • the time point at which the adjustment is made, and the predicted time point corresponds to the time point when the weight update iteration is completed.
  • the change trend value of the point position parameter is based on the weight value corresponding to the current prediction time point.
  • the moving average of the point position parameter during the iteration process, and the value corresponding to the previous prediction time point The moving average of the point position parameters in the weight iteration process is determined, or according to the point position parameter in the weight iteration process corresponding to the current predictive time point, and the point position in the weight iteration process corresponding to the last predictive time point
  • the moving average of the parameters is determined.
  • Equation 18 is:
  • M is the moving average of the point position parameter s increasing with training iterations.
  • M (t) is the moving average of the point position parameter s corresponding to the t-th predictive time point increasing with the training iteration
  • M (t) is obtained according to formula (19 ) .
  • s (t) is the point position parameter s corresponding to the t-th predictive time point.
  • M (t-1) is the sliding average value of the point position parameter s corresponding to the t-1th predictive time point
  • is the hyperparameter.
  • diff update1 measures the change trend of the point position parameter s, because the change of the point position parameter s is also reflected in disguised form in the change of the maximum value Z max in the current data to be quantified. The larger the diff update1 is, it indicates that the value range changes drastically, and an update frequency with a shorter interval is required, that is, the target iteration interval is smaller.
  • Step 602 Determine the corresponding target iteration interval according to the change trend value of the point position parameter.
  • the target iteration interval is determined according to equation (20).
  • the same data bit width is used in the quantization process within the same target iteration interval, and the data bit width used in the quantization process within different target iteration intervals may be the same or different.
  • I is the target iteration interval.
  • diff update1 is the change trend value of the point position parameter.
  • ⁇ and ⁇ are empirical values and can also be variable hyperparameters. Conventional hyperparameter optimization methods are all suitable for ⁇ and ⁇ , and the optimization scheme of hyperparameters will not be repeated here.
  • the predictive time point includes the first predictive time point, and the first predictive time point is determined according to the target iteration interval. Specifically, at the t-th predictive time point in the training or fine-tuning process, the weight data of the corresponding layer of the current iteration is quantized using the data bit width used in the quantization of the corresponding layer of the previous iteration to obtain the quantized fixed value. For the number of points, the quantization error diff bit is determined according to the weight data before quantization and the corresponding weight data before quantization. The quantization error diff bit is compared with the first threshold and the second threshold respectively, and the comparison result is used to determine whether to adjust the data bit width used in the quantization of the corresponding layer in the previous iteration.
  • the t-th first predictive time point corresponds to the 100th iteration
  • the data bit width used in the 99th iteration is n 1 .
  • the quantization error diff bit is confirmed according to the data bit width n 1 and the quantization error diff bit is compared with the first threshold and the second threshold to obtain the comparison result. If it is confirmed according to the comparison result that the data bit width n 1 does not need to be changed, use equation (20) to confirm that the target iteration interval is 8 iterations.
  • the 100th iteration is used as the initial iteration within the current target iteration interval, then the 100th to the first The 107th iteration is regarded as the current target iteration interval.
  • the 101st to 108th iterations are regarded as the current target iteration interval.
  • each generation still uses the data bit width n 1 used in the previous target iteration interval.
  • the data bit width used in quantization between different target iteration intervals can be the same. If the 100th to 107th iterations are used as the current target iteration interval, the 108th iteration within the next target iteration interval is regarded as the t+1 first predictive time point.
  • the quantization error diff bit is confirmed according to the data bit width n 1 , and the quantization error diff bit is compared with the first threshold and the second threshold to obtain the comparison result. According to the comparison result, it is determined that the data bit width n 1 needs to be changed to n 2 , and formula (20) is used to confirm that the target iteration interval is 55 iterations.
  • the 108th iteration to the 163rd iteration or the 109th iteration to the 163rd iteration are used as the target iteration interval, and the data bit width n 2 is used for each generation when quantizing within the target iteration interval.
  • the data bit width used in quantization can be different between different target iteration intervals.
  • formula (18) is applicable to obtain the change trend value of the point position parameter. If the first prediction time point at the current moment is the starting iteration of the current target iteration interval, then in formula (18), M (t) is the point position parameter corresponding to the time point corresponding to the starting iteration of the current target iteration interval s is the moving average of the increase of training iterations, s (t) is the point position parameter s corresponding to the time point of the initial iteration of the current target iteration interval, and M (t-1) is the start of the previous target iteration interval The moving average of the point position parameter s corresponding to the corresponding time point of the iteration increases with the training iteration.
  • M (t) is the point position parameter corresponding to the time point corresponding to the last iteration of the current target iteration interval s is the moving average value that increases with the training iteration
  • s (t) is the point position parameter s corresponding to the last iteration of the current target iteration interval
  • M (t-1) is the last target iteration interval
  • the prediction time point may also include the second prediction time point.
  • the second predictive time point is determined according to the data variation range curve. Based on the data fluctuation range of the big data in the neural network training process, the data fluctuation range curve as shown in FIG. 5a is obtained.
  • the data variation range curve shown in Fig. 5a it can be seen from the data variation range curve shown in Fig. 5a that from the start of training to the T-th iteration, the data variation range is very large every time the weight is updated.
  • the current iteration first uses the data bit width n 1 of the previous iteration to quantize, and the quantization result obtained and the corresponding data before quantization determine the corresponding quantization error.
  • the quantization error is respectively the same as the first
  • the threshold and the second threshold are compared, and the data bit width n 1 is adjusted according to the comparison result to obtain the data bit width n 2 .
  • Use the data bit width n 2 to quantify the weight data to be quantized involved in the current iteration.
  • the target iteration interval according to formula (20), thereby determining the first predictive time point, at the first predictive time point, determine whether to adjust the data bit width and how to adjust, and determine the next target iteration interval according to formula (20) To get the next first prediction time point. Since the interval between the start of training and the T-th iteration, the weight data changes before and after each iteration are very large, so that the weight data of the corresponding layer of each iteration does not have similarity. In order to meet the accuracy problem, quantify When the data of each layer of the current iteration cannot continue to use the corresponding quantization parameter of the corresponding layer of the previous iteration, the data bit width can be adjusted for the first T iterations.
  • the quantization is used for each iteration of the previous T iterations.
  • the data bit width of is different, and the target iteration interval is 1 iteration.
  • the target iteration interval of the previous T iterations can be preset in advance according to the law revealed by the data variation curve graph shown in Figure 5a, that is: according to the data variation curve before The target iteration interval of T iterations is directly preset, and there is no need to confirm through formula (20) that the weight update corresponding to each iteration of the previous T iterations is completed as the second predictive time point. This makes the resources of the artificial intelligence processor chip more reasonable.
  • the data change amplitude curve shown in Figure 5a has little change from the Tth iteration. In the middle and late stages of training, the quantization parameters are not reconfirmed for generations. At the Tth iteration or the T+1th iteration, use the current time The iteration corresponds to the data before quantization and the data after quantization to determine the quantization error. According to the quantization error, it is determined whether the data bit width needs to be adjusted and how to adjust, and the target iteration interval is also determined according to formula (20).
  • the confirmed target iteration interval is 55 iterations, this requires the time point corresponding to 55 iterations after the Tth iteration or the T+1th iteration as the first predictive time point to determine whether to adjust the data bit width and How to adjust, and determine the next target iteration interval according to formula (20), so as to determine the next first predictive time point, until all algebraic operations in the same epoch are completed.
  • adaptive adjustments are made to the data bit width or quantization parameters, and finally the quantized data is used to obtain a neural network whose accuracy meets the expectations.
  • the value of T is determined to be 130 according to the graph of the variation range of weight data shown in Fig. 5a (this value does not correspond to Fig. 5a.
  • the 130th iteration in the training process is regarded as the second prediction time point, and the current first prediction time point is the 100th iteration in the training process.
  • the formula (20) Determine the target iteration interval as 35 iterations. In the target iteration interval, train to the 130th iteration and reach the second predictive time point.
  • Equation (20) determines the target iteration interval.
  • the target iteration interval determined in this case is 42 iterations.
  • the 135th iteration corresponding to the first predictive time point determined when the target iteration interval is 35 iterations is within the target iteration interval of 42 iterations.
  • the second prediction time point is preset in advance according to the data variation curve.
  • the data bit width is directly adjusted according to the quantization error, and the adjusted data bit width is used to quantify the data to be quantized involved in the current iteration.
  • the target iteration interval is obtained according to formula (20) to determine the corresponding first predictive time point, and at each first predictive time point, it is determined whether to adjust the data bit width and how to adjust. In this way, while satisfying the accuracy of floating-point operations required by neural network operations, the resources of the artificial intelligence processor chip are reasonably used, which greatly improves the efficiency of quantization.
  • the step of determining the target iteration interval includes:
  • Step 701) Determine the change trend value of the position parameter of the corresponding point of the data to be quantified and the change trend value of the data bit width during the weight iteration process at the predicted time point; wherein, the predicted time point is used to judge Whether the data bit width needs to be adjusted at the time point, the pre-judged time point corresponds to the time point when the weight update iteration is completed.
  • the change trend value of the data bit width is determined by using the corresponding quantization error.
  • is the hyperparameter
  • diff bit is the quantization error
  • diff update2 is the change trend value of the data bit width.
  • diff update2 measures the changing trend of the data bit width n used in quantization. The larger the diff update2 , the more likely it is to update the fixed-point bit width, and a shorter update frequency is required.
  • the change trend value of the point position parameter involved in FIG. 7 can still be obtained according to formula (18), and the M (t) in formula (18) can be obtained according to formula (19).
  • diff update1 measures the change trend of the point position parameter s, because the change of the point position parameter s is also reflected in disguised form in the change of the maximum value Z max in the current data to be quantified. The larger the diff update1 is, it indicates that the value range changes drastically, and an update frequency with a shorter interval is required, that is, the target iteration interval is smaller.
  • Step 702 Determine the corresponding target iteration interval according to the change trend value of the point position parameter and the change trend value of the data bit width.
  • the target iteration interval is determined according to formula (22).
  • the same data bit width is used in the quantization process within the same target iteration interval, and the data bit width used in the quantization process within different target iteration intervals may be the same or different.
  • I is the target iteration interval.
  • ⁇ and ⁇ are hyperparameters.
  • diff update1 is the change trend value of the point position parameter.
  • diff update2 is the change trend value of the data bit width.
  • ⁇ and ⁇ are empirical values, and can also be variable hyperparameters. Conventional hyperparameter optimization methods are all suitable for ⁇ and ⁇ , and the optimization scheme of hyperparameters will not be repeated here.
  • diff update1 is used to measure the change of the point position parameter s, but the change of the point position parameter s caused by the change of the data bit width n should be ignored. Because this has already reflected the change of data bit width n in diff update2 . If this neglected operation is not done in diff update1 , then the target iteration interval I determined according to formula (22) is inaccurate, resulting in too many first prediction time points, and it is easy to do it frequently during training or fine-tuning The operation of whether the data bit width n is updated and how to update, thus causing the artificial intelligence processor chip's resources to not be reasonably used.
  • diff update1 is determined according to M (t) . Assuming that the data bit width corresponding to the t-1th predictive time point is n 1 , the corresponding point position parameter is s 1 , and the moving average of the point position parameter increasing with the training iteration is m 1 . The data to be quantized is quantized by using the data bit width n 1 to obtain the quantized fixed-point number.
  • the data bit width used in the quantization of the t-th predictive time point is n 2 .
  • M (t) one of the following two optimization methods can be selected.
  • the first method If the data bit width increases by
  • M (t) which is the point corresponding to the t-th predictive time point.
  • the data bit width n and the point position parameter s have a great influence on the quantization accuracy, and the second scaling factor f 2 and the offset O in the quantization parameter have little effect on the quantization accuracy.
  • the first scaling factor f 1 as mentioned above, if it belongs to the second case, consider 2 s ⁇ f 2 as a whole as the first scaling factor f 1 , because the point position parameter s has a great influence on the quantization accuracy. If large, the first scaling factor f 1 in this case has a great influence on quantization.
  • Step 801) Determine the change trend value of the position parameter of the corresponding point of the data to be quantified involved in the weight iteration process at the predictive time point; wherein the predictive time point is used to determine whether the quantization parameter needs to be The adjusted time point, the predicted time point corresponds to the time point when the weight update iteration is completed.
  • Step 802 Determine the corresponding target iteration interval according to the change trend value of the point position parameter.
  • the quantization parameter is preferably a point position parameter.
  • the technical solution is used to determine the quantization parameter, adjust the data bit width or quantization parameter according to the quantization error, and determine the target iteration interval for adjusting the data bit width or quantization parameter, so as to achieve the appropriate time point in the neural network operation process.
  • the data bit width or quantization parameters are adjusted so that the appropriate quantization parameters are used at the appropriate iteration time point to realize the artificial intelligence processor chip to perform the neural network operation to reach the fixed-point operation speed, and to improve the peak computing power of the artificial intelligence processor chip At the same time, it satisfies the accuracy of floating-point operations required for operations.
  • steps in the flowcharts of FIGS. 2, 6, 7, and 8 are displayed in sequence as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless specifically stated in this article, the execution of these steps is not strictly limited in order, and these steps can be executed in other orders. Moreover, at least some of the steps in Figure 2, Figure 6, Figure 7, and Figure 8 may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be performed at different times. Execution, the order of execution of these sub-steps or stages is not necessarily performed sequentially, but may be performed alternately or alternately with other steps or at least part of the sub-steps or stages of other steps.
  • the device 10 for determining the quantization parameter of the neural network may include a processor 110 and a memory 120.
  • the quantization parameter determination device 10 of the neural network in FIG. 9 only the constituent elements related to this embodiment are shown. Therefore, it is obvious to those of ordinary skill in the art that the quantization parameter determination device 10 of the neural network may also include common constituent elements different from those shown in FIG. 10. For example: fixed-point arithmetic.
  • the device 10 for determining the quantization parameter of a neural network can correspond to a computing device with various processing functions, for example, for generating a neural network, training or learning a neural network, quantizing a floating-point neural network into a fixed-point neural network, or retraining Function of neural network.
  • the quantization parameter determination apparatus 10 of a neural network may be implemented as various types of equipment, such as a personal computer (PC), server equipment, mobile equipment, and so on.
  • the processor 110 controls all functions of the quantitative parameter determination device 10 of the neural network.
  • the processor 110 controls all functions of the quantization parameter determination device 10 of the neural network by executing a program stored in the memory 120 on the quantization parameter determination device 10 of the neural network.
  • the processor 110 may be implemented by a central processing unit (CPU), a graphics processing unit (GPU), an application processor (AP), an artificial intelligence processor chip (IPU), etc., provided in the apparatus 10 for determining a quantitative parameter of a neural network.
  • CPU central processing unit
  • GPU graphics processing unit
  • AP application processor
  • IPU artificial intelligence processor chip
  • the memory 120 is hardware for storing various data processed in the quantization parameter determination device 10 of the neural network.
  • the memory 120 may store the processed data and the data to be processed in the device 10 for determining the quantitative parameter of the neural network.
  • the memory 120 can store the data set involved in the neural network operation process that the processor 110 has processed or to be processed, for example, the data of the untrained initial neural network, the intermediate data of the neural network generated during the training process, and all training is completed.
  • the memory 120 may store applications, driver programs, etc. to be driven by the quantization parameter determination device 10 of the neural network.
  • the memory 120 may store various programs related to the training algorithm, quantization algorithm, etc.
  • the memory 120 may be a DRAM, but the present disclosure is not limited thereto.
  • the memory 120 may include at least one of a volatile memory or a non-volatile memory.
  • Non-volatile memory can include read only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), flash memory, phase change RAM (PRAM), magnetic RAM (MRAM), resistance RAM (RRAM), ferroelectric RAM (FRAM), etc.
  • Volatile memory may include dynamic RAM (DRAM), static RAM (SRAM), synchronous DRAM (SDRAM), PRAM, MRAM, RRAM, ferroelectric RAM (FeRAM), etc.
  • the memory 120 may include a hard disk drive (HDD), a solid state drive (SSD), a high-density flash memory (CF), a secure digital (SD) card, a micro secure digital (Micro-SD) card, a mini secure digital (Mini -At least one of SD) card, xD card, caches or memory stick.
  • the processor 110 may generate a trained neural network by repeatedly training (learning) a given initial neural network.
  • the parameters of the initial neural network are a high-precision data representation format, such as a data representation format with 32-bit floating point precision.
  • the parameters may include various types of data input/output to/from the neural network, for example: input/output neurons, weights, biases, etc. of the neural network.
  • floating-point operations require relatively large amounts of operations and relatively frequent memory accesses.
  • most of the operations required for neural network processing are known as various convolution operations.
  • neural network high-precision data operations will make the resources of mobile devices underutilized.
  • the high-precision data involved in the neural network operation can be quantified and converted into low-precision fixed-point numbers.
  • the quantization parameter determination device 10 of the neural network performs conversion of the parameters of the trained neural network into fixed-point quantization with a specific number of bits, and the neural network
  • the network quantization parameter determination device 10 sends the corresponding quantization parameter to the device deploying the neural network, so that it is a fixed-point number arithmetic operation when the artificial intelligence processor chip performs training, fine-tuning and other operations.
  • the device deploying the neural network may be an autonomous vehicle, robot, smart phone, tablet device, augmented reality (AR) device, Internet of Things (IoT) device, etc. that perform voice recognition, image recognition, etc. by using the neural network, but the present disclosure does not Limited to this.
  • the processor 110 obtains data during the neural network operation from the memory 120.
  • the data includes at least one of neuron, weight, bias, and gradient.
  • the technical solution shown in FIG. 2 is used to determine the corresponding quantization parameter, and the quantization parameter is used to quantify the target data in the neural network operation process.
  • the processor 110 adjusts the data bit width n according to the quantization error diff bit , and the processor 110 can execute the program of the target iteration interval method shown in FIG. 6, FIG. 7 and FIG. 8 to determine the target iteration interval of the data bit width or The target iteration interval of the quantization parameter.
  • the method of this specification provides a neural network quantitative parameter determination device, and the specific functions implemented by the memory 120 and the processor 110 can be explained in comparison with the foregoing embodiments in this specification, and can achieve the foregoing embodiments. The technical effects of this will not be repeated here.
  • the processor 110 may be implemented in any suitable manner.
  • the processor 110 may take the form of, for example, a microprocessor or a processor, and a computer-readable medium storing computer-readable program codes (for example, software or firmware) executable by the (micro)processor, logic gates, switches, special purpose The form of integrated circuit (Application Specific Integrated Circuit, ASIC), programmable logic controller and embedded microcontroller, etc.
  • ASIC Application Specific Integrated Circuit
  • ASIC Application Specific Integrated Circuit
  • FIG. 10 a schematic diagram of the application of the neural network quantitative parameter determination device proposed in this application to the artificial intelligence processor chip.
  • the processor 110 performs a quantization operation to quantize the floating-point data involved in the neural network operation into fixed-point numbers, and artificial intelligence processing
  • the fixed-point arithmetic unit on the chip uses the fixed-point number obtained by quantization to perform training, fine-tuning or inference.
  • Artificial intelligence processor chips are dedicated hardware used to drive neural networks. Since the artificial intelligence processor chip is implemented with relatively low power or performance, the technical solution uses low-precision fixed-point numbers to implement neural network operations.
  • the technical solution can replace the floating-point arithmetic unit on the artificial intelligence processor chip with a fixed-point arithmetic unit, so that the artificial intelligence processor chip has lower power consumption. This is especially important for mobile devices.
  • the present technology opened a door leading to a large number of floating-point computation can not efficiently run code embedded systems, widely used in the world to make things possible.
  • the artificial intelligence processor chip may correspond to, for example, a neural processing unit (NPU), a tensor processing unit (TPU), a neural engine, etc., which are dedicated chips for driving neural networks, but the present disclosure is not limited to this.
  • NPU neural processing unit
  • TPU tensor processing unit
  • a neural engine etc., which are dedicated chips for driving neural networks, but the present disclosure is not limited to this.
  • the artificial intelligence processor chip can be implemented in a separate device that is independent of the quantitative parameter determination device 10 of the neural network, and the quantitative parameter determination device 10 of the neural network can also be used as a part of the functional module of the artificial intelligence processor chip.
  • the present disclosure is not limited to this.
  • the operating system of a general-purpose processor (such as CPU) generates instructions based on this technical solution, and sends the generated instructions to an artificial intelligence processor chip (such as GPU), and the artificial intelligence processor chip executes the instruction operations Realize the determination of the quantization parameters of the neural network and the quantification process.
  • the general-purpose processor directly determines the corresponding quantization parameter based on the technical solution
  • the general-purpose processor directly quantizes the corresponding target data according to the quantization parameter
  • the artificial intelligence processor chip uses the quantized data to perform fixed-point arithmetic operations.
  • general-purpose processors such as CPU
  • artificial intelligence processor chips such as GPU
  • the operating system of general-purpose processors (such as CPU) generates instructions based on this technical solution and copies target data at the same time Artificial intelligence processor chips (such as GPU) perform neural network operations, which can hide some time consumption. But the present disclosure is not limited to this.
  • the embodiment of the present application also provides a readable storage medium on which a computer program is stored, and when the computer program is executed, the method for determining the quantization parameter of the neural network described above is realized.
  • the quantization parameter is used by the artificial intelligence processor to quantify the data in the neural network operation process and convert high-precision data into low-precision data.
  • the precision fixed-point number can reduce all the space size of data storage involved in the process of neural network operation. For example: converting float32 to fix8 can reduce model parameters by 4 times. As the data storage space becomes smaller, a smaller space is used for neural network deployment, so that the on-chip memory on the artificial intelligence processor chip can hold more data, reducing the data access to the artificial intelligence processor chip, and improving computing performance.
  • FIG. 11 a functional block diagram of a device for determining quantitative parameters of a neural network proposed in this application.
  • the method includes:
  • the statistical result obtaining unit a is configured to obtain statistical results of each type of data to be quantified; wherein the data to be quantified includes at least one of neurons, weights, gradients, and biases of the neural network;
  • the quantization parameter determination unit b is used to determine the corresponding quantization parameter by using the statistical results of each type of data to be quantized and the data bit width; wherein the quantization parameter is used by the artificial intelligence processor to correspondingly quantify the data in the neural network operation process.
  • the quantization parameter determination device of the neural network further includes:
  • the first quantization unit is configured to quantize the data to be quantized by using the corresponding quantization parameter.
  • the quantization parameter determination device of the neural network further includes:
  • the second quantization unit is configured to quantify the target data by using the corresponding quantization parameter; wherein the characteristics of the target data and the characteristics of the data to be quantized have similarities.
  • the neural network operation process includes at least one operation of neural network training, neural network inference, and neural network fine-tuning.
  • the statistical result obtained by the statistical unit is the maximum value and the minimum value in each type of data to be quantified.
  • the statistical result obtained by the statistical unit is the maximum absolute value of each type of data to be quantified.
  • the statistical unit determines the maximum absolute value according to the maximum value and the minimum value in each type of data to be quantized.
  • the quantization parameter determination unit determines the quantization parameter according to the maximum value, the minimum value and the data bit width in each type of data to be quantized.
  • the quantization parameter determination unit determines the quantization parameter according to the maximum absolute value of each type of data to be quantized and the data bit width.
  • the quantization parameter determined by the quantization parameter determination unit is a point position parameter or a first scaling coefficient.
  • the quantization parameter determination unit determines the first zoom factor according to the point position parameter and the second zoom factor; wherein the point position parameter used in determining the first zoom factor is a known fixed value, or The result of multiplying the point position parameter and the corresponding second scaling factor as a whole is used as the first scaling factor for data quantization in the neural network operation process.
  • the quantization parameter determined by the quantization parameter determination unit includes a point position parameter and a second scaling coefficient.
  • the quantization parameter determination unit determines the second scaling factor according to the point position parameter, the statistical result, and the data bit width.
  • the quantization parameter determined by the quantization parameter determination unit further includes an offset.
  • the quantization parameter determination unit determines the offset according to the statistical result of each type of data to be quantized.
  • the data bit width used by the quantization parameter determining unit is a preset value.
  • the quantization parameter determination unit includes an adjustment module and a quantization error determination module; wherein,
  • the adjustment module is configured to adjust the data bit width according to the corresponding quantization error
  • the quantization error determination module is configured to determine the quantization error according to the quantized data and the corresponding data before quantization.
  • the adjustment module is specifically used for:
  • the quantization error is compared with a threshold, and the data bit width is adjusted according to the comparison result; wherein the threshold includes at least one of a first threshold and a second threshold.
  • the adjustment module includes a first adjustment sub-module, wherein the first adjustment sub-module is used for:
  • the data bit width is increased.
  • the adjustment module includes a second adjustment sub-module, wherein the second adjustment sub-module is used for:
  • the data bit width is reduced.
  • the adjustment module includes a third adjustment sub-module, wherein the third adjustment sub-module is used for:
  • the data bit width remains unchanged.
  • the quantization error determination module includes:
  • a quantization interval determining sub-module configured to determine a quantization interval according to the data bit width
  • the first quantization error determination submodule is configured to determine the quantization error according to the quantization interval, the number of quantized data, and the corresponding data before quantization.
  • the quantization error determination module includes:
  • the dequantization data determining sub-module is used to dequantize the quantized data to obtain the dequantized data; wherein the data format of the dequantized data is the same as the data format of the corresponding data before quantization;
  • the second quantization error determination sub-module is configured to determine the quantization error according to the quantized data and corresponding inverse quantization data.
  • the data before quantization used by the quantization error determining module is the data to be quantized.
  • the data before quantization used by the quantization error determination module is the data to be quantized involved in the weight update iteration process within the target iteration interval; wherein, the target iteration interval includes at least one weight The value is updated iterations, and the same data bit width is used in the quantization process within the same target iteration interval.
  • the quantization parameter determination device of the neural network further includes a first target iteration interval determination unit; wherein, the first target iteration interval determination unit includes:
  • the first change trend value determination module is used to determine the change trend value of the point position parameter of the to-be-quantified data involved in the weight update iteration process at the predicted time point; wherein the predicted time point is used to determine whether The time point at which the data bit width needs to be adjusted, and the predictive time point corresponds to the time point when the weight update iteration is completed;
  • the first target iteration interval module is configured to determine the corresponding target iteration interval according to the change trend value of the point position parameter.
  • the first target iteration interval determination unit includes:
  • the second change trend value determination module is used to determine the change trend value of the point position parameter of the data to be quantified and the change trend value of the data bit width involved in the weight update iteration process at the predicted time point; wherein, the predicted The judgment time point is a time point for judging whether the data bit width needs to be adjusted, and the prediction time point corresponds to the time point when the weight update iteration is completed;
  • the second target iteration interval module is configured to determine the corresponding target iteration interval according to the change trend value of the point position parameter and the change trend value of the data bit width.
  • the first target iteration interval determination unit further includes a first pre-determined time point determination unit; wherein,
  • the first prediction time point determining unit is configured to determine the first prediction time point according to the target iteration interval.
  • the first target iteration interval determination unit further includes a second pre-determined time point determination unit; wherein, the second pre-determined time point determination unit is configured to determine the second pre-determined time point according to the data change amplitude curve. Judging the time point; wherein the data variation range curve is obtained by statistically calculating the data variation range during the weight update iteration process.
  • the first change trend value determination module and the second change trend value determination module are based on the sliding average value of the point position parameter corresponding to the current prediction time point, and the value corresponding to the previous prediction time point.
  • the sliding average value of the point position parameter determines the change trend value of the point position parameter.
  • the first change trend value determination module and the second change trend value determination module are based on the value of the point position parameter corresponding to the current predictive time point and the point position parameter corresponding to the previous predictive time point.
  • the sliding average value determines the change trend value of the point position parameter.
  • the first change trend value determination module and the second change trend value determination module both include:
  • the point position parameter determination sub-module corresponding to the current predictive time point is used to determine the point position parameter corresponding to the current predictive time point according to the point position parameter corresponding to the last predictive time point and the adjustment value of the data bit width ;
  • the adjustment result determination sub-module is configured to adjust the sliding average value of the point position parameter corresponding to the last pre-judged time point according to the adjustment value of the data bit width to obtain an adjustment result;
  • the first sliding average determination sub-module is configured to determine the sliding average of the point position parameters corresponding to the current prediction time point according to the point position parameters corresponding to the current prediction time point and the adjustment result.
  • the first change trend value determination module and the second change trend value determination module both include:
  • the intermediate result determination sub-module is used to determine the moving average of the point position parameter corresponding to the current predictive time point according to the moving average of the point position parameter corresponding to the last predictive time point and the point position parameter corresponding to the last predictive time point Intermediate result of value;
  • the second sliding average determination sub-module is configured to determine the point position corresponding to the current prediction time point according to the intermediate result of the sliding average of the point position parameter corresponding to the current prediction time point and the adjustment value of the data bit width The moving average of the parameter.
  • the second change trend value determination module determines the change trend value of the data bit width according to the corresponding quantization error.
  • the first target iteration interval determining unit further includes:
  • a quantization error determination module configured to determine a corresponding quantization error; wherein the data before quantization corresponding to the quantization error is the data to be quantized involved in the weight update iteration process corresponding to the pre-determined time point;
  • the data bit width determination module is used to determine the data bit width used in the quantization process within the target iteration interval according to the corresponding quantization error.
  • the data bit width determination module is specifically configured to:
  • the quantization error is compared with the threshold, and the data bit width used in the quantization process in the last target iteration interval is adjusted according to the comparison result, and the adjustment result is used as the data bit width used in the quantization process in the current target iteration interval.
  • the data before quantization used by the quantization error determining module is the data to be quantized involved in the weight update iteration within the target iteration interval; wherein, the target iteration interval includes at least one weight The iteration is updated, and the same quantization parameter is used in the quantization process within the same target iteration interval.
  • the quantization parameter determination device of the neural network further includes a second target iteration interval determination unit; wherein, the second target iteration interval determination unit includes:
  • the third change trend value determination module is used to determine the change trend value of the point position parameter of the to-be-quantified data involved in the weight update iteration process at the predicted time point; wherein, the predicted time point is used to determine whether The time point at which the quantization parameter needs to be adjusted, and the predictive time point corresponds to the time point when the weight update iteration is completed;
  • the third target iteration interval module is configured to determine the corresponding target iteration interval according to the change trend value of the point position parameter.
  • the quantization parameter determination unit determines the point position parameter according to the statistical result and the data bit width.
  • the above device embodiments are only illustrative, and the device of the present disclosure may also be implemented in other ways.
  • the division of the units/modules in the foregoing embodiment is only a logical function division, and there may be other division methods in actual implementation.
  • multiple units, modules, or components may be combined or integrated into another system, or some features may be omitted or not implemented.
  • the units or modules described as separate components may or may not be physically separate.
  • a component described as a unit or module may be a physical unit or not a physical unit, that is, it may be located in one device, or may be distributed on multiple devices.
  • the solutions of the embodiments of the present disclosure can be implemented by selecting some or all of the units according to actual needs.
  • the functional units/modules in the various embodiments of the present disclosure may be integrated into one unit/module, or each unit/module may exist alone physically, or two or more units/modules may exist.
  • the modules are integrated together.
  • the above-mentioned integrated units/modules can be implemented in the form of hardware or software program modules.
  • the above device embodiments are only illustrative, and the device of the present disclosure may also be implemented in other ways.
  • the division of the units/modules in the foregoing embodiment is only a logical function division, and there may be other division methods in actual implementation.
  • multiple units, modules, or components may be combined or integrated into another system, or some features may be omitted or not implemented.
  • the units or modules described as separate components may or may not be physically separate.
  • a component described as a unit or module may be a physical unit or not a physical unit, that is, it may be located in one device, or may be distributed on multiple devices.
  • the solutions of the embodiments of the present disclosure can be implemented by selecting some or all of the units according to actual needs.
  • the functional units/modules in the various embodiments of the present disclosure may be integrated into one unit/module, or each unit/module may exist alone physically, or two or more units/modules may exist.
  • the modules are integrated together.
  • the above-mentioned integrated units/modules can be implemented in the form of hardware or software program modules.
  • the hardware may be a digital circuit, an analog circuit, and so on.
  • the physical realization of the hardware structure includes but is not limited to transistors, memristors and so on.
  • the artificial intelligence processor may be any appropriate hardware processor, such as: CPU, GPU, FPGA, DSP, ASIC, and so on.
  • the storage unit may be any suitable magnetic storage medium or magneto-optical storage medium, such as: resistive random access memory (RRAM), dynamic random access memory (DRAM), Static random access memory SRAM (Static Random-Access Memory), enhanced dynamic random access memory EDRAM (Enhanced Dynamic Random Access Memory), high-bandwidth memory HBM (High-Bandwidth Memory), hybrid storage cube HMC (Hybrid Memory Cube), etc. Wait.
  • RRAM resistive random access memory
  • DRAM dynamic random access memory
  • Static random access memory SRAM Static Random-Access Memory
  • enhanced dynamic random access memory EDRAM Enhanced Dynamic Random Access Memory
  • high-bandwidth memory HBM High-Bandwidth Memory
  • hybrid storage cube HMC Hybrid Memory Cube
  • the integrated unit/module is implemented in the form of a software program module and sold or used as an independent product, it can be stored in a computer readable memory.
  • the technical solution of the present disclosure essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, the computer software product is stored in a memory, A number of instructions are included to enable a computer device (which may be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present disclosure.
  • the aforementioned memory includes: U disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), mobile hard disk, magnetic disk or optical disk and other various media that can store program codes.
  • the present disclosure also discloses an artificial intelligence chip, which includes the above-mentioned neural network quantitative parameter determination device.
  • the present disclosure also discloses a board card, which includes a storage device, an interface device, a control device, and the aforementioned artificial intelligence chip; wherein, the artificial intelligence chip and the storage device, the control device, and The interface devices are respectively connected; the storage device is used to store data; the interface device is used to realize data transmission between the artificial intelligence chip and external equipment; the control device is used to The state of the smart chip is monitored.
  • Fig. 12 shows a structural block diagram of a board card according to an embodiment of the present disclosure.
  • the board card may include other supporting components in addition to the chip 389 mentioned above.
  • the supporting components include but are not limited to: Interface device 391 and control device 392;
  • the storage device 390 is connected to the artificial intelligence chip through a bus for storing data.
  • the storage device may include multiple groups of storage units 393. Each group of the storage unit and the artificial intelligence chip are connected through a bus. It can be understood that each group of the storage unit may be DDR SDRAM (English: Double Data Rate SDRAM, double-rate synchronous dynamic random access memory).
  • the storage device may include 4 groups of the storage units. Each group of the storage unit may include a plurality of DDR4 particles (chips).
  • the artificial intelligence chip may include four 72-bit DDR4 controllers. In the 72-bit DDR4 controller, 64 bits are used for data transmission and 8 bits are used for ECC verification. It can be understood that when DDR4-3200 particles are used in each group of the storage units, the theoretical bandwidth of data transmission can reach 25600MB/s.
  • each group of the storage unit includes a plurality of double-rate synchronous dynamic random memories arranged in parallel.
  • DDR can transmit data twice in one clock cycle.
  • a controller for controlling the DDR is provided in the chip for controlling the data transmission and data storage of each storage unit.
  • the interface device is electrically connected with the artificial intelligence chip.
  • the interface device is used to implement data transmission between the artificial intelligence chip and an external device (such as a server or a computer).
  • the interface device may be a standard PCIE interface.
  • the data to be processed is transferred from the server to the chip through a standard PCIE interface to realize data transfer.
  • the interface device may also be other interfaces. This disclosure does not limit the specific manifestations of the other interfaces mentioned above, as long as the interface unit can realize the switching function.
  • the calculation result of the artificial intelligence chip is still transmitted by the interface device back to an external device (such as a server).
  • the control device is electrically connected with the artificial intelligence chip.
  • the control device is used to monitor the state of the artificial intelligence chip.
  • the artificial intelligence chip and the control device may be electrically connected through an SPI interface.
  • the control device may include a single-chip microcomputer (Micro Controller Unit, MCU).
  • MCU Micro Controller Unit
  • the artificial intelligence chip may include multiple processing chips, multiple processing cores, or multiple processing circuits, and can drive multiple loads. Therefore, the artificial intelligence chip can be in different working states such as multiple load and light load.
  • the control device can realize the regulation of the working states of multiple processing chips, multiple processing and or multiple processing circuits in the artificial intelligence chip.
  • an electronic device which includes the aforementioned artificial intelligence chip.
  • Electronic equipment includes data processing devices, robots, computers, printers, scanners, tablets, smart terminals, mobile phones, driving recorders, navigators, sensors, cameras, servers, cloud servers, cameras, cameras, projectors, watches, headsets , Mobile storage, wearable devices, vehicles, household appliances, and/or medical equipment.
  • the transportation means include airplanes, ships, and/or vehicles;
  • the household appliances include televisions, air conditioners, microwave ovens, refrigerators, rice cookers, humidifiers, washing machines, electric lights, gas stoves, and range hoods;
  • the medical equipment includes nuclear magnetic resonance, B-ultrasound and/or electrocardiograph.
  • a method for determining quantitative parameters of a neural network comprising:
  • the statistical results and data bit widths of each type of data to be quantified are used to determine the corresponding quantization parameters; wherein the quantization parameters are used by the artificial intelligence processor to correspondingly quantify the data in the neural network operation process.
  • the data to be quantized is quantized using the corresponding quantization parameter.
  • the target data is quantified by using the corresponding quantization parameter; wherein the characteristics of the target data and the characteristics of the data to be quantized have similarities.
  • the neural network operation process includes at least one operation of neural network training, neural network inference, and neural network fine-tuning.
  • the quantization parameter includes a point position parameter and a second scaling factor.
  • the quantization error is compared with a threshold, and the data bit width is adjusted according to the comparison result; wherein the threshold includes at least one of a first threshold and a second threshold.
  • the data bit width is increased.
  • the data bit width is reduced.
  • the step of adjusting the data bit width includes:
  • the data bit width remains unchanged.
  • the method for obtaining the quantization error includes:
  • the quantization error is determined according to the quantization interval, the number of quantized data, and the corresponding data before quantization.
  • the method of obtaining the quantization error includes:
  • the quantization error is determined according to the quantized data and the corresponding inverse quantization data.
  • the predictive time point determine the change trend value of the point position parameter of the data to be quantified involved in the weight update iteration process; wherein the predictive time point is used to determine whether the data bit width needs to be adjusted Time point, the predicted time point corresponds to the time point when the weight update iteration is completed;
  • the target iteration interval is determined according to the change trend value of the point position parameter.
  • the predictive time point determine the change trend value of the point position parameter of the data to be quantified and the change trend value of the data bit width involved in the weight update iteration process; wherein the predictive time point is used to determine whether it is necessary to The time point at which the data bit width is adjusted, and the predictive time point corresponds to the time point when the weight update iteration is completed;
  • the target iteration interval is determined according to the change trend value of the point position parameter and the change trend value of the data bit width.
  • the pre-determined time point further includes a second pre-determined time point; wherein, the second pre-determined time point is determined according to a data variation range curve; the data variation range curve It is obtained by statistically calculating the magnitude of data change during the weight update iteration process.
  • the change trend value of the point position parameter is based on the sliding average value of the point position parameter corresponding to the current prediction time point, and the value corresponding to the previous prediction time point
  • the moving average of the point position parameter is determined.
  • the change trend value of the point position parameter is based on the point position parameter corresponding to the current prediction time point and the point position parameter corresponding to the previous prediction time point.
  • the moving average is determined.
  • the step of determining the moving average of the point position parameter corresponding to the current pre-determined time point includes:
  • the sliding average value of the point position parameter corresponding to the current prediction time point is determined according to the intermediate result of the sliding average value of the point position parameters corresponding to the current prediction time point and the adjustment value of the data bit width.
  • the step of determining the data bit width used in the quantization process within the target iteration interval includes:
  • the data bit width used in the quantization process within the target iteration interval is determined.
  • the step of determining the data bit width used in the quantization process within the target iteration interval includes:
  • the quantization error is compared with the threshold, and the data bit width used in the quantization process in the last target iteration interval is adjusted according to the comparison result, and the adjustment result is used as the data bit width used in the quantization process in the current target iteration interval.
  • the predictive time point determine the change trend value of the point position parameter of the data to be quantified involved in the weight update iteration process; wherein the predictive time point is the time used to determine whether the quantization parameter needs to be adjusted Point, the predicted time point corresponds to the time point when the weight update iteration is completed;
  • the target iteration interval is determined according to the change trend value of the point position parameter.
  • a device for determining quantitative parameters of a neural network comprising a memory and a processor.
  • the memory stores a computer program that can be run on the processor.
  • the processor implements clause A1 to clause A39 when the computer program is executed. The steps of any one of the methods.
  • a device for determining quantitative parameters of a neural network comprising:
  • a statistical result obtaining unit configured to obtain a statistical result of each type of data to be quantified; wherein the data to be quantified includes at least one of the neuron, weight, gradient, and bias of the neural network;
  • the quantization parameter determination unit is used to determine the corresponding quantization parameter by using the statistical results of each type of data to be quantized and the data bit width; wherein the quantization parameter is used for the artificial intelligence processor to correspondingly quantify the data in the neural network operation process.
  • the first quantization unit is configured to quantize the data to be quantized by using the corresponding quantization parameter.
  • the second quantization unit is configured to quantify the target data by using the corresponding quantization parameter; wherein the characteristics of the target data and the characteristics of the data to be quantized have similarities.
  • the quantization parameter determination unit determines the first scaling factor according to a point position parameter and a second scaling factor; wherein the point position parameter used when determining the first scaling factor is a known fixed The value or the result of multiplying the point position parameter and the corresponding second scaling factor as a whole is used as the first scaling factor for data quantization in the neural network operation process.
  • the quantization parameter determination unit includes an adjustment module and a quantization error determination module;
  • the quantization error determination module is configured to determine the quantization error according to the quantized data and corresponding data before quantization
  • the adjustment module is configured to adjust the data bit width according to the corresponding quantization error.
  • the quantization error is compared with a threshold, and the data bit width is adjusted according to the comparison result; wherein the threshold includes at least one of a first threshold and a second threshold.
  • the data bit width is increased.
  • the data bit width is reduced.
  • the data bit width remains unchanged.
  • a quantization interval determining sub-module configured to determine a quantization interval according to the data bit width
  • the first quantization error determination submodule is configured to determine the quantization error according to the quantization interval, the number of quantized data, and the corresponding data before quantization.
  • the dequantization data determining sub-module is used to dequantize the quantized data to obtain the dequantized data; wherein the data format of the dequantized data is the same as the data format of the corresponding data before quantization;
  • the second quantization error determination sub-module is configured to determine the quantization error according to the quantized data and corresponding inverse quantization data.
  • the quantization parameter determination device of the neural network further includes a first target iteration interval determination unit; wherein the first target iteration interval determination unit includes:
  • the first change trend value determination module is used to determine the change trend value of the point position parameter of the to-be-quantified data involved in the weight update iteration process at the predicted time point; wherein the predicted time point is used to determine whether The time point at which the data bit width needs to be adjusted, and the predictive time point corresponds to the time point when the weight update iteration is completed;
  • the first target iteration interval module is configured to determine the corresponding target iteration interval according to the change trend value of the point position parameter.
  • the second change trend value determination module is used to determine the change trend value of the point position parameter of the data to be quantified and the change trend value of the data bit width involved in the weight update iteration process at the predicted time point; wherein, the predicted The judgment time point is a time point for judging whether the data bit width needs to be adjusted, and the prediction time point corresponds to the time point when the weight update iteration is completed;
  • the second target iteration interval module is configured to determine the corresponding target iteration interval according to the change trend value of the point position parameter and the change trend value of the data bit width.
  • the first prediction time point determining unit is configured to determine the first prediction time point according to the target iteration interval.
  • the first target iteration interval determination unit further includes a second pre-determined time point determination unit; wherein the second pre-determined time point determination unit is used to determine a data change amplitude curve Determine the second pre-judgment time point; wherein, the data variation range curve is obtained by statistics of the data variation range during the weight update iteration process.
  • the point position parameter determination sub-module corresponding to the current predictive time point is used to determine the point position parameter corresponding to the current predictive time point according to the point position parameter corresponding to the last predictive time point and the adjustment value of the data bit width ;
  • the adjustment result determination sub-module is configured to adjust the sliding average value of the point position parameter corresponding to the last pre-judged time point according to the adjustment value of the data bit width to obtain an adjustment result;
  • the first sliding average determination sub-module is configured to determine the sliding average of the point position parameters corresponding to the current prediction time point according to the point position parameters corresponding to the current prediction time point and the adjustment result.
  • the intermediate result determination sub-module is used to determine the moving average of the point position parameter corresponding to the current predictive time point according to the moving average of the point position parameter corresponding to the last predictive time point and the point position parameter corresponding to the last predictive time point Intermediate result of value;
  • the second sliding average determination sub-module is configured to determine the point position corresponding to the current prediction time point according to the intermediate result of the sliding average of the point position parameter corresponding to the current prediction time point and the adjustment value of the data bit width The moving average of the parameter.
  • a quantization error determination module configured to determine a corresponding quantization error; wherein the data before quantization corresponding to the quantization error is the data to be quantized involved in the weight update iteration process corresponding to the pre-determined time point;
  • the data bit width determination module is used to determine the data bit width used in the quantization process within the target iteration interval according to the corresponding quantization error.
  • the quantization error is compared with the threshold, and the data bit width used in the quantization process in the last target iteration interval is adjusted according to the comparison result, and the adjustment result is used as the data bit width used in the quantization process in the current target iteration interval.
  • the quantization parameter determination device of the neural network further includes a second target iteration interval determination unit; wherein the second target iteration interval determination unit includes:
  • the third change trend value determination module is used to determine the change trend value of the point position parameter of the to-be-quantified data involved in the weight update iteration process at the predicted time point; wherein, the predicted time point is used to determine whether The time point at which the quantization parameter needs to be adjusted, and the predictive time point corresponds to the time point when the weight update iteration is completed;
  • the third target iteration interval module is configured to determine the corresponding target iteration interval according to the change trend value of the point position parameter.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • General Engineering & Computer Science (AREA)
  • Computing Systems (AREA)
  • Data Mining & Analysis (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Computational Linguistics (AREA)
  • Molecular Biology (AREA)
  • General Health & Medical Sciences (AREA)
  • Neurology (AREA)
  • Quality & Reliability (AREA)
  • Probability & Statistics with Applications (AREA)
  • Compression, Expansion, Code Conversion, And Decoders (AREA)
  • Image Analysis (AREA)
  • Compression Or Coding Systems Of Tv Signals (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

一种神经网络的量化参数确定方法及相关产品,相关产品中的板卡包括:存储器件(390)、接口装置(391)和控制器件(392)以及人工智能芯片(389);其中,所述人工智能芯片与所述存储器件、所述控制器件以及所述接口装置分别连接;所述存储器件,用于存储数据;所述接口装置,用于实现所述人工智能芯片与外部设备之间的数据传输;所述控制器件,用于对所述人工智能芯片的状态进行监控。所述板卡可以用于执行人工智能运算。

Description

一种神经网络的量化参数确定方法及相关产品
相关申请:
本申请要求2019年6月27日提交,申请号为201910570125.0,发明名称为“一种神经网络的量化参数确定方法及相关产品”的优先权。
本申请要求2019年6月12日提交,申请号为201910505239.7,发明名称为“神经网络量化方法及装置以及相关产品”的优先权。
本申请要求2019年6月18日提交,申请号为201910528537.8,发明名称为“量化参数调整方法、装置及相关产品”的优先权。
本申请要求2019年6月14日提交,申请号201910515355.7,发明名称为“一种神经网络运算方法及装置以及相关产品”的优先权。
技术领域
本公开的实施例涉及一种神经网络的量化参数确定方法及相关产品。
背景技术
神经网络(neural network,NN)是一种模仿生物神经网络的结构和功能的数学模型或计算模型。神经网络通过样本数据的训练,不断修正网络权值和阈值使误差函数沿负梯度方向下降,逼近期望输出。它是一种应用较为广泛的识别分类模型,多用于函数逼近、模型识别分类、数据压缩和时间序列预测等。
在实际运用中,神经网络的数据常用32Bit,现有的神经网络的数据占用的比特位较多,虽然确保了精度,但是需要较高的存储空间以及处理带宽,提高了成本。
发明内容
为了解决上述所述的技术问题,本公开提出一种神经网络的量化参数确定方法及相关产品。
为实现上述目的,本公开提供一种神经网络的量化参数确定方法,所述方法包括:
获得每种待量化数据的统计结果;其中,所述待量化数据包括所述神经网络的神经元、 权值、梯度、偏置中的至少一种数据;
利用每种待量化数据的统计结果以及数据位宽确定对应量化参数;其中,所述量化参数用于人工智能处理器对神经网络运算过程中的数据进行对应量化。
为实现上述目的,本公开提供一种神经网络的量化参数确定装置,包括存储器及处理器,所述存储器上存储有可在处理器上运行的计算机程序,所述处理器执行所述计算机程序时实现上述所述方法的步骤。
为实现上述目的,本公开提供一种计算机可读存储介质,其上存储有计算机程序,其中,所述计算机程序被处理器执行时实现上述所述的方法的步骤。
为实现上述目的,本公开提供一种神经网络的量化参数确定设备,所述设备包括:
统计结果获取单元,用于获得每种待量化数据的统计结果;其中,所述待量化数据包括所述神经网络的神经元、权值、梯度、偏置中的至少一种数据;
量化参数确定单元,用于利用每种待量化数据的统计结果以及数据位宽确定对应量化参数;其中,所述量化参数用于人工智能处理器对神经网络运算过程中的数据进行对应量化。
在神经网络运算过程中,量化时利用本公开的技术方案确定量化参数,该量化参数用于人工智能处理器对神经网络运算过程中的数据进行量化,将高精度数据转换为低精度定点数,可以减少神经网络运算过程中涉及的数据存储所有的空间大小。例如:float32转化为fix8可以将模型参数减少4倍。由于数据存储空间变小,使得神经网络部署时使用更小的空间,使得人工智能处理器芯片上的片上内存可以容纳更多的数据,减少了人工智能处理器芯片访存数据,提高计算性能。
附图说明
为了更清楚地说明本公开实施例的技术方案,下面将对实施例的附图作简单地介绍,显而易见地,下面描述中的附图仅仅涉及本公开的一些实施例,而非对本公开的限制。
图1为神经网络结构示意图;
图2为本申请提出的一种神经网络的量化参数确定方法流程图;
图3为对称的定点数表示示意图;
图4为引入偏移量的定点数表示示意图;
图5a为训练过程中神经网络的权值数据变动幅度曲线图之一;
图5b为训练过程中神经网络的权值数据变动幅度曲线图之二;
图6为确定目标迭代间隔的方法流程图之一;
图7为确定目标迭代间隔的方法流程图之二;
图8为确定目标迭代间隔的方法流程图之三;
图9为本申请提出的一种神经网络的量化参数确定装置的硬件配置的框图;
图10为本申请提出的神经网络的量化参数确定装置应用于人工智能处理器芯片的应用示意图;
图11为本申请提出的一种神经网络的量化参数确定设备的功能框图;
图12为本申请实施例的板卡的结构框图。
具体实施方式
下面将结合本披露实施例中的附图,对本披露实施例中的技术方案进行清楚、完整地描述,显然,所描述的实施例是本披露一部分实施例,而不是全部的实施例。基于本披露中的实施例,本领域技术人员在没有做出创造性劳动前提下所获得的所有其他实施例,都属于本披露保护的范围。
应当理解,本披露的权利要求、说明书及附图中的术语“第一”、“第二”、“第三”和“第四”等是用于区别不同对象,而不是用于描述特定顺序。本披露的说明书和权利要求书中使用的术语“包括”和“包含”指示所描述特征、整体、步骤、操作、元素和/或组件的存在,但并不排除一个或多个其它特征、整体、步骤、操作、元素、组件和/或其集合的存在或添加。
还应当理解,在此本披露说明书中所使用的术语仅仅是出于描述特定实施例的目的,而并不意在限定本披露。如在本披露说明书和权利要求书中所使用的那样,除非上下文清楚地指明其它情况,否则单数形式的“一”、“一个”及“该”意在包括复数形式。还应当进一步理解,在本披露说明书和权利要求书中使用的术语“和/或”是指相关联列出的项中的一个或多个的任何组合以及所有可能组合,并且包括这些组合。
如在本说明书和权利要求书中所使用的那样,术语“如果”可以依据上下文被解释为 “当...时”或“一旦”或“响应于确定”或“响应于检测到”。类似地,短语“如果确定”或“如果检测到[所描述条件或事件]”可以依据上下文被解释为意指“一旦确定”或“响应于确定”或“一旦检测到[所描述条件或事件]”或“响应于检测到[所描述条件或事件]”。
技术术语定义:
浮点数:IEEE浮点标准用V=(-1)∧sign*mantissa*2∧E的形式表示一个数。其中,sign为符号位,0表示正数,1代表负数;E表示阶码,对浮点数进行加权,权值是2的E次幂(可能是负数次幂);mantissa表示尾数,mantissa是一个二进制小数,其范围是1~2-ε,或者是0-ε。浮点数表示在计算机中的表示分为三个字段,分别对这些字段进行编码:
(1)一个单独的符号位s直接编码符号s。
(2)k位的阶码字段编码阶码,exp=e(k-1)......e(1)e(0)。
(3)n位的小数字段mantissa,编码尾数。但编码结果依赖阶码阶段是否全为0。
定点数:由共享指数(exponent)、符号位(sign)、尾数(mantissa)三部分构成。其中,共享指数是说指数在需要量化的一个实数集合内共享;符号位标志了定点数的正负。尾数决定了定点数的有效数字位数,即精度。以8bit定点数类型为例,其数值计算方法为:
value=(-1) sign×(mantissa)×2 (exponent-127)
二进制小数:任意十进制数都可以用公式∑j*10 i表示。例如十进制数12.34,以公式1表示为:12.34=1*10 1+2*10 0+3*10 -1+4*10 -2,小数点的左边计为10的正次幂,小数点的右边计为10的负次幂。同理,二进制小数也可以用这种方式表达,小数点左边为2的正次幂,小数点的右边计为2的负次幂,十进制小数5.75可用二进制小数101.11表示,该二进制小数表示为5.75=1*2 2+0*2 1+1*2 0+1*2 -1+1*2 -2
溢出:在定点运算器中,数的表示有一定的范围。在运算过程中。若数的大小超出了定点数能表示的范围,称为“溢出”。
KL(Kullback–Leibler divergence)散度:又称为相对熵(relative entropy)、信息散度(information divergence)、信息增益(information gain)。KL散度是两个概率分布P和Q之间差别的非对称性的度量。KL散度是用来度量使用基于Q的编码来编码来自P的样本平均所需的额外的比特个数。典型情况下,P表示数据的真实分布,Q表示数据的理论分布、模型分布、或P的近似分布。
数据位宽:数据用多少个比特位来表示。
量化:将以往用32bit或者64bit表达的高精度数转换成占用较少内存空间的定点数的过程,高精度数转换为定点数的过程就会在精度上引起一定的损失。
下面结合附图,对本公开实施例提供的一种神经网络的量化参数确定方法及相关产品的具体实施方式进行详细说明。
神经网络(neural network,NN)是一种模仿生物神经网络的结构和功能的数学模型,神经网络由大量的神经元连接进行计算。因此,神经网络是一种计算模型,由大量的节点(或称“神经元”)相互的连接构成。每个节点代表一种特定的输出函数,称为激活函数(activation function)。每两个神经元之间的连接都代表一个通过该连接信号的加权值,称之为权值,这相当于神经网络的记忆。神经网络的输出则依神经元之间的连接方式以及权值和激活函数的不同而不同。在神经网络中,神经元是神经网络的基本单位。它获得一定数量的输入和一个偏置,当信号(值)到达时会乘以一个权值。连接是将一个神经元连接到另一层或同一层的另一个神经元,连接伴随着与之相关联的权值。另外,偏置是神经元的额外输入,它始终为1,并具有自己的连接权值。这确保即使所有的输入都为空(全部为0),神经元也会激活。
在应用中,如果不对神经网络中的神经元应用一个非线性函数,神经网络只是一个线性函数而已,那么它并不比单个神经元强大。如果让一个神经网络的输出结果在0到1之间,例如,在猫狗鉴别的例子中,可以把接近于0的输出视为猫,将接近于1的输出视为狗。为了完成这个目标,在神经网络中引入激活函数,比如:sigmoid激活函数。关于这个激活函数,只需要知道它的返回值是一个介于0到1的数字。因此,激活函数用于将非线性引入神经网络,它会将神经网络运算结果缩小到较小的范围内。实际上,激活函数怎样表达并不重要,重要的是通过一些权值将一个非线性函数参数化,可以通过改变这些权值来改变这个非线性函数。
如图1所示,为神经网络结构示意图。在图1所示的神经网络中,包括三层,分别为输入层、隐含层以及输出层,图1所示的隐含层为5层。其中,神经网络的最左边一层被称为输入层,输入层的神经元被称为输入神经元。输入层作为神经网络中的第一层,接受需要输入信号(值)并将它们传递到下一层。它一般不对输入信号(值)做操作,并且没有关联的权值和偏置。在图1所示的神经网络中,有4个输入信号x1,x2,x3,x4。
隐含层包含神经元(节点)。在图1所示的神经网络中,有5个隐含层。第一隐含层有4个神经元(节点),第2层有5个神经元,第3层有6个神经元,第4层有4个神经元,第5层有3个神经元。最后,隐含层将神经元的运算值传递给输出层。图1所示的神经网 络将5个隐含层中每个神经元之间进行完全连接,即每个隐含层的每一个神经元都与下一层的每一个神经元有连接。需要说明的是,并不是每个神经网络的隐含层是完全连接的。
图1神经网络的最右边一层被称为输出层,输出层的神经元被称为输出神经元。输出层接收来自最后一个隐含层的输出。在图1所示的神经网络中,输出层有3个神经元,有3个输出信号y1,y2,y3。
在实际应用中,预先给大量的样本数据(包含输入和输出)对初始神经网络进行训练,训练完成后,获得训练后的神经网络。该神经网络对于将来的真实环境的输入能给出一个正确的输出。
在开始讨论神经网络的训练之前,需要定义损失函数。损失函数是一个衡量神经网络在执行某个特定任务的表现函数。在有些实施例中,损失函数可以如此得到:在训练某神经网络过程中,对每一个样本数据,都沿着神经网络传递得到输出值,然后将这个输出值与期望值做差再求平方,这样计算出来的损失函数就是预测值与真实值之间的距离,而训练神经网络目的就是将这个距离或损失函数的取值减小。在某些实施例中,损失函数可以表示为:
Figure PCTCN2019106754-appb-000001
上式中,y代表期望值,
Figure PCTCN2019106754-appb-000002
指样本数据集合中每个样本数据通过神经网络得到的实际结果,i是样本数据集合中每个样本数据的索引。
Figure PCTCN2019106754-appb-000003
表示期望值y与实际结果
Figure PCTCN2019106754-appb-000004
之间的误差值。m为样本数据集合中样本数据的个数。还是以猫狗鉴别为例。有一个数据集,由猫和狗的图片组成,如果图片是狗,对应的标签是1,如果图片是猫,对应的标签是0。这个标签就是对应上述公式中的期望值y,在向神经网络传递每一张样本图片的时候,实际是想通过神经网络获得识别结果。为了计算损失函数,必须遍历样本数据集中的每一张样本图片,获得每一张样本图片对应的实际结果
Figure PCTCN2019106754-appb-000005
然后按照上面的定义计算损失函数。如果损失函数比较大,那么说明神经网络还没有训练好,需要对权值进一步调整。
在开始训练神经网络的时候,要对权值进行随机初始化。显然,初始化的神经网络并不会提供一个很好的结果。在训练的过程中,假设以一个很糟糕的神经网络开始,通过训练,可以得到一个具有高准确率的网络。
神经网络的训练过程分为两个阶段,第一阶段是信号的正向处理,从输入层经过隐含层,最后到达输出层。第二阶段是反向传播梯度,从输出层到隐含层,最后到输入层,根据梯度依次调节神经网络中每层的权值和偏置。
在正向处理的过程中,将输入值输入到神经网络的输入层,并从神经网络的输出层得到所谓的预测值的输出。当输入值提供给神经网络的输入层时,它没有进行任何操作。在隐含层中,第二个隐含层从第一个隐含层获取预测中间结果值并进行计算操作和激活操作,然后将得到的预测中间结果值传递给下一个隐含层。在后面的层中执行相同的操作,最后在神经网络的输出层得到输出值。
正向处理后,得到一个被称为预测值的输出值。为了计算误差,将预测值与实际输出值进行比较,获得对应的误差值。反向传播使用微分学的链式法则,在链条法则中,首先计算对应神经网络的最后一层权值的误差值的导数。称这些导数为梯度,然后使用这些梯度来计算神经网络中的倒数第二层的梯度。重复此过程,直到得到神经网络中每个权值对应的梯度。最后,将神经网络中每个权值减去对应的梯度,从而对权值进行一次更新,以达到减少误差值的目的。
对于神经网络来说,微调是载入训练过的神经网络,微调过程与训练过程相同,分为两个阶段,第一阶段是信号的正向处理,第二阶段是反向传播梯度,对训练过的神经网络的权值进行更新。训练与微调的不同之处在于,训练是随机对初始化的神经网络进行处理,从头开始训练神经网络,而微调不是。
在神经网络进行训练或微调过程中,神经网络每经过一次信号的正向处理以及对应一次误差的反向传播过程,神经网络中的权值利用梯度进行一次更新,此时称为一次迭代(iteration)。为了获得精度符合预期的神经网络,在训练过程中需要很庞大的样本数据集。在这种情况下,一次性将样本数据集输入计算机是不可能的。因此,为了解决这个问题,需要把样本数据集分成多个块,每块传递给计算机,每块数据集正向处理后,对应更新一次神经网络的权值。当一个完整的样本数据集通过了神经网络一次正向处理并且对应返回了一次权值更新,这个过程称为一个周期(epoch)。实际中,在神经网络中传递一次完整的数据集是不够的,需要将完整的数据集在同一神经网络中传递多次,即需要多个周期,最终获得精度符合预期的神经网络。
在神经网络进行训练或微调过程中,一般希望速度越快越好,准确率越高越好。神经网络的数据通过高精度数据格式表示,比如:浮点数,所以在训练或微调过程中,涉及的数据均为高精度数据格式,然后再将训练好的神经网络进行量化。以量化对象是整个神经网络的权值、且量化后的权值均为8bit定点数为例,由于一个神经网络中常常有数百万连接,几乎所有空间都被神经元连接的权值所占据。况且这些权值都是不同的浮点数。每层权值都趋向于某个确定区间的正态分布,例如(-3.0,3.0)。将神经网络中每层的权值对应的最大 值和最小值保存下来,将每个浮点数值采用8bit定点数表示。其中,在最大值、最小值范围内区间线性划分256个量化间隔,每个量化间隔用一个8bit定点数表示。例如:在(-3.0,3.0)区间内,字节0表示-3.0,字节255表示3.0。以此类推,字节128表示0。
对于高精度数据格式表示的数据,以浮点数为例,根据计算机体系结构可知,基于浮点数的运算表示法则、定点数的运算表示法则,对于同样长度的定点运算和浮点运算,浮点运算计算模式更为复杂,需要更多的逻辑器件来构成浮点运算器。这样从体积上来说,浮点运算器的体积比定点运算器的体积要大。并且,浮点运算器需要消耗更多的资源去处理,使得定点运算和浮点运算二者之间的功耗差距通常是数量级的。简言之,浮点运算器占用的芯片面积和功耗相比于定点运算器都要大很多倍。
但是,浮点运算又有其不可取代性。首先,定点运算虽然直观,但是固定的小数点位置决定了固定位数的整数部分和小数部分,不利于同时表达特别大的数或者特别小的数,可能出现溢出的情况。
此外,具体使用人工智能处理器芯片做训练或微调时,通常更倾向于浮点运算器,主要是因为在有监督学习的神经网络中,只有浮点运算才能记录和捕捉到训练时很小的增量。那么,如何在不增加人工智能处理器芯片面积和功耗的前提下,如何大幅提升芯片做训练的运算能力是目前急需解决的问题。
对于本领域的技术人员来说,使用低位宽表示的定点数进行训练,经实践反馈可知需要使用高于8bit定点数来处理反向传播梯度,这使得使用低位宽表示的定点数实现训练的过程异常复杂。如何将定点运算器代替浮点运算器,达到定点运算的快速度,提升人工智能处理器芯片的峰值算力的同时满足运算所需的浮点运算的精度是本说明书解决的技术问题。
基于上述技术问题的描述,神经网络的一个特性是对输入噪声容忍度很高。如果考虑识别照片中物体,神经网络能忽略主要噪声,把注意力放在重要的相似性。该功能意味着神经网络可以把低精度计算作为一种噪声源,在容纳较少信息的数值格式下仍然能产生准确的预测结果。要做好低精度训练或微调,就要找到一个具有适普性的数据表示,既能改善数据的溢出情况,又能更好的表达目标区间范围内0附近的数据。因此,这个数据表示需要自适应性,能随着训练或微调的过程进行调整。
基于上述描述,如图2所示,为本申请提出的一种神经网络的量化参数确定方法流程图。利用图2所示的技术方案确定的量化参数用于对待量化数据进行数据表示,从而确认量化后的定点数。所述量化后的定点数用于进行神经网络的训练、微调或推理。所述方法 包括:
步骤201):获得每种待量化数据的统计结果;其中,所述待量化数据包括所述神经网络的神经元、权值、梯度、偏置中的至少一种数据。
正如前文所述,在训练或微调神经网络的过程中,神经网络每层包括四种数据,分别为神经元、权值、梯度和偏置。在推理过程中,神经网络每层包括三种数据,分别为神经元、权值和偏置。这些数据均以高精度数据格式表示,本说明书以浮点数作为高精度数据为例。需要明确的是,以浮点数为例仅仅是例举的部分情况,而不是穷举,本领域技术人员在理解本技术方案的精髓的情况下,可能会在本申请技术方案的基础上产生其它的变形或者变换,比如:高精度数据可以是表示范围大且表示的最小精度小的高数据位宽的定点数,采用本技术方案均可转为低数据位宽的定点数。但只要其实现的功能以及达到的技术效果与本申请类似,那么均应当属于本申请的保护范围。
不管什么样的神经网络结构,在训练或微调神经网络的过程中,待量化数据包括神经网络的神经元、权值、梯度、偏置中的至少一种数据,在推理过程中,待量化数据包括神经网络的神经元、权值、偏置中的至少一种数据。以待量化数据是权值为例,待量化数据可以是神经网络中某一层的所有权值,也可以是神经网络中某一层的部分权值。如果该层是卷积层时,待量化数据也可以是该卷积层中以通道为单位的所有权值或者部分权值,该通道是该卷积层的所有通道或者部分通道。需要强调的是,只有卷积层有通道的概念,在卷积层中,仅有权值分层按照通道的方式进行量化。
下面以待量化数据是神经网络中目标层的神经元和权值这两种数据为例,并详细描述本技术方案。在本步骤中,对目标层中每层的神经元和权值分别进行统计,获得每种待量化数据的最大值、最小值,也可以获得每种待量化数据的绝对值最大值。其中,目标层作为神经网络中需要量化的层,可以是一层,也可以是多层。以一层为单位,每种待量化数据的绝对值最大值可以通过每种待量化数据中的最大值和最小值方式确认。也可以先将每种待量化数据求绝对值,遍历求绝对值后的结果,获得每种待量化数据的绝对值最大值。
在实际应用中,按照每种待量化数据中的最大值和最小值方式获取每种待量化数据的绝对值最大值的原因在于,量化时,常规情况下会将目标层中每层的待量化数据对应的最 大值和最小值保存下来,无需消耗更多的资源去对待量化数据求绝对值,直接基于保存的待量化数据对应的最大值和最小值来获取绝对值最大值即可。
步骤202):利用每种待量化数据的统计结果以及数据位宽确定对应量化参数;其中,所述量化参数用于人工智能处理器对神经网络运算过程中的数据进行对应量化。
在本步骤中,量化参数可以分以下六种情况。第一种情况:量化参数是点位置参数s。这种情况下,可以利用如下的公式(1)对待量化数据进行量化,得到量化数据I x
Figure PCTCN2019106754-appb-000006
其中,s为点位置参数,I x为数据x量化后的n位二进制表示值,F x为数据x量化前的浮点值,round为进行四舍五入的取整运算。需要说明的是,此处不仅仅局限于round这一种取整运算,也可以采用其他的取整运算方法,例如:采用向上取整、向下取整、向零取整等取整运算,替换公式(1)中的round取整运算。此时,用n位定点数可以表示浮点数的最大值A为2 s(2 n-1-1),那么n位定点数可以表示待量化数据的数域中最大值为2 s(2 n-1-1),n位定点数可以表示待量化数据的数域中最小值为-2 s(2 n-1-1)。由式(1)可知,采用第一种情况对应的量化参数对待量化数据进行量化时,量化间隔为2 s,量化间隔记为C。
设Z为待量化数据的数域中所有浮点数的绝对值最大值,则A需要包含Z,且Z要大于
Figure PCTCN2019106754-appb-000007
因此有如下公式(2)约束:
2 s(2 n-1-1)≥Z>2 s-1(2 n-1-1)   (2)
因此,
Figure PCTCN2019106754-appb-000008
得到
Figure PCTCN2019106754-appb-000009
Figure PCTCN2019106754-appb-000010
根据式(3)对数据x量化后的n位二进制表示值I x进行反量化,获得反量化数据
Figure PCTCN2019106754-appb-000011
其中,所述反量化数据
Figure PCTCN2019106754-appb-000012
的数据格式与对应的量化前的数据F x的数据格式相同,均为浮点值。
Figure PCTCN2019106754-appb-000013
第二种情况:量化参数是第一缩放系数f 1。这种情况下,可以利用如下的公式(4)对待量化数据进行量化,得到量化数据I x
Figure PCTCN2019106754-appb-000014
其中,f 1为第一缩放系数,I x为数据x量化后的n位二进制表示值,F x为数据x量化前的浮点值,round为进行四舍五入的取整运算。需要说明的是,此处不仅仅局限于round这一种取整运算,也可以采用其他的取整运算方法,例如:采用向上取整、向下取整、向零取整等取整运算,替换公式(4)中的round取整运算。由式(4)可知,采用第二种情况对应的量化参数对待量化数据进行量化时,量化间隔为f 1,量化间隔记为C。
对于第一缩放系数f 1来说,有一种情况,即:点位置参数s为固定已知值,不再发生变化,设2 s=T,T为固定值,那么,用n位定点数可以表示浮点数的最大值A为(2 n-1-1)×T。这种情况下,最大值A取决于数据位宽n。设Z为待量化数据的数域中所有数的绝对值最大值,则
Figure PCTCN2019106754-appb-000015
此时Z=(2 n-1-1)×f 1。n位定点数可以表示待量化数据的数域中最大值为(2 n-1-1)×f 1,n位定点数可以表示待量化数据的数域中最小值为-(2 n-1-1)×f 1。还有一种情况,在工程应用中,2 s×f 2作为一个整体当做第一缩放系数f 1。此时,就可以当做不存在独立的点位置参数s。其中,f 2为第二缩放系数。设Z为待量化数据的数域中所有数的绝对值最大值,则
Figure PCTCN2019106754-appb-000016
此时Z=(2 n-1-1)×f 1。n位定点数可以表示待量化数据的数域中最大值为(2 n-1-1)×f 1,n位定点数可以表示待量化数据的数域中最小值为-(2 n-1-1)×f 1
根据式(5)对数据x量化后的n位二进制表示值I x进行反量化,获得反量化数据
Figure PCTCN2019106754-appb-000017
其中,所述反量化数据
Figure PCTCN2019106754-appb-000018
的数据格式与对应的量化前的数据F x的数据格式相同,均为浮点值。
Figure PCTCN2019106754-appb-000019
第三种情况:量化参数是点位置参数s和第二缩放系数f 2。这种情况下,可以利用如下的公式(6)对待量化数据进行量化,得到量化数据I x
Figure PCTCN2019106754-appb-000020
其中,s为点位置参数,f 2为第二缩放系数,
Figure PCTCN2019106754-appb-000021
I x为数据x量化后的n位二进制表示值,F x为数据x量化前的浮点值,round为进行四舍五入的取整运算。需要说明的是,此处不仅仅局限于round这一种取整运算,也可以采用其他的取整运算方法,例如:采用向上取整、向下取整、向零取整等取整运算,替换公式(6)中的round取整运算。用n位定点数可以表示的待量化数据的数域中的最大值A为2 s(2 n-1-1)。由式(6)可知,采用第三种情况对应的量化参数对待量化数据进行量化时,量化间隔为2 s×f 2,量化间隔记为C。
设Z为待量化数据的数域中所有数的绝对值最大值,此时,根据公式(2)可得:
Figure PCTCN2019106754-appb-000022
Figure PCTCN2019106754-appb-000023
Figure PCTCN2019106754-appb-000024
时,根据公式(2),Z可以无损精确表示。当f 2=1时,公式(6)与公式(1),
Figure PCTCN2019106754-appb-000025
n位定点数可以表示待量化数据的数域中最大值为(2 n-1-1)×2 s×f 2,n位定点数可以表示待量化数据的数域中最小值为-(2 n-1-1)×2 s×f 2
根据式(7)对数据x量化后的n位二进制表示值I x进行反量化,获得反量化数据
Figure PCTCN2019106754-appb-000026
其中,所述反量化数据
Figure PCTCN2019106754-appb-000027
的数据格式与对应的量化前的数据F x的数据格式相同,均为浮点值。
Figure PCTCN2019106754-appb-000028
如图3所示,对称的定点数表示示意图。图3所示的待量化数据的数域是以“0”为对称中心分布。Z为待量化数据的数域中所有浮点数的绝对值最大值,在图3中,A为n位定 点数可以表示的浮点数的最大值,浮点数A转换为定点数是2 n-1-1。为了避免溢出,A需要包含Z。实际中,神经网络运算过程中的浮点数据趋向于某个确定区间的正态分布,但是并不一定满足以“0”为对称中心的分布,这时用定点数表示时,容易出现溢出情况。为了改善这一情况,量化参数中引入偏移量,如图4所示。在图4中,待量化数据的数域不是以“0”为对称中心分布,Z min是待量化数据的数域中所有浮点数的最小值,Z max是待量化数据的数域中所有浮点数的最大值。P为Z min~Z max之间的中心点,将待量化数据的数域整体偏移,使得平移后的待量化数据的数域以“0”为对称中心分布,平移后的待量化数据的数域中的绝对值最大值为Z。由图4可知,偏移量为“0”点到“P”点之间的水平距离,该距离称为偏移量O。其中,
Figure PCTCN2019106754-appb-000029
基于上述关于偏移量O的描述,出现第四种量化参数的情况。第四种情况:量化参数包括点位置参数和偏移量。这种情况下,可以利用如下的公式(8)对待量化数据进行量化,得到量化数据I x
Figure PCTCN2019106754-appb-000030
其中,s为点位置参数,O为偏移量,
Figure PCTCN2019106754-appb-000031
I x为数据x量化后的n位二进制表示值,F x为数据x量化前的浮点值,round为进行四舍五入的取整运算。需要说明的是,此处不仅仅局限于round这一种取整运算,也可以采用其他的取整运算方法,例如:采用向上取整、向下取整、向零取整等取整运算,替换公式(8)中的round取整运算。此时,用n位定点数可以表示浮点数的最大值A为2 s(2 n-1-1),那么n位定点数可以表示待量化数据的数域中最大值为2 s(2 n-1-1)+O,n位定点数可以表示待量化数据的数域中最小值为-2 s(2 n-1-1)+O。由式(8)可知,采用第四种情况对应的量化参数对待量化数据进行量化时,量化间隔为2 s,量化间隔记为C。
设Z为待量化数据的数域中所有浮点数的绝对值最大值,
Figure PCTCN2019106754-appb-000032
则A需要包 含Z,且Z要大于
Figure PCTCN2019106754-appb-000033
根据公式(2)获得
Figure PCTCN2019106754-appb-000034
进而得到
Figure PCTCN2019106754-appb-000035
根据式(9)对数据x量化后的n位二进制表示值I x进行反量化,获得反量化数据
Figure PCTCN2019106754-appb-000036
其中,所述反量化数据
Figure PCTCN2019106754-appb-000037
的数据格式与对应的量化前的数据F x的数据格式相同,均为浮点值。
Figure PCTCN2019106754-appb-000038
基于上述关于偏移量O的描述,出现第五种量化参数的情况。第五种情况:量化参数包括第一缩放系数f 1和偏移量O。这种情况下,可以利用如下的公式(10)对待量化数据进行量化,得到量化数据I x
Figure PCTCN2019106754-appb-000039
其中,f 1为第一缩放系数,O为偏移量,I x为数据x量化后的n位二进制表示值,F x为数据x量化前的浮点值,round为进行四舍五入的取整运算。需要说明的是,此处不仅仅局限于round这一种取整运算,也可以采用其他的取整运算方法,例如:采用向上取整、向下取整、向零取整等取整运算,替换公式(10)中的round取整运算。此时,有一种情况,即:点位置参数s为固定已知值,不再发生变化,设2 s=T,T为固定值。那么,用n位定点数可以表示浮点数的最大值A为(2 n-1-1)×T。这种情况下,最大值A取决于数据位宽n。设Z为待量化数据的数域中所有数的绝对值最大值,则
Figure PCTCN2019106754-appb-000040
此时Z=(2 n-1-1)×f 1。n位定点数可以表示待量化数据的数域中最大值为(2 n-1-1)×f 1,n位定点数可以表示待量化数据的数域中最小值为-(2 n-1-1)×f 1。还有一种情况,在工程应用中,2 s×f 2作为一个整体当作第一缩放系数f 1。此时,就可以当作不存在独立的点位置参数s。 其中,f 2为第二缩放系数。设Z为待量化数据的数域中所有数的绝对值最大值,则
Figure PCTCN2019106754-appb-000041
此时Z=(2 n-1-1)×f 1。n位定点数可以表示待量化数据的数域中最大值为(2 n-1-1)×f 1+O,n位定点数可以表示待量化数据的数域中最小值为-(2 n-1-1)×f 1+O。
由式(10)可知,采用第五种情况对应的量化参数对待量化数据进行量化时,量化间隔为f 1,量化间隔记为C。
根据式(11)对数据x量化后的n位二进制表示值I x进行反量化,获得反量化数据
Figure PCTCN2019106754-appb-000042
其中,所述反量化数据
Figure PCTCN2019106754-appb-000043
的数据格式与对应的量化前的数据F x的数据格式相同,均为浮点值。
Figure PCTCN2019106754-appb-000044
基于上述关于偏移量O的描述,出现第六种量化参数的情况。第六种情况:量化参数包括点位置参数、第二缩放系数f 2和偏移量O。这种情况下,可以利用如下的公式(12)对待量化数据进行量化,得到量化数据I x
Figure PCTCN2019106754-appb-000045
其中,s为点位置参数,偏移量O,f 2为第二缩放系数,
Figure PCTCN2019106754-appb-000046
I x为数据x量化后的n位二进制表示值,F x为数据x量化前的浮点值,round为进行四舍五入的取整运算。需要说明的是,此处不仅仅局限于round这一种取整运算,也可以采用其他的取整运算方法,例如:采用向上取整、向下取整、向零取整等取整运算,替换公式(12)中的round取整运算。用n位定点数可以表示的待量化数据的数域中的最大值A为2 s(2 n-1-1)。由式(12)可知,采用第六种情况对应的量化参数对待量化数据进行量化时,量化间隔为2 s×f 2,量化间隔记为C。
设Z为待量化数据的数域中所有数的绝对值最大值,此时,根据公式(2)可得:
Figure PCTCN2019106754-appb-000047
Figure PCTCN2019106754-appb-000048
Figure PCTCN2019106754-appb-000049
时,根据公式(2),Z可以无损精确表示。当f 2=1时,
Figure PCTCN2019106754-appb-000050
n位定点数可以表示待量化数据的数域中最大值为(2 n-1-1)×2 s×f 2+O,n位定点数可以表示待量化数据的数域中最小值为-(2 n-1-1)×2 s×f 2+O。
根据式(13)对数据x量化后的n位二进制表示值I x进行反量化,获得反量化数据
Figure PCTCN2019106754-appb-000051
其中,所述反量化数据
Figure PCTCN2019106754-appb-000052
的数据格式与对应的量化前的数据F x的数据格式相同,均为浮点值。
Figure PCTCN2019106754-appb-000053
上述详细描述了6种量化参数的确定过程,仅仅是实施例说明。量化参数的种类在不同的实施例中可以与以上的描述不同。由公式(1)~公式(13)可知,点位置参数和缩放系数均与数据位宽有关。不同的数据位宽,导致点位置参数和缩放系数不同,从而影响量化精度。在训练或微调过程中,在一定的迭代(iterations)的次数范围内,使用相同的数据位宽量化对神经网络运算的总体精度影响不大。超过一定的迭代次数,再使用同一数据位宽量化就无法满足训练或微调对精度的要求。这就需要随着训练或微调的过程对数据位宽n进行调整。简单地,可以人为设置数据位宽n。在不同的迭代次数范围内,调用提前设置的对应的数据位宽n。但是,前文已经提到,使用低位宽表示的定点数实现训练的过程异常复杂。这种人为提前设置数据位宽的调整方式基本上是不符合实际应用的需求。
在本技术方案中,根据量化误差diff bit对数据位宽n进行调整。进一步详说,将量化误差diff bit与阈值进行比较,并获取比较结果。其中,阈值包括第一阈值和第二阈值,且第一阈值大于第二阈值,比较结果有三种情况,第一种情况为:量化误差diff bit大于等于所述第一阈值,这种情况下,对所述数据位宽进行增加。第二种情况为:量化误差diff bit小于等于所述第二阈值,这种情况下,对所述数据位宽进行减少。第三种情况为:量化误差diff bit处于所述第一阈值和所述第二阈值之间,这种情况下,所述数据位宽保持不变。在实际应用 中,第一阈值和第二阈值可以为经验值,也可以为可变的超参数。常规的超参数的优化方法均适于第一阈值和第二阈值,这里不再赘述超参数的优化方案。
需要强调的是,可以将数据位宽按照固定的位数步长进行调整,也可以根据量化误差与误差阈值之间的差值的不同,按照可变的调整步长调整数据位宽,最终根据神经网络运算过程的实际需要,将数据位宽调整的更长或更短。比如:当前卷积层的数据位宽n为16,根据量化误差diff bit将数据位宽n调整为12。也就是说,在实际应用中,数据位宽n取值为12而不必取值为16即可满足神经网络运算过程中对精度的需求,这样在精度允许范围内可以大大提到定点运算速度,从而提升了人工智能处理器芯片的资源利用率。
对于量化误差diff bit来说,根据量化后的数据与对应的量化前的数据确定量化误差。在实际应用中,有三种量化误差确定方式,均可适用于本技术方案。第一种方式:根据量化间隔、量化后的数据的个数以及对应的量化前的数据按照公式(14)确定量化误差。
Figure PCTCN2019106754-appb-000054
其中,C为量化时对应的量化间隔,m为量化后获得的量化数据的个数,F i为待量化对应的浮点值,其中,i为待量化数据集合中数据的下标。
第二种方式:根据量化后的数据以及对应的反量化数据按照公式(15)确定量化误差diff bit
Figure PCTCN2019106754-appb-000055
其中,F i为待量化对应的浮点值,其中,i为待量化数据集合中数据的下标。
Figure PCTCN2019106754-appb-000056
为浮点值对应的反量化数据。
第三种方式:根据量化后的数据以及对应的反量化数据按照公式(16)确定量化误差diff bit
Figure PCTCN2019106754-appb-000057
其中,F i为待量化对应的浮点值,其中,i为待量化数据集合中数据的下标。
Figure PCTCN2019106754-appb-000058
为浮 点值对应的反量化数据。
需要强调的是,上述获取量化误差diff bit的方式仅仅是例举的部分情况,而不是穷举,本领域技术人员在理解本申请技术方案的精髓的情况下,可能会在本申请技术方案的基础上产生其它的变形或者变换,凡是支持根据量化后的数据与对应的量化前的数据确定量化误差的变形公式,但只要其实现的功能以及达到的技术效果与本申请类似,那么均应当属于本申请的保护范围。
对于数据位宽来说,图5a为训练过程中神经网络的权值数据变动幅度曲线图之一。图5b为训练过程中神经网络的权值数据变动幅度曲线图之二。在图5a和图5b中,横坐标表示是迭代数,纵坐标表示是权值取对数后的最大值。图5a所示的权值数据变动幅度曲线展示神经网络的任一卷积层同一周期(epoch)内在不同迭代对应的权值数据变动情况。在图5b中,conv0层对应权值数据变动幅度曲线A,conv1层对应权值数据变动幅度曲线B,conv2层对应权值数据变动幅度曲线C,conv3层对应权值数据变动幅度曲线D,conv4层对应权值数据变动幅度曲线e。由图5a和图5b可知,同一个周期(epoch)内,在训练初期,每次迭代权值变化幅度比较大。在训练中后期,每次迭代权值的变化幅度不会太大。此种情况下,在训练中后期,因为每次迭代前后权值数据变化幅度不大,使得每代的对应层的权值数据之间在一定的迭代间隔内具有相似性,在神经网络训练过程中每层涉及的数据量化时可以采用上一次迭代时对应层量化时使用的数据位宽。但是,在训练初期,由于每次迭代前后权值数据的变化幅度比较大,为了满足量化所需的浮点运算的精度,在训练初期的每一次迭代,利用上一次迭代对应层量化时采用的数据位宽对当前代的对应层的权值数据进行量化,或者基于当前层预设的数据位宽n对当前层的权值数据进行量化,获得量化后的定点数。根据量化后的权值数据和对应的量化前的权值数据,确定量化误差diff bit,根据量化误差diff bit与阈值的比较结果,对上一次迭代对应层量化时采用的数据位宽n或者当前层预设的数据位宽n进行调整,将调整后的数据位宽应用于当前次迭代的对应层的权值数据的量化。进一步地,在训练或微调过程中,神经网络的每层之间的权值数据相互独立,不具备相似性。因权值数据不具备相似性使得每层之间的神经元数据也相互独立,不具备相似性。因此,在神经网络训练或微调过程中,神经网络的每一次迭代内的每层的数据位宽只适用 用于对应的神经网络层。
上述以权值数据为例,在神经网络训练或微调过程中,神经元数据和梯度数据分别对应的数据位宽亦如此,此处不再赘述。
在神经网络推理过程中,神经网络的每层之间的权值数据相互独立,不具备相似性。因权值数据不具备相似性使得每层之间的神经元数据也相互独立,不具备相似性。因此,在神经网络推理过程中,神经网络的每层的数据位宽应用于对应层。在实际应用中,推理过程中每次的输入神经元数据很有可能不相同或者不相似,并且,由于神经网络的每层之间的权值数据相互独立,那么神经网络隐含层中的每层的输入神经元数据就不相似。量化时,上一层的输入神经元数据使用的数据位宽就不适于应用于当前层的输入神经元数据。基于此,为了满足量化所需的浮点运算的精度,在推理时,利用上一层的输入神经元数据量化时采用的数据位宽对当前层的输入神经元数据进行量化,或者基于当前层预设的数据位宽n对当前层的输入神经元数据进行量化,获得量化后的定点数。根据量化前的输入神经元数据和对应的量化后的输入神经元数据,确定量化误差diff bit,根据量化误差diff bit与阈值的比较结果,对上一层的输入神经元数据量化时采用的数据位宽n或者当前层预设的数据位宽n进行调整,将调整后的数据位宽应用于当前层的输入神经元数据的量化。权值数据对应的数据位宽亦如此,此处不再赘述。
对于量化参数来说,由图5a可知,同一个周期(epoch)内,在训练初期,每次迭代权值变化幅度比较大。在训练中后期,由于每一次迭代前后权值数据变化幅度不大,使得每次迭代的对应层的权值数据之间在一定的迭代间隔内具有相似性,这样量化时当前次迭代的每层的数据可以延用上一次迭代的对应层的对应数据的量化参数,在训练的中后期不用代代都重新确认量化参数,仅仅在训练初期的每次迭代的每层确认量化参数,这样仍然能够满足神经网络运算所需的浮点运算的精度,大大提高了量化时的效率。进一步地,在训练或微调过程中,神经网络的每层之间的权值数据相互独立,不具备相似性。因权值数据不具备相似性使得每层之间的神经元数据也相互独立,不具备相似性。因此,在神经网络训练或微调过程中,神经网络的每一次迭代内的每层的量化参数应用于对应层的对应待量化数据。
上述以权值数据为例,在神经网络训练或微调过程中,神经元数据和梯度数据分别对应的量化参数亦如此,此处不再赘述。
在神经网络推理过程中,神经网络的每层之间的权值数据相互独立,不具备相似性。因权值数据不具备相似性使得每层之间的神经元数据也相互独立,不具备相似性。因此,在神经网络推理过程中,神经网络的每层的量化参数应用于对应层的待量化数据。比如:神经网络的当前层为卷积层,根据卷积层的待量化数据按照图2所示的技术方案获得了当前卷积层的待量化数据的量化参数,该量化参数只能应用于当前的卷积层,而不能应用于该神经网络的其他层,即使其它层为卷积层也不适用。
综上所述,数据位宽和量化参数的延用策略基于数据之间的相似性来确定,如果数据之间具有相似性,则数据位宽和量化参数可以延用,如果数据之间不具有相似性,则需要对数据位宽或量化参数进行调整。数据之间的相似性的度量通常采用KL散度来衡量,也可以采用下式(17)来衡量。
abs max(A)≈abs max(B)且mean(A)≈mean(B)    (17)
在某些实施例中,如果数据A和数据B满足式(17),则判断数据A和数据B之间具有相似性。
需要说明的是,关于上述描述的量化误差的确认方法、数据位宽的调整方法、数据位宽和量化参数的延用策略均仅仅是例举的部分情况,而不是穷举,比如:上述的量化误差的确认方法、数据位宽的调整方法、数据位宽和量化参数的延用策略均适用于神经网络的微调过程。还有,关于数据之间的相似性的度量,上述列举了KL散度以及公式(17)对相似性的度量方法,仅仅是例举的部分情况,而不是穷举,比如:直方图匹配法、矩阵分解法、基于特征点的图像相似度计算法、邻近度度量标准法等等。本领域技术人员在理解本申请技术方案的精髓的情况下,可能会在本申请技术方案的基础上产生其它的变形或者变换,但只要其实现的功能以及达到的技术效果与本申请类似,那么均应当属于本申请的保护范围。
综上所述,在训练中后期,因为每次迭代前后权值数据变化幅度不大,使得每次迭代 的对应层的权值数据之间在一定的迭代间隔内具有相似性,为了使得本技术方案在训练或微调中具有更好的适普性,满足人工智能处理器芯片的资源达到合理的应用,需要一种策略确定迭代间隔,使得在该迭代间隔范围内,每一次迭代的对应层的数据位宽n保持不变,超过该迭代间隔,数据位宽n就要发生变化,就无需代代确定是否对数据位宽n进行调整。同理,量化参数亦如此,从而达到提升人工智能处理器芯片的峰值算力的同时满足量化所需的浮点运算的精度。
如图6所示,为确定目标迭代间隔的方法流程图之一。在图6所示的技术方案中,目标迭代间隔包括至少一次权值更新迭代,且同一目标迭代间隔内量化过程中采用相同的所述数据位宽。所述目标迭代间隔的确定步骤包括:
步骤601):在预判时间点,确定权值迭代过程中所述待量化数据对应点位置参数的变化趋势值;其中,所述预判时间点是用于判断是否需要对所述数据位宽进行调整的时间点,所述预判时间点对应权值更新迭代完成时的时间点。
在本步骤中,按照公式(18),所述点位置参数的变化趋势值根据当前预判时间点对应的权值迭代过程中的点位置参数的滑动平均值、上一预判时间点对应的权值迭代过程中的点位置参数的滑动平均值确定,或者根据当前预判时间点对应的权值迭代过程中的点位置参数、上一预判时间点对应的权值迭代过程中的点位置参数的滑动平均值确定。式18的表达式为:
diff update1=|M (t)-M (t-1)|=α|s (t)-M (t-1)|   (18)
公式18中,M为点位置参数s随着训练迭代增加的滑动平均值。其中,M (t)为第t个预判时间点对应的点位置参数s随着训练迭代增加的滑动平均值,根据公式(19)获得M (t)。s (t)为第t个预判时间点对应的点位置参数s。M (t-1)为第t-1个预判时间点对应的点位置参数s的滑动平均值,α为超参数。diff update1衡量点位置参数s变化趋势,由于点位置参数s的变化也变相体现在当前待量化数据中数据最大值Z max的变化情况。diff update1越大,说明数值范围变化剧烈,需要间隔更短的更新频率,即目标迭代间隔更小。
M (t)←α×s (t-1)+(1-α)×M (t-1)   (19)
步骤602):根据所述点位置参数的变化趋势值确定对应所述目标迭代间隔。
在本技术方案中,根据式(20)确定目标迭代间隔。对于目标迭代间隔来说,同一目标迭代间隔内量化过程中采用相同的所述数据位宽,不同目标迭代间隔内量化过程中采用的数据位宽可以相同,也可以不同。
Figure PCTCN2019106754-appb-000059
式(20)中,I为目标迭代间隔。diff update1为点位置参数的变化趋势值。β、γ为经验值,也可以为可变的超参数。常规的超参数的优化方法均适于β、γ,这里不再赘述超参数的优化方案。
对于本技术方案来说,预判时间点包括第一预判时间点,根据目标迭代间隔确定第一预判时间点。具体地,在训练或微调过程中的第t个预判时间点,利用上一次迭代对应层量化时采用的数据位宽对当前次迭代的对应层的权值数据进行量化,获得量化后的定点数,根据量化前的权值数据和对应的量化前的权值数据,确定量化误差diff bit。将量化误差diff bit分别与第一阈值和第二阈值进行比较,利用比较结果确定是否对上一次迭代对应层量化时采用的数据位宽进行调整。假如:第t个第一预判时间点对应第100次迭代,第99次迭代使用的数据位宽为n 1。在第100次迭代,根据数据位宽n 1确认量化误差diff bit,将量化误差diff bit与第一阈值、第二阈值进行比较,获得比较结果。如果根据比较结果确认数据位宽n 1无需改变,利用式(20)确认目标迭代间隔为8次迭代,当第100次迭代作为当前目标迭代间隔内的起始迭代,那么第100次迭代~第107次迭代作为当前目标迭代间隔,当第100次迭代作为上一目标迭代间隔的最末迭代,那么第101次迭代~第108次迭代作为当前目标迭代间隔。在当前目标迭代间隔内量化时每代仍然延用上一个目标迭代间隔所使用的数据位宽n 1。这种情况,不同的目标迭代间隔之间量化时所使用的数据位宽可以相同。如果以第100次迭代~第107次迭代作为当前的目标迭代间隔,那么下一个目标迭代间隔内的第108次迭代作为第t+1个第一预判时间点,如果第101次迭代~第108次迭代作为当前的目标迭代间隔,那么当前的目标迭代间隔内的第108次迭代作为第t+1个第一预判时间点。在第t+1个第一预判时间点,根据数据位宽n 1确认量化误差diff bit,将量化误差diff bit与第 一阈值、第二阈值进行比较,获得比较结果。根据比较结果确定数据位宽n 1需要更改为n 2,并利用公式(20)确认目标迭代间隔为55次迭代。那么第108次迭代~第163次迭代或者第109次迭代~第163次迭代作为目标迭代间隔,在该目标迭代间隔内量化时每代使用数据位宽n 2。这种情况下,不同的目标迭代间隔之间量化时所使用的数据位宽可以不同。
对于本技术方案来说,不管第一预判时间点是目标迭代间隔内的起始迭代还是最末迭代,均适于公式(18)来获得点位置参数的变化趋势值。如果当前时刻的第一预判时间点为当前目标迭代间隔的起始迭代,那么在公式(18)中,M (t)为当前目标迭代间隔的起始迭代对应时间点所对应的点位置参数s随着训练迭代增加的滑动平均值,s (t)为当前目标迭代间隔的起始迭代对应时间点所对应的点位置参数s,M (t-1)为上一目标迭代间隔的起始迭代对应时间点所对应的点位置参数s随着训练迭代增加的滑动平均值。如果当前时刻的第一预判时间点为当前目标迭代间隔的最末迭代,那么在公式(18)中,M (t)为当前目标迭代间隔的最末迭代对应时间点所对应的点位置参数s随着训练迭代增加的滑动平均值,s (t)为当前目标迭代间隔的最末迭代对应时间点所对应的点位置参数s,M (t-1)为上一目标迭代间隔的最末迭代对应时间点所对应的点位置参数s随着训练迭代增加的滑动平均值。
对于本技术方案来说,在包括第一预判时间点的基础上,预判时间点还可以包括第二预判时间点。第二预判时间点是根据数据变动幅度曲线确定的。基于大数据在神经网络训练过程中数据变动幅度情况,获得如图5a所示的所述数据变动幅度曲线。
以权值数据为例,由图5a所示的数据变动幅度曲线可知,从训练开始到第T次迭代的迭代间隔周期内,每次权值更新时,数据变动幅度非常大。在当前预判时间点,量化时,当前次迭代先利用上一次迭代的数据位宽n 1进行量化,获得的量化结果与对应的量化前的数据确定对应的量化误差,量化误差分别与第一阈值、第二阈值进行比较,根据比较结果对数据位宽n 1进行调整,获得数据位宽n 2。利用数据位宽n 2对当前次迭代涉及的待量化权值数据进行量化。然后根据式(20)确定目标迭代间隔,从而确定第一预判时间点,在第一预判时间点再判断是否调整数据位宽以及如何调整,并根据公式(20)确定下一目标迭代间隔来获得下一个第一预判时间点。由于训练开始到第T次迭代的迭代间隔周期内,每一次迭代前后权值数据变化幅度非常大,使得每次迭代的对应层的权值数据之间不具有相 似性,为了满足精度问题,量化时当前次迭代的每层的数据不能延用上一次迭代的对应层的对应量化参数,在前T次迭代可以代代调整数据位宽,此时,量化时前T次迭代中每次迭代使用的数据位宽均不同,目标迭代间隔为1次迭代。为了人工智能处理器芯片的资源达到最优化利用,前T次迭代的目标迭代间隔可以根据图5a所示的数据变动幅度曲线图所揭示的规律提前预设好,即:根据数据变动幅度曲线前T次迭代的目标迭代间隔直接预设,无需经过公式(20)确认前T次迭代的每次迭代对应的权值更新迭代完成时的时间点作为第二预判时间点。从而使得人工智能处理器芯片的资源更为合理的利用。图5a所示的数据变动幅度曲线从第T次迭代开始变动幅度不大,在训练的中后期不用代代都重新确认量化参数,在第T次迭代或者第T+1次迭代,利用当前次迭代对应量化前的数据以及量化后的数据确定量化误差,根据量化误差确定对数据位宽是否需要调整以及如何调整,还要根据公式(20)确定目标迭代间隔。如果确认的目标迭代间隔为55次迭代,这就要求从第T次迭代或第T+1次迭代之后隔55次迭代对应的时间点作为第一预判时间点再判断是否调整数据位宽以及如何调整,并根据公式(20)确定下一目标迭代间隔,从而确定下一个第一预判时间点,直至同一周期(epoch)内所有代运算完成。在此基础上,在每个周期(epoch)之后,再对数据位宽或量化参数做适应性调整,最终使用量化后的数据获得精度符合预期的神经网络。
特别地,假如:根据图5a所示的权值数据变动幅度曲线图确定T取值为130(这个数值与图5a不对应,为方便描述,仅仅是假设T取值为130,不限于在假设值。),那么训练过程中的第130次迭代作为第二预判时间点,当前的第一预判时间点为训练过程中的第100次迭代,在第100次迭代,经公式(20)确定目标迭代间隔为35次迭代。在该目标迭代间隔内,训练至第130次迭代,到达第二预判时间点,此时就要在第130次迭代对应的时间点确定对数据位宽是否需要调整以及如何调整,还要根据公式(20)确定目标迭代间隔。假如该情况下确定的目标迭代间隔为42次迭代。就要从第130次迭代起至第172次迭代作为目标迭代间隔,目标迭代间隔为35次迭代时确定的第一预判时间点对应的第135次迭代处于目标迭代间隔为42次迭代内,在第135次迭代,可以再根据公式(20)判断是否需要调整数据位宽以及如何调整。也可以不在第135次迭代做评估预判,直接到第172次迭代 再执行是否需要调整数据位宽的评估以及如何调整。总之,是否在第135次迭代进行评估和预判均适于本技术方案。
综上,根据数据变动幅度曲线提前预设第二预判时间点,在训练或微调的初期,无需花费人工智能处理器芯片的资源来确定目标迭代间隔,在预设好的第二预判时间点上直接根据量化误差来调整数据位宽,并利用调整好的数据位宽来量化当前次迭代涉及的待量化数据。在训练或微调的中后期,根据公式(20)获得目标迭代间隔,从而确定对应的第一预判时间点,在每个第一预判时间点上确定是否调整数据位宽以及如何调整。这样在能够满足神经网络运算所需的浮点运算的精度的同时合理利用人工智能处理器芯片的资源,大大提高了量化时的效率。
在实际中,为了获得更准确的数据位宽的目标迭代间隔,不仅仅根据点位置参数的变化趋势值diff update1,可以同时考虑点位置参数的变化趋势值diff update1和数据位宽的变化趋势值diff update2。如图7所示,为确定目标迭代间隔的方法流程图之二。所述目标迭代间隔的确定步骤包括:
步骤701):在预判时间点,确定权值迭代过程中所述待量化数据对应点位置参数的变化趋势值、数据位宽的变化趋势值;其中,所述预判时间点是用于判断是否需要对所述数据位宽进行调整的时间点,所述预判时间点对应权值更新迭代完成时的时间点。
需要强调的是,图6所示的关于基于点位置参数的变化趋势值确定数据位宽的目标迭代间隔的技术方案内容适用于图7所示的技术方案,这里不再赘述。
在本步骤中,根据式(21)来利用对应所述量化误差确定所述数据位宽的变化趋势值。
Figure PCTCN2019106754-appb-000060
公式(21)中,δ为超参数,diff bit为量化误差;diff update2为数据位宽的变化趋势值。diff update2衡量量化时采用的数据位宽n的变化趋势,diff update2越大越有可能需要更新定点的位宽,需要间隔更短的更新频率。
对于图7中涉及的点位置参数的变化趋势值仍然可根据公式(18)获得,对于公式(18)中的M (t)根据公式(19)获得。diff update1衡量点位置参数s变化趋势,由于点位置参数s的变化也变相体现在当前待量化数据中数据最大值Z max的变化情况。diff update1越大,说明数值 范围变化剧烈,需要间隔更短的更新频率,即目标迭代间隔更小。
步骤702):根据所述点位置参数的变化趋势值和所述数据位宽的变化趋势值确定对应所述目标迭代间隔。
在本技术方案中,根据公式(22)确定目标迭代间隔。对于目标迭代间隔来说,同一目标迭代间隔内量化过程中采用相同的所述数据位宽,不同目标迭代间隔内量化过程中采用的数据位宽可以相同,也可以不同。
Figure PCTCN2019106754-appb-000061
公式(22)中,I为目标迭代间隔。β、γ为超参数。diff update1为点位置参数的变化趋势值。diff update2为数据位宽的变化趋势值。β、γ为经验值,也可以为可变的超参数。常规的超参数的优化方法均适于β、γ,这里不再赘述超参数的优化方案。
对于本技术方案来说,diff update1是用来衡量点位置参数s的变化情况,但是由数据位宽n的变化而导致的点位置参数s的变化是要忽略掉的。因为这已经在diff update2中体现过了数据位宽n的变化。如果在diff update1中不做这个忽略的操作,那么根据公式(22)确定的目标迭代间隔I是不准确的,造成第一预判时间点过多,在训练或微调过程中,易频繁的做数据位宽n是否更新以及如何更新的操作,从而造成人工智能处理器芯片的资源没有合理利用。
基于上述描述,diff update1根据M (t)确定。假设第t-1个预判时间点对应的数据位宽为n 1,对应的点位置参数为s 1,点位置参数随着训练迭代增加的滑动平均值为m 1。利用数据位宽n 1对待量化数据进行量化,获得量化后的定点数。根据量化前的数据和对应的量化后的数据,确定量化误差diff bit,根据量化误差diff bit与阈值的比较结果,将数据位宽n 1调整为n 2,数据位宽调整了|n 1-n 2|位,第t个预判时间点量化时使用的数据位宽为n 2。为了忽略由数据位宽的变化而导致的点位置参数的变化,在确定M (t)时可以选出下述两种优化方式中的其中一种即可。第一种方式:如果数据位宽增加了|n 1-n 2|位,则s (t-1)取值为s 1-|n 1-n 2|,M (t-1)取值为m 1-|n 1-n 2|,将s (t-1)、M (t-1)代入公式(19)中,获得M (t),即为第t个预判时间点对应的点位置参数随着训练迭代增加的滑动平均值。如果数据位宽减少了|n 1-n 2|位,则s (t-1) 取值为s 1+|n 1-n 2|,M (t-1)取值为m 1+|n 1-n 2|,将s (t-1)、M (t-1)代入公式(19)中,获得M (t),即为第t个预判时间点对应的点位置参数随着训练迭代增加的滑动平均值。第二种方式:不管数据位宽是增加了|n 1-n 2|位还是减少了|n 1-n 2|,s (t-1)取值为s 1,M (t-1)取值为m 1,将s (t-1)、M (t-1)代入公式(19)中,获得M (t)。在数据位宽增加|n 1-n 2|位时,将M (t)减去|n 1-n 2|,在数据位宽减少|n 1-n 2|位时,将M (t)加上|n 1-n 2|,结果作为第t个预判时间点对应的点位置参数随着训练迭代增加的滑动平均值。这两种方式是等价的,均可以忽略由数据位宽的变化而导致的点位置参数的变化,获得更为精准的目标迭代间隔,从而提高人工智能处理器芯片的资源利用率。
在实际应用中,数据位宽n和点位置参数s对量化精度影响很大,量化参数中的第二缩放系数f 2以及偏移量O对量化精度影响不大。对于第一缩放系数f 1来说,上文已经提及,如果属于第二种情况,将2 s×f 2作为一个整体当做第一缩放系数f 1,由于点位置参数s对量化精度影响很大,那么此种情况下的第一缩放系数f 1对量化影响很大。所以,在本技术方案中,不管数据位宽n是否发生变化、点位置参数s可变的情况下,确定点位置参数s的目标迭代间隔也是一件非常有意义的事情,图6所示的技术方案的思想可应用于确定点位置参数s的目标迭代间隔。因此,确定点位置参数s的目标迭代间隔的方法如图8所示。包括:
步骤801):在预判时间点,确定权值迭代过程中涉及的待量化数据对应点位置参数的变化趋势值;其中,所述预判时间点是用于判断是否需要对所述量化参数进行调整的时间点,所述预判时间点对应权值更新迭代完成时的时间点。
步骤802):根据所述点位置参数的变化趋势值确定对应所述目标迭代间隔。
需要强调的是,图6所示的关于基于点位置参数的变化趋势值确定量化参数的目标迭代间隔的技术方案内容适用于图8所示的技术方案,这里不再赘述。对于图8所示的技术方案来说,量化参数优选为点位置参数。
需要说明的是,关于上述确定数据位宽的目标迭代间隔和量化参数的目标迭代间隔均仅仅是例举的部分情况,而不是穷举,本领域技术人员在理解本申请技术方案的精髓的情况下,可能会在本申请技术方案的基础上产生其它的变形或者变换,比如:在确定数据位宽的目标迭代间隔内再确定量化参数的目标迭代间隔也适用于图6、图7和图8所示的技 术方案。但只要其实现的功能以及达到的技术效果与本申请类似,那么均应当属于本申请的保护范围。
利用本技术方案确定量化参数,根据量化误差对数据位宽或量化参数进行调整,并确定了对数据位宽或量化参数是否调整的目标迭代间隔,达到神经网络运算过程中在适合的时间点对数据位宽或量化参数进行调整,使得在合适的迭代时间点使用合适的量化参数,实现人工智能处理器芯片执行神经网络运算达到定点运算的速度,提升了人工智能处理器芯片的峰值算力的同时满足运算所需的浮点运算的精度。
需要说明的是,对于前述的各方法实施例,为了简单描述,故将其都表述为一系列的动作组合,但是本领域技术人员应该知悉,本披露并不受所描述的动作顺序的限制,因为依据本披露,某些步骤可以采用其他顺序或者同时进行。其次,本领域技术人员也应该知悉,说明书中所描述的实施例均属于可选实施例,所涉及的动作和模块并不一定是本披露所必须的。
进一步需要说明的是,虽然图2、图6、图7、图8的流程图中的各个步骤按照箭头的指示依次显示,但是这些步骤并不是必然按照箭头指示的顺序依次执行。除非本文中有明确的说明,这些步骤的执行并没有严格的顺序限制,这些步骤可以以其它的顺序执行。而且,图2、图6、图7、图8中的至少一部分步骤可以包括多个子步骤或者多个阶段,这些子步骤或者阶段并不必然是在同一时刻执行完成,而是可以在不同的时刻执行,这些子步骤或者阶段的执行顺序也不必然是依次进行,而是可以与其它步骤或者其它步骤的子步骤或者阶段的至少一部分轮流或者交替地执行。
如图9所示,为本申请提出的一种神经网络的量化参数确定装置的硬件配置的框图。在图9中,神经网络的量化参数确定装置10可以包括处理器110和存储器120。在图9的神经网络的量化参数确定装置10中,仅示出了与本实施例有关的组成元素。因此,对于本领域普通技术人员而言显而易见的是:神经网络的量化参数确定装置10还可以包括与图10中所示的组成元素不同的常见组成元素。比如:定点运算器。
神经网络的量化参数确定装置10可以对应于具有各种处理功能的计算设备,例如,用于生成神经网络、训练或学习神经网络、将浮点型神经网络量化为定点型神经网络、或者重新训练神经网络的功能。例如,神经网络的量化参数确定装置10可以被实现为各种类型的设备,例如个人计算机(PC)、服务器设备、移动设备等。
处理器110控制神经网络的量化参数确定装置10的所有功能。例如,处理器110通过 执行神经网络的量化参数确定装置10上的存储器120中存储的程序,来控制神经网络的量化参数确定装置10的所有功能。处理器110可以由神经网络的量化参数确定装置10中提供的中央处理单元(CPU)、图形处理单元(GPU)、应用处理器(AP)、人工智能处理器芯片(IPU)等来实现。然而,本公开并不限于此。
存储器120是用于存储神经网络的量化参数确定装置10中处理的各种数据的硬件。例如,存储器120可以存储神经网络的量化参数确定装置10中的处理过的数据和待处理的数据。存储器120可存储处理器110已处理或要处理的神经网络运算过程中涉及的数据集,例如,未训练的初始神经网络的数据、在训练过程中生成的神经网络的中间数据、完成了所有训练的神经网络的数据、经量化的神经网络的数据等。此外,存储器120可以存储要由神经网络的量化参数确定装置10驱动的应用、驱动程序等。例如:存储器120可以存储与将由处理器110执行的神经网络的训练算法、量化算法等有关的各种程序。存储器120可以是DRAM,但是本公开不限于此。存储器120可以包括易失性存储器或非易失性存储器中的至少一种。非易失性存储器可以包括只读存储器(ROM)、可编程ROM(PROM)、电可编程ROM(EPROM)、电可擦除可编程ROM(EEPROM)、闪存、相变RAM(PRAM)、磁性RAM(MRAM)、电阻RAM(RRAM)、铁电RAM(FRAM)等。易失性存储器可以包括动态RAM(DRAM)、静态RAM(SRAM)、同步DRAM(SDRAM)、PRAM、MRAM、RRAM、铁电RAM(FeRAM)等。在实施例中,存储器120可以包括硬盘驱动器(HDD)、固态驱动器(SSD)、高密度闪存(CF)、安全数字(SD)卡、微安全数字(Micro-SD)卡、迷你安全数字(Mini-SD)卡、极限数字(xD)卡、高速缓存(caches)或记忆棒中的至少一项。
处理器110可以通过反复训练(学习)给定的初始神经网络来生成经训练的神经网络。在这种状态下,在保证神经网络的处理准确度的意义上,初始神经网络的参数为高精度数据表示格式,例如具有32比特浮点精度的数据表示格式。参数可以包括向/从神经网络输入/输出的各种类型的数据,例如:神经网络的输入/输出神经元、权值、偏置等。与定点运算相比,浮点运算过程中需要相对大量的运算和相对频繁的存储器访问。具体而言,神经网络处理所需的大部分运算已知为各种卷积运算。因此,在具有相对低的处理性能的移动设备(诸如智能电话、平板电脑、可穿戴设备等、嵌入式设备等)中,神经网络高精度数据运算会使得移动设备的资源没有充分利用。结果是,为了在允许的精度损失范围内驱动神经网络运算,充分减少上述设备中的运算量,可以对在神经网络运算过程中涉及的高精度数据进行量化,转换为低精度的定点数。
考虑到部署神经网络的例如移动设备、嵌入式设备等设备的处理性能,神经网络的量 化参数确定装置10执行将经训练的神经网络的参数转换为具有特定比特数的定点型的量化,并且神经网络的量化参数确定装置10向部署神经网络的设备发送对应的量化参数,使得在人工智能处理器芯片执行训练、微调等运算操作时为定点数运算操作。部署神经网络的设备可以是通过使用神经网络来执行语音识别、图像识别等的自主车辆、机器人、智能电话、平板设备、增强现实(AR)设备、物联网(IoT)设备等,但是本公开不限于此。
处理器110从存储器120中获取神经网络运算过程中的数据。该数据包括神经元、权值、偏置和梯度中的至少一种数据,利用图2所示的技术方案确定对应的量化参数,利用量化参数对神经网络运算过程中的目标数据进行量化。将量化后的数据执行神经网络运算操作。该运算操作包括但不限于训练、微调、推理。
处理器110根据量化误差diff bit对数据位宽n进行调整,且处理器110可以执行图6、图7和图8所示的目标迭代间隔的方法的程序去确定数据位宽的目标迭代间隔或量化参数的目标迭代间隔。
综上,本说明书实施方式提供的一种神经网络的量化参数确定装置,其存储器120和处理器110实现的具体功能,可以与本说明书中的前述实施方式相对照解释,并能够达到前述实施方式的技术效果,这里便不再赘述。
在本实施方式中,所述处理器110可以按任何适当的方式实现。例如,所述处理器110可以采取例如微处理器或处理器以及存储可由该(微)处理器执行的计算机可读程序代码(例如软件或固件)的计算机可读介质、逻辑门、开关、专用集成电路(Application Specific Integrated Circuit,ASIC)、可编程逻辑控制器和嵌入微控制器的形式等等。
如图10所示,为本申请提出的神经网络的量化参数确定装置应用于人工智能处理器芯片的应用示意图。参考图10,如上所述,在诸如PC、服务器等神经网络的量化参数确定装置10中,处理器110执行量化操作,将神经网络运算过程中涉及的浮点数据量化为定点数,人工智能处理器芯片上的定点运算器采用量化获得的定点数执行训练、微调或推理。人工智能处理器芯片是用于驱动神经网络的专用硬件。由于人工智能处理器芯片是以相对较低的功率或性能实现的,利用本技术方案采用低精度的定点数实现神经网络运算,与高精度数据相比,读取低精度的定点数时所需内存带宽更小,可以更好的使用人工智能处理器芯片的caches,避免访存瓶颈。同时,在人工智能处理器芯片上执行SIMD指令时,在一个时钟周期内实现更多计算,达到更快地执行神经网络运算操作。
进一步地,面对同样长度的定点运算和高精度数据运算,尤其是定点运算和浮点运算之间比对可知,浮点运算计算模式更为复杂,需要更多的逻辑器件来构成浮点运算器。这 样从体积上来说,浮点运算器的体积比定点运算器的体积要大。并且,浮点运算器需要消耗更多的资源去处理,达到定点运算和浮点运算二者之间的功耗差距通常是数量级的。
综上所述,本技术方案能够让人工智能处理器芯片上的浮点运算器更换为定点运算器,使得人工智能处理器芯片的功耗更低。这一点对于移动设备尤其重要。也就是说,本技术方案打开了一扇通向大量不能高效运行浮点计算代码的 嵌入式系统的大门,让物联网世界广泛应用成为可能。
在本技术方案中,人工智能处理器芯片可以对应于例如神经处理单元(NPU)、张量处理单元(TPU)、神经引擎等,它们是用于驱动神经网络的专用芯片,但是本公开不限于此。
在本技术方案中,人工智能处理器芯片可以在独立于神经网络的量化参数确定装置10的单独设备中实现,神经网络的量化参数确定装置10也可以作为人工智能处理器芯片的一部分功能模块来实现。但是本公开不限于此。
在本技术方案中,通用处理器(比如CPU)的操作系统基于本技术方案生成指令,将生成的指令发送至人工智能处理器芯片(比如GPU)上,由人工智能处理器芯片去执行指令操作实现神经网络的量化参数的确定以及量化过程。还有一种应用情况,通用处理器基于本技术方案直接确定对应的量化参数,通用处理器直接根据量化参数将对应的目标数据进行量化,人工智能处理器芯片利用量化后的数据执行定点运算操作。更甚者,通用处理器(比如CPU)和人工智能处理器芯片(比如GPU)流水化操作,通用处理器(比如CPU)的操作系统基于本技术方案生成指令,且对目标数据进行拷贝的同时人工智能处理器芯片(比如GPU)进行神经网络运算操作,这样可以把某些时间消耗隐藏起来。但是本公开不限于此。
在本实施例中,本申请实施例还提供一种可读存储介质,其上存储有计算机程序,所述计算机程序被执行时实现上述所述的神经网络的量化参数确定方法。
由上可见,在神经网络运算过程中,量化时利用本公开的技术方案确定量化参数,该量化参数用于人工智能处理器对神经网络运算过程中的数据进行量化,将高精度数据转换为低精度定点数,可以减少神经网络运算过程中涉及的数据存储所有的空间大小。例如:float32转化为fix8可以将模型参数减少4倍。由于数据存储空间变小,使得神经网络部署时使用更小的空间,使得人工智能处理器芯片上的片上内存可以容纳更多的数据,减少了人工智能处理器芯片访存数据,提高计算性能。
本领域技术人员也知道,除了以纯计算机可读程序代码方式实现客户端和服务器以外,完全可以通过将方法步骤进行逻辑编程来使得客户端和服务器以逻辑门、开关、专用集成 电路、可编程逻辑控制器和嵌入微控制器等的形式来实现相同功能。因此这种客户端和服务器可以被认为是一种硬件部件,而对其内包括的用于实现各种功能的装置也可以视为硬件部件内的结构。或者甚至,可以将用于实现各种功能的装置视为既可以是实现方法的软件模块又可以是硬件部件内的结构。
如图11所示,为本申请提出的一种神经网络的量化参数确定设备的功能框图。所述方法包括:
统计结果获取单元a,用于获取每种待量化数据的统计结果;其中,所述待量化数据包括所述神经网络的神经元、权值、梯度、偏置中的至少一种数据;
量化参数确定单元b,用于利用每种待量化数据的统计结果以及数据位宽确定对应量化参数;其中,所述量化参数用于人工智能处理器对神经网络运算过程中的数据进行对应量化。
在本实施例中,可选地,所述神经网络的量化参数确定设备还包括:
第一量化单元,用于利用对应量化参数对所述待量化数据进行量化。
在本实施例中,可选地,所述神经网络的量化参数确定设备还包括:
第二量化单元,用于利用对应量化参数对目标数据进行量化;其中,所述目标数据的特征与所述待量化数据的特征之间具有相似性。
在本实施例中,所述神经网络运算过程包括神经网络训练、神经网络推理、神经网络微调中的至少一种运算。
在本实施例中,所述统计单元获得的统计结果为每种待量化数据中的最大值和最小值。
在本实施例中,所述统计单元获得的统计结果为每种待量化数据中的绝对值最大值。
在本实施例中,所述统计单元根据每种待量化数据中的最大值和最小值确定所述绝对值最大值。
在本实施例中,所述量化参数确定单元根据每种待量化数据中的最大值、最小值以及所述数据位宽确定量化参数。
在本实施例中,所述量化参数确定单元根据每种待量化数据中的绝对值最大值、所述数据位宽确定量化参数。
在本实施例中,所述量化参数确定单元确定的所述量化参数为点位置参数或第一缩放 系数。
在本实施例中,所述量化参数确定单元根据点位置参数和第二缩放系数确定所述第一缩放系数;其中,确定第一缩放系数时使用的点位置参数为已知固定值,或所述点位置参数和对应的所述第二缩放系数相乘的结果整体作为第一缩放系数应用于神经网络运算过程中的数据量化。
在本实施例中,所述量化参数确定单元确定的所述量化参数包括点位置参数和第二缩放系数。
在本实施例中,所述量化参数确定单元根据所述点位置参数、所述统计结果、所述数据位宽确定所述第二缩放系数。
在本实施例中,所述量化参数确定单元确定的所述量化参数还包括偏移量。
在本实施例中,所述量化参数确定单元根据每种待量化数据的统计结果确定所述偏移量。
在本实施例中,所述量化参数确定单元使用的数据位宽是预设值。
在本实施例中,所述量化参数确定单元包括调整模块和量化误差确定模块;其中,
所述调整模块,用于根据对应的量化误差对数据位宽进行调整;
所述量化误差确定模块,用于根据量化后的数据与对应的量化前的数据确定所述量化误差。
在本实施例中,所述调整模块具体用于:
所述量化误差与阈值进行比较,根据比较结果,调整所述数据位宽;其中,所述阈值包括第一阈值和第二阈值中的至少一个。
在本实施例中,所述调整模块包括第一调整子模块,其中,所述第一调整子模块用于:
所述量化误差大于等于所述第一阈值,则对所述数据位宽进行增加。
在本实施例中,所述调整模块包括第二调整子模块,其中,所述第二调整子模块用于:
所述量化误差小于等于所述第二阈值,则对所述数据位宽进行减少。
在本实施例中,所述调整模块包括第三调整子模块,其中,所述第三调整子模块用于:
所述量化误差处于所述第一阈值和所述第二阈值之间,则所述数据位宽保持不变。
在本实施例中,所述量化误差确定模块包括:
量化间隔确定子模块,用于根据所述数据位宽确定量化间隔;
第一量化误差确定子模块,用于根据所述量化间隔、所述量化后的数据的个数与对应的量化前的数据确定量化误差。
在本实施例中,所述量化误差确定模块包括:
反量化数据确定子模块,用于对量化后的数据进行反量化,获得反量化数据;其中,所述反量化数据的数据格式与对应的量化前的数据的数据格式相同;
第二量化误差确定子模块,用于根据所述量化后的数据以及对应的反量化数据确定量化误差。
在本实施例中,所述量化误差确定模块使用的所述量化前的数据是所述待量化数据。
在本实施例中,所述量化误差确定模块使用的所述量化前的数据是在目标迭代间隔内的权值更新迭代过程中涉及的待量化数据;其中,所述目标迭代间隔包括至少一次权值更新迭代,且同一目标迭代间隔内量化过程中采用相同的所述数据位宽。
在本实施例中,所述神经网络的量化参数确定设备还包括第一目标迭代间隔确定单元;其中,所述第一目标迭代间隔确定单元包括:
第一变化趋势值确定模块,用于在预判时间点,确定权值更新迭代过程中涉及的待量化数据的点位置参数的变化趋势值;其中,所述预判时间点是用于判断是否需要对所述数据位宽进行调整的时间点,所述预判时间点对应权值更新迭代完成时的时间点;
第一目标迭代间隔模块,用于根据所述点位置参数的变化趋势值确定对应所述目标迭代间隔。
在本实施例中,所述第一目标迭代间隔确定单元包括:
第二变化趋势值确定模块,用于在预判时间点,确定权值更新迭代过程中涉及的待量化数据的点位置参数的变化趋势值、数据位宽的变化趋势值;其中,所述预判时间点是用于判断是否需要对所述数据位宽进行调整的时间点,所述预判时间点对应权值更新迭代完成时的时间点;
第二目标迭代间隔模块,用于根据所述点位置参数的变化趋势值和所述数据位宽的变 化趋势值确定对应所述目标迭代间隔。
在本实施例中,所述第一目标迭代间隔确定单元还包括第一预判时间点确定单元;其中,
所述第一预判时间点确定单元,用于根据所述目标迭代间隔确定所述第一预判时间点。
在本实施例中,所述第一目标迭代间隔确定单元还包括第二预判时间点确定单元;其中,所述第二预判时间点确定单元,用于根据数据变动幅度曲线确定第二预判时间点;其中,所述数据变动幅度曲线是对权值更新迭代过程中数据变动幅度情况进行统计获得的。
在本实施例中,所述第一变化趋势值确定模块和所述第二变化趋势值确定模块均根据当前预判时间点对应的点位置参数的滑动平均值、上一预判时间点对应的点位置参数的滑动平均值确定所述点位置参数的变化趋势值。
在本实施例中,所述第一变化趋势值确定模块和所述第二变化趋势值确定模块均根据当前预判时间点对应的点位置参数、上一预判时间点对应的点位置参数的滑动平均值确定所述点位置参数的变化趋势值。
在本实施例中,所述第一变化趋势值确定模块和所述第二变化趋势值确定模块均包括:
当前预判时间点对应的点位置参数确定子模块,用于根据上一预判时间点对应的点位置参数与所述数据位宽的调整值确定所述当前预判时间点对应的点位置参数;
调整结果确定子模块,用于根据所述数据位宽的调整值对所述上一预判时间点对应的点位置参数的滑动平均值进行调整,获得调整结果;
第一滑动平均值确定子模块,用于根据所述当前预判时间点对应的点位置参数、所述调整结果确定当前预判时间点对应的点位置参数的滑动平均值。
在本实施例中,所述第一变化趋势值确定模块和所述第二变化趋势值确定模块均包括:
中间结果确定子模块,用于根据上一预判时间点对应的点位置参数与上一预判时间点对应的点位置参数的滑动平均值确定当前预判时间点对应的点位置参数的滑动平均值的中间结果;
第二滑动平均值确定子模块,用于根据当前预判时间点对应的点位置参数的滑动平均值的中间结果与所述数据位宽的调整值确定所述当前预判时间点对应的点位置参数的滑动 平均值。
在本实施例中,所述第二变化趋势值确定模块根据对应所述量化误差确定数据位宽的变化趋势值。
在本实施例中,所述第一目标迭代间隔确定单元还包括:
量化误差确定模块,用于确定对应量化误差;其中,所述量化误差对应的量化前的数据是所述预判时间点对应的权值更新迭代过程中涉及的待量化数据;
数据位宽确定模块,用于根据对应量化误差,确定所述目标迭代间隔内量化过程中采用的数据位宽。
在本实施例中,所述数据位宽确定模块具体用于:
所述量化误差与阈值进行比较,根据比较结果,对上一目标迭代间隔内量化过程中采用的数据位宽进行调整,调整结果作为当前目标迭代间隔内量化过程中采用的数据位宽。
在本实施例中,所述量化误差确定模块使用的所述量化前的数据是在目标迭代间隔内的权值更新迭代时涉及的待量化数据;其中,所述目标迭代间隔包括至少一次权值更新迭代,且同一目标迭代间隔内量化过程中采用相同的所述量化参数。
在本实施例中,所述神经网络的量化参数确定设备还包括第二目标迭代间隔确定单元;其中,所述第二目标迭代间隔确定单元包括:
第三变化趋势值确定模块,用于在预判时间点,确定权值更新迭代过程中涉及的待量化数据的点位置参数的变化趋势值;其中,所述预判时间点是用于判断是否需要对所述量化参数进行调整的时间点,所述预判时间点对应权值更新迭代完成时的时间点;
第三目标迭代间隔模块,用于根据所述点位置参数的变化趋势值确定对应所述目标迭代间隔。
在本实施例中,所述量化参数确定单元根据统计结果、和所述数据位宽确定所述点位置参数。
应该理解,上述的装置实施例仅是示意性的,本披露的设备还可通过其它的方式实现。例如,上述实施例中所述单元/模块的划分,仅仅为一种逻辑功能划分,实际实现时可以有另外的划分方式。例如,多个单元、模块或组件可以结合,或者可以集成到另一个系统,或一些特征可以忽略或不执行。
所述作为分离部件说明的单元或模块可以是物理上分开的,也可以不是物理上分开的。作为单元或模块说明的部件可以是物理单元,也可以不是物理单元,即可以位于一个装置中,或者也可以分布到多个装置上。本披露中实施例的方案可以根据实际的需要选择其中的部分或者全部单元来实现。
另外,若无特别说明,在本披露各个实施例中的各功能单元/模块可以集成在一个单元/模块中,也可以是各个单元/模块单独物理存在,也可以两个或两个以上单元/模块集成在一起。上述集成的单元/模块既可以采用硬件的形式实现,也可以采用软件程序模块的形式实现。
应该理解,上述的装置实施例仅是示意性的,本披露的装置还可通过其它的方式实现。例如,上述实施例中所述单元/模块的划分,仅仅为一种逻辑功能划分,实际实现时可以有另外的划分方式。例如,多个单元、模块或组件可以结合,或者可以集成到另一个系统,或一些特征可以忽略或不执行。
所述作为分离部件说明的单元或模块可以是物理上分开的,也可以不是物理上分开的。作为单元或模块说明的部件可以是物理单元,也可以不是物理单元,即可以位于一个装置中,或者也可以分布到多个装置上。本披露中实施例的方案可以根据实际的需要选择其中的部分或者全部单元来实现。
另外,若无特别说明,在本披露各个实施例中的各功能单元/模块可以集成在一个单元/模块中,也可以是各个单元/模块单独物理存在,也可以两个或两个以上单元/模块集成在一起。上述集成的单元/模块既可以采用硬件的形式实现,也可以采用软件程序模块的形式实现。
所述集成的单元/模块如果以硬件的形式实现时,该硬件可以是数字电路,模拟电路等等。硬件结构的物理实现包括但不局限于晶体管,忆阻器等等。若无特别说明,所述人工智能处理器可以是任何适当的硬件处理器,比如:CPU、GPU、FPGA、DSP和ASIC等等。若无特别说明,所述存储单元可以是任何适当的磁存储介质或者磁光存储介质,比如:阻变式存储器RRAM(Resistive Random Access Memory)、动态随机存取存储器DRAM(Dynamic Random Access Memory)、静态随机存取存储器SRAM(Static Random-Access Memory)、增强动态随机存取存储器EDRAM(Enhanced Dynamic Random Access Memory)、高带宽内存HBM(High-Bandwidth Memory)、混合存储立方HMC(Hybrid Memory Cube)等等。
所述集成的单元/模块如果以软件程序模块的形式实现并作为独立的产品销售或使用时,可以存储在一个计算机可读取存储器中。基于这样的理解,本披露的技术方案本质上或者说对现有技术做出贡献的部分或者该技术方案的全部或部分可以以软件产品的形式体现出来,该计算机软件产品存储在一个存储器中,包括若干指令用以使得一台计算机设备(可为个人计算机、服务器或者网络设备等)执行本披露各个实施例所述方法的全部或部分步骤。而前述的存储器包括:U盘、只读存储器(ROM,Read-Only Memory)、随机存取存储器(RAM,Random Access Memory)、移动硬盘、磁碟或者光盘等各种可以存储程序代码的介质。
在本技术方案中,本披露还公开了一种人工智能芯片,其包括了上述神经网络的量化参数确定设备。
在本技术方案中,本披露还公开了一种板卡,其包括存储器件、接口装置和控制器件以及上述人工智能芯片;其中,所述人工智能芯片与所述存储器件、所述控制器件以及所述接口装置分别连接;所述存储器件,用于存储数据;所述接口装置,用于实现所述人工智能芯片与外部设备之间的数据传输;所述控制器件,用于对所述人工智能芯片的状态进行监控。
图12示出根据本披露实施例的板卡的结构框图,参阅图12,上述板卡除了包括上述芯片389以外,还可以包括其他的配套部件,该配套部件包括但不限于:存储器件390、接口装置391和控制器件392;
所述存储器件390与所述人工智能芯片通过总线连接,用于存储数据。所述存储器件可以包括多组存储单元393。每一组所述存储单元与所述人工智能芯片通过总线连接。可以理解,每一组所述存储单元可以是DDR SDRAM(英文:Double Data Rate SDRAM,双倍速率同步动态随机存储器)。
DDR不需要提高时钟频率就能加倍提高SDRAM的速度。DDR允许在时钟脉冲的上升沿和下降沿读出数据。DDR的速度是标准SDRAM的两倍。在一个实施例中,所述存储装置可以包括4组所述存储单元。每一组所述存储单元可以包括多个DDR4颗粒(芯片)。在一个实施例中,所述人工智能芯片内部可以包括4个72位DDR4控制器,上述72位DDR4控制器中64bit用于传输数据,8bit用于ECC校验。可以理解,当每一组所述存储单元中采用DDR4-3200颗粒时,数据传输的理论带宽可达到25600MB/s。
在一个实施例中,每一组所述存储单元包括多个并联设置的双倍速率同步动态随机存 储器。DDR在一个时钟周期内可以传输两次数据。在所述芯片中设置控制DDR的控制器,用于对每个所述存储单元的数据传输与数据存储的控制。
所述接口装置与所述人工智能芯片电连接。所述接口装置用于实现所述人工智能芯片与外部设备(例如服务器或计算机)之间的数据传输。例如在一个实施例中,所述接口装置可以为标准PCIE接口。比如,待处理的数据由服务器通过标准PCIE接口传递至所述芯片,实现数据转移。优选的,当采用PCIE 3.0 X 16接口传输时,理论带宽可达到16000MB/s。在另一个实施例中,所述接口装置还可以是其他的接口,本披露并不限制上述其他的接口的具体表现形式,所述接口单元能够实现转接功能即可。另外,所述人工智能芯片的计算结果仍由所述接口装置传送回外部设备(例如服务器)。
所述控制器件与所述人工智能芯片电连接。所述控制器件用于对所述人工智能芯片的状态进行监控。具体的,所述人工智能芯片与所述控制器件可以通过SPI接口电连接。所述控制器件可以包括单片机(Micro Controller Unit,MCU)。如所述人工智能芯片可以包括多个处理芯片、多个处理核或多个处理电路,可以带动多个负载。因此,所述人工智能芯片可以处于多负载和轻负载等不同的工作状态。通过所述控制装置可以实现对所述人工智能芯片中多个处理芯片、多个处理和或多个处理电路的工作状态的调控。
在一种可能的实现方式中,公开了一种电子设备,其包括了上述人工智能芯片。电子设备包括数据处理装置、机器人、电脑、打印机、扫描仪、平板电脑、智能终端、手机、行车记录仪、导航仪、传感器、摄像头、服务器、云端服务器、相机、摄像机、投影仪、手表、耳机、移动存储、可穿戴设备、交通工具、家用电器、和/或医疗设备。
所述交通工具包括飞机、轮船和/或车辆;所述家用电器包括电视、空调、微波炉、冰箱、电饭煲、加湿器、洗衣机、电灯、燃气灶、油烟机;所述医疗设备包括核磁共振仪、B超仪和/或心电图仪。
依据以下条款可更好地理解前述内容:
A1.一种神经网络的量化参数确定方法,所述方法包括:
获取每种待量化数据的统计结果;其中,所述待量化数据包括所述神经网络的神经元、权值、梯度、偏置中的至少一种数据;
利用每种待量化数据的统计结果以及数据位宽确定对应量化参数;其中,所述量化参数用于人工智能处理器对神经网络运算过程中的数据进行对应量化。
A2.如条款A1所述的方法,所述方法还包括:
利用对应量化参数对所述待量化数据进行量化。
A3.如条款A1或条款A 2所述的方法,所述方法还包括:
利用对应量化参数对目标数据进行量化;其中,所述目标数据的特征与所述待量化数据的特征之间具有相似性。
A4.如条款A1所述的方法,所述神经网络运算过程包括神经网络训练、神经网络推理、神经网络微调中的至少一种运算。
A5.如条款A1所述的方法,所述统计结果为每种待量化数据中的最大值和最小值。
A6.如条款A1所述的方法,所述统计结果为每种待量化数据中的绝对值最大值。
A7.如条款A6所述的方法,所述绝对值最大值根据每种待量化数据中的最大值和最小值确定。
A8.如条款A5所述的方法,所述量化参数根据每种待量化数据中的最大值、最小值以及所述数据位宽确定。
A9.如条款A6或条款A7所述的方法,所述量化参数根据每种待量化数据中的绝对值最大值、所述数据位宽确定。
A10.如条款A1所述的方法,所述量化参数为点位置参数或第一缩放系数。
A11.如条款A10所述的方法,所述第一缩放系数根据点位置参数和第二缩放系数确定;其中,确定第一缩放系数时使用的点位置参数为已知固定值,或所述点位置参数和对应的所述第二缩放系数相乘的结果整体作为第一缩放系数应用于神经网络运算过程中的数据量化。
A12.如条款A10所述的方法,所述量化参数包括点位置参数和第二缩放系数。
A13.如条款A12所述的方法,所述第二缩放系数根据所述点位置参数、所述统计结果、所述数据位宽确定。
A14.如条款A10~条款A12任一项条款所述的方法,所述量化参数还包括偏移量。
A15如条款A14所述的方法,所述偏移量根据每种待量化数据的统计结果确定。
A16.如条款A1所述的方法,所述数据位宽是预设值。
A17.如条款A1所述的方法,所述数据位宽根据对应的量化误差进行调整;其中,所述量化误差是根据量化后的数据与对应的量化前的数据确定。
A18.如条款A17所述的方法,所述数据位宽的调整步骤包括:
所述量化误差与阈值进行比较,根据比较结果,调整所述数据位宽;其中,所述阈值包括第一阈值和第二阈值中的至少一个。
A19.如条款A18所述的方法,调整所述数据位宽的步骤包括:
所述量化误差大于等于所述第一阈值,则对所述数据位宽进行增加。
A20.如条款A18所述的方法,调整所述数据位宽的步骤包括:
所述量化误差小于等于所述第二阈值,则对所述数据位宽进行减少。
A21.如条款A18所述的方法,调整所述数据位宽的步骤包括:
所述量化误差处于所述第一阈值和所述第二阈值之间,则所述数据位宽保持不变。
A22.如条款A17所述的方法,所述量化误差的获取方法包括:
根据所述数据位宽确定量化间隔;
根据所述量化间隔、所述量化后的数据的个数与对应的量化前的数据确定量化误差。
A23.如条款A17所述的方法,所述量化误差的获取方法包括:
对量化后的数据进行反量化,获得反量化数据;其中,所述反量化数据的数据格式与对应的量化前的数据的数据格式相同;
根据所述量化后的数据以及对应的反量化数据确定量化误差。
A24.如条款A17所述的方法,所述量化前的数据是所述待量化数据。
A25.如条款A17所述的方法,所述量化前的数据是在目标迭代间隔内的权值更新迭代过程中涉及的待量化数据;其中,所述目标迭代间隔包括至少一次权值更新迭代,且同一目标迭代间隔内量化过程中采用相同的所述数据位宽。
A26.如条款A25所述的方法,所述目标迭代间隔的确定步骤包括:
在预判时间点,确定权值更新迭代过程中涉及的待量化数据的点位置参数的变化趋势值;其中,所述预判时间点是用于判断是否需要对所述数据位宽进行调整的时间点,所述预判时间点对应权值更新迭代完成时的时间点;
根据所述点位置参数的变化趋势值确定对应所述目标迭代间隔。
A27.如条款A25所述的方法,所述目标迭代间隔的确定步骤包括:
在预判时间点,确定权值更新迭代过程中涉及的待量化数据的点位置参数的变化趋势值、数据位宽的变化趋势值;其中,所述预判时间点是用于判断是否需要对所述数据位宽进行调整的时间点,所述预判时间点对应权值更新迭代完成时的时间点;
根据所述点位置参数的变化趋势值和所述数据位宽的变化趋势值确定对应所述目标迭代间隔。
A28.如条款A26或条款A27所述的方法,所述预判时间点包括第一预判时间点;其中,所述第一预判时间点是根据所述目标迭代间隔确定的。
A29.如条款A28所述的方法,所述预判时间点还包括第二预判时间点;其中,所述第二预判时间点是根据数据变动幅度曲线确定的;所述数据变动幅度曲线是对权值更新迭代过程中数据变动幅度情况进行统计获得的。
A30.如条款A26~条款A29任一项条款所述的方法,所述点位置参数的变化趋势值根据当前预判时间点对应的点位置参数的滑动平均值、上一预判时间点对应的点位置参数的滑动平均值确定。
A31.如条款A26~条款A29任一项条款所述的方法,所述点位置参数的变化趋势值根据当前预判时间点对应的点位置参数、上一预判时间点对应的点位置参数的滑动平均值确定。
A32.如条款A30所述的方法,所述当前预判时间点对应的点位置参数的滑动平均值的确定步骤包括:
根据上一预判时间点对应的点位置参数与所述数据位宽的调整值确定所述当前预判时间点对应的点位置参数;
根据所述数据位宽的调整值对所述上一预判时间点对应的点位置参数的滑动平均值进行调整,获得调整结果;
根据所述当前预判时间点对应的点位置参数、所述调整结果确定当前预判时间点对应的点位置参数的滑动平均值。
A33.如条款A30所述的方法,所述当前预判时间点对应的点位置参数的滑动平均值的 确定步骤包括:
根据上一预判时间点对应的点位置参数与上一预判时间点对应的点位置参数的滑动平均值确定当前预判时间点对应的点位置参数的滑动平均值的中间结果;
根据当前预判时间点对应的点位置参数的滑动平均值的中间结果与所述数据位宽的调整值确定所述当前预判时间点对应的点位置参数的滑动平均值。
A34.如条款A27所述的方法,所述数据位宽的变化趋势值根据对应所述量化误差确定。
A35.如条款A26~条款A29任一项条款所述的方法,所述目标迭代间隔内量化过程中采用的数据位宽的确定步骤包括:
确定对应量化误差;其中,所述量化误差对应的量化前的数据是所述预判时间点对应的权值更新迭代过程中涉及的待量化数据;
根据对应量化误差,确定所述目标迭代间隔内量化过程中采用的数据位宽。
A36.如条款A35所述的方法,确定所述目标迭代间隔内量化过程中采用的数据位宽的步骤包括:
所述量化误差与阈值进行比较,根据比较结果,对上一目标迭代间隔内量化过程中采用的数据位宽进行调整,调整结果作为当前目标迭代间隔内量化过程中采用的数据位宽。
A37.如条款A17所述的方法,所述量化前的数据是在目标迭代间隔内的权值更新迭代时涉及的待量化数据;其中,所述目标迭代间隔包括至少一次权值更新迭代,且同一目标迭代间隔内量化过程中采用相同的所述量化参数。
A38.如条款A37所述的方法,所述目标迭代间隔的确定步骤包括:
在预判时间点,确定权值更新迭代过程中涉及的待量化数据的点位置参数的变化趋势值;其中,所述预判时间点是用于判断是否需要对所述量化参数进行调整的时间点,所述预判时间点对应权值更新迭代完成时的时间点;
根据所述点位置参数的变化趋势值确定对应所述目标迭代间隔。
A39.如条款A10~条款A15任一条款所述的方法,所述点位置参数根据统计结果、和所述数据位宽确定。
B40.一种神经网络的量化参数确定装置,包括存储器及处理器,所述存储器上存储有 可在处理器上运行的计算机程序,所述处理器执行所述计算机程序时实现条款A1~条款A39中任一项所述方法的步骤。
C41.一种神经网络的量化参数确定设备,所述方法包括:
统计结果获取单元,用于获取每种待量化数据的统计结果;其中,所述待量化数据包括所述神经网络的神经元、权值、梯度、偏置中的至少一种数据;
量化参数确定单元,用于利用每种待量化数据的统计结果以及数据位宽确定对应量化参数;其中,所述量化参数用于人工智能处理器对神经网络运算过程中的数据进行对应量化。
C42.如条款C41所述的设备,所述神经网络的量化参数确定设备还包括:
第一量化单元,用于利用对应量化参数对所述待量化数据进行量化。
C43.如条款C41或条款C42所述的设备,所述神经网络的量化参数确定设备还包括:
第二量化单元,用于利用对应量化参数对目标数据进行量化;其中,所述目标数据的特征与所述待量化数据的特征之间具有相似性。
C44.如条款C41所述的设备,所述统计单元获得的统计结果为每种待量化数据中的最大值和最小值。
C45.如条款C41所述的设备,所述统计单元获得的统计结果为每种待量化数据中的绝对值最大值。
C46.如条款C45所述的设备,所述统计单元根据每种待量化数据中的最大值和最小值确定所述绝对值最大值。
C47.如条款C44所述的设备,所述量化参数确定单元根据每种待量化数据中的最大值、最小值以及所述数据位宽确定所述量化参数。
C48.如条款C45或条款C46所述的设备,所述量化参数确定单元根据每种待量化数据中的绝对值最大值、所述数据位宽确定所述量化参数。
C49.如条款C41所述的设备,所述量化参数确定单元确定的所述量化参数为点位置参数或第一缩放系数。
C50.如条款C49所述的设备,所述量化参数确定单元根据点位置参数和第二缩放系数 确定所述第一缩放系数;其中,确定第一缩放系数时使用的点位置参数为已知固定值,或所述点位置参数和对应的所述第二缩放系数相乘的结果整体作为第一缩放系数应用于神经网络运算过程中的数据量化。
C51.如条款C49所述的设备,所述量化参数确定单元确定的所述量化参数包括点位置参数和第二缩放系数。
C52.如条款C51所述的设备,所述量化参数确定单元根据所述点位置参数、所述统计结果、所述数据位宽确定所述第二缩放系数。
C53.如条款C49~条款C51任一项条款所述的设备,所述量化参数确定单元确定的所述量化参数还包括偏移量。
C54.如条款C53所述的设备,所述量化参数确定单元根据每种待量化数据的统计结果确定所述偏移量。
C55.如条款C41所述的设备,所述量化参数确定单元使用的所述数据位宽是预设值。
C56.如条款C41所述的设备,所述量化参数确定单元包括调整模块和量化误差确定模块;其中,
所述量化误差确定模块,用于根据量化后的数据与对应的量化前的数据确定所述量化误差;
所述调整模块,用于根据对应的量化误差进行调整所述数据位宽。
C57.如条款C56所述的设备,所述调整模块具体用于:
所述量化误差与阈值进行比较,根据比较结果,调整所述数据位宽;其中,所述阈值包括第一阈值和第二阈值中的至少一个。
C60.如条款C57所述的设备,所述调整模块包括第一调整子模块,其中,所述第一调整子模块用于:
所述量化误差大于等于所述第一阈值,则对所述数据位宽进行增加。
C61.如条款C57所述的设备,所述调整模块包括第二调整子模块,其中,所述第二调整子模块用于:
所述量化误差小于等于所述第二阈值,则对所述数据位宽进行减少。
C60.如条款C57所述的设备,所述调整模块包括第三调整子模块,其中,所述第三调整子模块用于:
所述量化误差处于所述第一阈值和所述第二阈值之间,则所述数据位宽保持不变。
C61.如条款C56所述的设备,所述量化误差确定模块包括:
量化间隔确定子模块,用于根据所述数据位宽确定量化间隔;
第一量化误差确定子模块,用于根据所述量化间隔、所述量化后的数据的个数与对应的量化前的数据确定量化误差。
C62.如条款C56所述的设备,所述量化误差确定模块包括:
反量化数据确定子模块,用于对量化后的数据进行反量化,获得反量化数据;其中,所述反量化数据的数据格式与对应的量化前的数据的数据格式相同;
第二量化误差确定子模块,用于根据所述量化后的数据以及对应的反量化数据确定量化误差。
C63.如条款C56所述的设备,所述量化误差确定模块使用的所述量化前的数据是所述待量化数据。
C64.如条款C56所述的设备,所述量化误差确定模块使用的所述量化前的数据是在目标迭代间隔内的权值更新迭代过程中涉及的待量化数据;其中,所述目标迭代间隔包括至少一次权值更新迭代,且同一目标迭代间隔内量化过程中采用相同的所述数据位宽。
C65.如条款C64所述的设备,所述神经网络的量化参数确定设备还包括第一目标迭代间隔确定单元;其中,所述第一目标迭代间隔确定单元包括:
第一变化趋势值确定模块,用于在预判时间点,确定权值更新迭代过程中涉及的待量化数据的点位置参数的变化趋势值;其中,所述预判时间点是用于判断是否需要对所述数据位宽进行调整的时间点,所述预判时间点对应权值更新迭代完成时的时间点;
第一目标迭代间隔模块,用于根据所述点位置参数的变化趋势值确定对应所述目标迭代间隔。
C66.如条款C64所述的设备,所述第一目标迭代间隔确定单元包括:
第二变化趋势值确定模块,用于在预判时间点,确定权值更新迭代过程中涉及的待量 化数据的点位置参数的变化趋势值、数据位宽的变化趋势值;其中,所述预判时间点是用于判断是否需要对所述数据位宽进行调整的时间点,所述预判时间点对应权值更新迭代完成时的时间点;
第二目标迭代间隔模块,用于根据所述点位置参数的变化趋势值和所述数据位宽的变化趋势值确定对应所述目标迭代间隔。
C67.如条款C65或条款C66所述的设备,所述第一目标迭代间隔确定单元还包括第一预判时间点确定单元;其中,
所述第一预判时间点确定单元,用于根据所述目标迭代间隔确定所述第一预判时间点。
C68.如条款C67所述的设备,所述第一目标迭代间隔确定单元还包括第二预判时间点确定单元;其中,所述第二预判时间点确定单元,用于根据数据变动幅度曲线确定第二预判时间点;其中,所述数据变动幅度曲线是对权值更新迭代过程中数据变动幅度情况进行统计获得的。
C69.如条款C65~条款C68任一项条款所述的设备,所述第一变化趋势值确定模块和所述第二变化趋势值确定模块均根据当前预判时间点对应的点位置参数的滑动平均值、上一预判时间点对应的点位置参数的滑动平均值确定所述点位置参数的变化趋势值。
C70.如条款C65~条款C68任一项条款所述的设备,所述第一变化趋势值确定模块和所述第二变化趋势值确定模块均根据当前预判时间点对应的点位置参数、上一预判时间点对应的点位置参数的滑动平均值确定所述点位置参数的变化趋势值。
C71.如条款C69所述的设备,所述第一变化趋势值确定模块和所述第二变化趋势值确定模块均包括:
当前预判时间点对应的点位置参数确定子模块,用于根据上一预判时间点对应的点位置参数与所述数据位宽的调整值确定所述当前预判时间点对应的点位置参数;
调整结果确定子模块,用于根据所述数据位宽的调整值对所述上一预判时间点对应的点位置参数的滑动平均值进行调整,获得调整结果;
第一滑动平均值确定子模块,用于根据所述当前预判时间点对应的点位置参数、所述调整结果确定当前预判时间点对应的点位置参数的滑动平均值。
C72.如条款C69所述的设备,所述第一变化趋势值确定模块和所述第二变化趋势值确定模块均包括:
中间结果确定子模块,用于根据上一预判时间点对应的点位置参数与上一预判时间点对应的点位置参数的滑动平均值确定当前预判时间点对应的点位置参数的滑动平均值的中间结果;
第二滑动平均值确定子模块,用于根据当前预判时间点对应的点位置参数的滑动平均值的中间结果与所述数据位宽的调整值确定所述当前预判时间点对应的点位置参数的滑动平均值。
C73.如条款C66所述的设备,所述第二变化趋势值确定模块根据对应所述量化误差确定数据位宽的变化趋势值。
C74.如条款C65~条款C68任一项条款所述的设备,所述第一目标迭代间隔确定单元还包括:
量化误差确定模块,用于确定对应量化误差;其中,所述量化误差对应的量化前的数据是所述预判时间点对应的权值更新迭代过程中涉及的待量化数据;
数据位宽确定模块,用于根据对应量化误差,确定所述目标迭代间隔内量化过程中采用的数据位宽。
C75.如条款C74所述的设备,所述数据位宽确定模块具体用于:
所述量化误差与阈值进行比较,根据比较结果,对上一目标迭代间隔内量化过程中采用的数据位宽进行调整,调整结果作为当前目标迭代间隔内量化过程中采用的数据位宽。
C76.如条款C56所述的设备,所述量化误差确定模块使用的所述量化前的数据是在目标迭代间隔内的权值更新迭代时涉及的待量化数据;其中,所述目标迭代间隔包括至少一次权值更新迭代,且同一目标迭代间隔内量化过程中采用相同的所述量化参数。
C77.如条款C76所述的设备,所述神经网络的量化参数确定设备还包括第二目标迭代间隔确定单元;其中,所述第二目标迭代间隔确定单元包括:
第三变化趋势值确定模块,用于在预判时间点,确定权值更新迭代过程中涉及的待量化数据的点位置参数的变化趋势值;其中,所述预判时间点是用于判断是否需要对所述量 化参数进行调整的时间点,所述预判时间点对应权值更新迭代完成时的时间点;
第三目标迭代间隔模块,用于根据所述点位置参数的变化趋势值确定对应所述目标迭代间隔。
C78.如条款C49~条款C54任一条款所述的设备,所述量化参数确定单元根据统计结果、和所述数据位宽确定所述点位置参数。
D79.一种计算机可读存储介质,所述计算机可读存储介质中存储有计算机程序,所述计算机程序被执行时,实现如条款A1~A39中任一项所述的方法的步骤。
以上已经描述了本披露的各实施例,上述说明是示例性的,并非穷尽性的,并且也不限于所披露的各实施例。在不偏离所说明的各实施例的范围和精神的情况下,对于本技术领域的普通技术人员来说许多修改和变更都是显而易见的。本文中所用术语的选择,旨在最好地解释各实施例的原理、实际应用或对市场中的技术的改进,或者使本技术领域的其它普通技术人员能理解本文披露的各实施例。

Claims (67)

  1. 一种神经网络的量化参数确定方法,其特征在于,所述方法包括:
    获取每种待量化数据的统计结果;其中,所述待量化数据包括所述神经网络的神经元、权值、梯度、偏置中的至少一种数据;
    利用每种待量化数据的统计结果以及数据位宽确定对应量化参数;其中,所述量化参数用于人工智能处理器对神经网络运算过程中的数据进行对应量化;所述量化参数为点位置参数。
  2. 如权利要求1所述的方法,其特征在于,所述方法还包括:
    利用对应量化参数对所述待量化数据进行量化。
  3. 如权利要求1或2所述的方法,其特征在于,所述方法还包括:
    利用对应量化参数对目标数据进行量化;其中,所述目标数据的特征与所述待量化数据的特征之间具有相似性。
  4. 如权利要求1所述的方法,其特征在于,所述神经网络运算过程包括神经网络训练、神经网络推理、神经网络微调中的至少一种运算。
  5. 如权利要求1所述的方法,其特征在于,所述统计结果为每种待量化数据中的最大值和最小值。
  6. 如权利要求1所述的方法,其特征在于,所述统计结果为每种待量化数据中的绝对值最大值。
  7. 如权利要求6所述的方法,其特征在于,所述绝对值最大值根据每种待量化数据中的最大值和最小值确定。
  8. 如权利要求5所述的方法,其特征在于,所述量化参数根据每种待量化数据中的最大值、最小值以及所述数据位宽确定。
  9. 如权利要求6或7所述的方法,其特征在于,所述量化参数根据每种待量化数据中的绝对值最大值、所述数据位宽确定。
  10. 如权利要求1所述的方法,其特征在于,所述数据位宽是预设值。
  11. 如权利要求1所述的方法,其特征在于,所述数据位宽根据对应的量化误差进行调整;其中,所述量化误差是根据量化后的数据与对应的量化前的数据确定。
  12. 如权利要求11所述的方法,其特征在于,所述数据位宽的调整步骤包括:
    所述量化误差与阈值进行比较,根据比较结果,调整所述数据位宽;其中,所述阈值包括第一阈值和第二阈值中的至少一个。
  13. 如权利要求12所述的方法,其特征在于,调整所述数据位宽的步骤包括:
    所述量化误差大于等于所述第一阈值,则对所述数据位宽进行增加。
  14. 如权利要求12所述的方法,其特征在于,调整所述数据位宽的步骤包括:
    所述量化误差小于等于所述第二阈值,则对所述数据位宽进行减少。
  15. 如权利要求12所述的方法,其特征在于,调整所述数据位宽的步骤包括:
    所述量化误差处于所述第一阈值和所述第二阈值之间,则所述数据位宽保持不变。
  16. 如权利要求11所述的方法,其特征在于,所述量化误差的获取方法包括:
    根据所述数据位宽确定量化间隔;
    根据所述量化间隔、所述量化后的数据的个数与对应的量化前的数据确定量化误差。
  17. 如权利要求11所述的方法,其特征在于,所述量化误差的获取方法包括:
    对量化后的数据进行反量化,获得反量化数据;其中,所述反量化数据的数据格式与对应的量化前的数据的数据格式相同;
    根据所述量化后的数据以及对应的反量化数据确定量化误差。
  18. 如权利要求11所述的方法,其特征在于,所述量化前的数据是所述待量化数据。
  19. 如权利要求11所述的方法,其特征在于,所述量化前的数据是在目标迭代间隔内的权值更新迭代过程中涉及的待量化数据;其中,所述目标迭代间隔包括至少一次权值更新迭代,且同一目标迭代间隔内量化过程中采用相同的所述数据位宽。
  20. 如权利要求19所述的方法,其特征在于,所述目标迭代间隔的确定步骤包括:
    在预判时间点,确定权值更新迭代过程中涉及的待量化数据的点位置参数的变化趋势值;其中,所述预判时间点是用于判断是否需要对所述数据位宽进行调整的时间点,所述预判时间点对应权值更新迭代完成时的时间点;
    根据所述点位置参数的变化趋势值确定对应所述目标迭代间隔。
  21. 如权利要求19所述的方法,其特征在于,所述目标迭代间隔的确定步骤包括:
    在预判时间点,确定权值更新迭代过程中涉及的待量化数据的点位置参数的变化趋势值、数据位宽的变化趋势值;其中,所述预判时间点是用于判断是否需要对所述数据位宽进行调整的时间点,所述预判时间点对应权值更新迭代完成时的时间点;
    根据所述点位置参数的变化趋势值和所述数据位宽的变化趋势值确定对应所述目标迭代间隔。
  22. 如权利要求20或21所述的方法,其特征在于,所述预判时间点包括第一预判时间点;其中,所述第一预判时间点是根据所述目标迭代间隔确定的。
  23. 如权利要求22所述的方法,其特征在于,所述预判时间点还包括第二预判时间点;其中,所述第二预判时间点是根据数据变动幅度曲线确定的;所述数据变动幅度曲线是对权值更新迭代过程中数据变动幅度情况进行统计获得的。
  24. 如权利要求20~23任一项权利要求所述的方法,其特征在于,所述点位置参数的变化趋势值根据当前预判时间点对应的点位置参数的滑动平均值、上一预判时间点对应的点位置参数的滑动平均值确定。
  25. 如权利要求20~23任一项权利要求所述的方法,其特征在于,所述点位置参数的变化趋势值根据当前预判时间点对应的点位置参数、上一预判时间点对应的点位置参数的滑动平均值确定。
  26. 如权利要求24所述的方法,其特征在于,所述当前预判时间点对应的点位置参数的滑动平均值的确定步骤包括:
    根据上一预判时间点对应的点位置参数与所述数据位宽的调整值确定所述当前预判时间点对应的点位置参数;
    根据所述数据位宽的调整值对所述上一预判时间点对应的点位置参数的滑动平均值进行调整,获得调整结果;
    根据所述当前预判时间点对应的点位置参数、所述调整结果确定当前预判时间点对应的点位置参数的滑动平均值。
  27. 如权利要求24所述的方法,其特征在于,所述当前预判时间点对应的点位置参数的滑动平均值的确定步骤包括:
    根据上一预判时间点对应的点位置参数与上一预判时间点对应的点位置参数的滑动平均值确定当前预判时间点对应的点位置参数的滑动平均值的中间结果;
    根据当前预判时间点对应的点位置参数的滑动平均值的中间结果与所述数据位宽的调整值确定所述当前预判时间点对应的点位置参数的滑动平均值。
  28. 如权利要求21所述的方法,其特征在于,所述数据位宽的变化趋势值根据对应所述量化误差确定。
  29. 如权利要求20~23任一项权利要求所述的方法,其特征在于,所述目标迭代间隔内量化过程中采用的数据位宽的确定步骤包括:
    确定对应量化误差;其中,所述量化误差对应的量化前的数据是所述预判时间点对应的权值更新迭代过程中涉及的待量化数据;
    根据对应量化误差,确定所述目标迭代间隔内量化过程中采用的数据位宽。
  30. 如权利要求29所述的方法,其特征在于,确定所述目标迭代间隔内量化过程中采用的数据位宽的步骤包括:
    所述量化误差与阈值进行比较,根据比较结果,对上一目标迭代间隔内量化过程中采用的数据位宽进行调整,调整结果作为当前目标迭代间隔内量化过程中采用的数据位宽。
  31. 如权利要求11所述的方法,其特征在于,所述量化前的数据是在目标迭代间隔内的权值更新迭代时涉及的待量化数据;其中,所述目标迭代间隔包括至少一次权值更新迭代,且同一目标迭代间隔内量化过程中采用相同的所述量化参数。
  32. 如权利要求31所述的方法,其特征在于,所述目标迭代间隔的确定步骤包括:
    在预判时间点,确定权值更新迭代过程中涉及的待量化数据的点位置参数的变化趋势值;其中,所述预判时间点是用于判断是否需要对所述量化参数进行调整的时间点,所述预判时间点对应权值更新迭代完成时的时间点;
    根据所述点位置参数的变化趋势值确定对应所述目标迭代间隔。
  33. 如权利要求1所述的方法,其特征在于,所述点位置参数根据统计结果、和所述数据位宽确定。
  34. 一种神经网络的量化参数确定装置,包括存储器及处理器,所述存储器上存储有可 在处理器上运行的计算机程序,所述处理器执行所述计算机程序时实现1~33中任一项所述方法的步骤。
  35. 一种神经网络的量化参数确定设备,所述设备包括:
    统计结果获取单元,用于获取每种待量化数据的统计结果;其中,所述待量化数据包括所述神经网络的神经元、权值、梯度、偏置中的至少一种数据;
    量化参数确定单元,用于利用每种待量化数据的统计结果以及数据位宽确定对应量化参数;其中,所述量化参数用于人工智能处理器对神经网络运算过程中的数据进行对应量化;所述量化参数为点位置参数。
  36. 如权利要求35所述的设备,其特征在于,所述神经网络的量化参数确定设备还包括:
    第一量化单元,用于利用对应量化参数对所述待量化数据进行量化。
  37. 如权利要求35或36所述的设备,其特征在于,所述神经网络的量化参数确定设备还包括:
    第二量化单元,用于利用对应量化参数对目标数据进行量化;其中,所述目标数据的特征与所述待量化数据的特征之间具有相似性。
  38. 如权利要求35所述的设备,其特征在于,所述统计单元获得的统计结果为每种待量化数据中的最大值和最小值。
  39. 如权利要求35所述的设备,其特征在于,所述统计单元获得的统计结果为每种待量化数据中的绝对值最大值。
  40. 如权利要求39所述的设备,其特征在于,所述统计单元根据每种待量化数据中的最大值和最小值确定所述绝对值最大值。
  41. 如权利要求39所述的设备,其特征在于,所述量化参数确定单元根据每种待量化数据中的最大值、最小值以及所述数据位宽确定所述量化参数。
  42. 如权利要求39或40所述的设备,其特征在于,所述量化参数确定单元根据每种待量化数据中的绝对值最大值、所述数据位宽确定所述量化参数。
  43. 如权利要求35所述的设备,其特征在于,所述量化参数确定单元使用的所述数据 位宽是预设值。
  44. 如权利要求35所述的设备,其特征在于,所述量化参数确定单元包括调整模块和量化误差确定模块;其中,
    所述量化误差确定模块,用于根据量化后的数据与对应的量化前的数据确定所述量化误差;
    所述调整模块,用于根据对应的量化误差调整所述数据位宽。
  45. 如权利要求44所述的设备,其特征在于,所述调整模块具体用于:
    所述量化误差与阈值进行比较,根据比较结果,调整所述数据位宽;其中,所述阈值包括第一阈值和第二阈值中的至少一个。
  46. 如权利要求45所述的设备,其特征在于,所述调整模块包括第一调整子模块,其中,所述第一调整子模块用于:
    所述量化误差大于等于所述第一阈值,则对所述数据位宽进行增加。
  47. 如权利要求45所述的设备,其特征在于,所述调整模块包括第二调整子模块,其中,所述第二调整子模块用于:
    所述量化误差小于等于所述第二阈值,则对所述数据位宽进行减少。
  48. 如权利要求45所述的设备,其特征在于,所述调整模块包括第三调整子模块,其中,所述第三调整子模块用于:
    所述量化误差处于所述第一阈值和所述第二阈值之间,则所述数据位宽保持不变。
  49. 如权利要求44所述的设备,其特征在于,所述量化误差确定模块包括:
    量化间隔确定子模块,用于根据所述数据位宽确定量化间隔;
    第一量化误差确定子模块,用于根据所述量化间隔、所述量化后的数据的个数与对应的量化前的数据确定量化误差。
  50. 如权利要求44所述的设备,其特征在于,所述量化误差确定模块包括:
    反量化数据确定子模块,用于对量化后的数据进行反量化,获得反量化数据;其中,所述反量化数据的数据格式与对应的量化前的数据的数据格式相同;
    第二量化误差确定子模块,用于根据所述量化后的数据以及对应的反量化数据确定量 化误差。
  51. 如权利要求44所述的设备,其特征在于,所述量化误差确定模块使用的所述量化前的数据是所述待量化数据。
  52. 如权利要求44所述的设备,其特征在于,所述量化误差确定模块使用的所述量化前的数据是在目标迭代间隔内的权值更新迭代过程中涉及的待量化数据;其中,所述目标迭代间隔包括至少一次权值更新迭代,且同一目标迭代间隔内量化过程中采用相同的所述数据位宽。
  53. 如权利要求52所述的设备,其特征在于,所述神经网络的量化参数确定设备还包括第一目标迭代间隔确定单元;其中,所述第一目标迭代间隔确定单元包括:
    第一变化趋势值确定模块,用于在预判时间点,确定权值更新迭代过程中涉及的待量化数据的点位置参数的变化趋势值;其中,所述预判时间点是用于判断是否需要对所述数据位宽进行调整的时间点,所述预判时间点对应权值更新迭代完成时的时间点;
    第一目标迭代间隔模块,用于根据所述点位置参数的变化趋势值确定对应所述目标迭代间隔。
  54. 如权利要求52所述的设备,其特征在于,所述第一目标迭代间隔确定单元包括:
    第二变化趋势值确定模块,用于在预判时间点,确定权值更新迭代过程中涉及的待量化数据的点位置参数的变化趋势值、数据位宽的变化趋势值;其中,所述预判时间点是用于判断是否需要对所述数据位宽进行调整的时间点,所述预判时间点对应权值更新迭代完成时的时间点;
    第二目标迭代间隔模块,用于根据所述点位置参数的变化趋势值和所述数据位宽的变化趋势值确定对应所述目标迭代间隔。
  55. 如权利要求53或54所述的设备,其特征在于,所述第一目标迭代间隔确定单元还包括第一预判时间点确定单元;其中,
    所述第一预判时间点确定单元,用于根据所述目标迭代间隔确定所述第一预判时间点。
  56. 如权利要求55所述的设备,其特征在于,所述第一目标迭代间隔确定单元还包括第二预判时间点确定单元;其中,所述第二预判时间点确定单元,用于根据数据变动幅度 曲线确定第二预判时间点;其中,所述数据变动幅度曲线是对权值更新迭代过程中数据变动幅度情况进行统计获得的。
  57. 如权利要求53~56任一项权利要求所述的设备,其特征在于,所述第一变化趋势值确定模块和所述第二变化趋势值确定模块均根据当前预判时间点对应的点位置参数的滑动平均值、上一预判时间点对应的点位置参数的滑动平均值确定所述点位置参数的变化趋势值。
  58. 如权利要求53~56任一项权利要求所述的设备,其特征在于,所述第一变化趋势值确定模块和所述第二变化趋势值确定模块均根据当前预判时间点对应的点位置参数、上一预判时间点对应的点位置参数的滑动平均值确定所述点位置参数的变化趋势值。
  59. 如权利要求57所述的设备,其特征在于,所述第一变化趋势值确定模块和所述第二变化趋势值确定模块均包括:
    当前预判时间点对应的点位置参数确定子模块,用于根据上一预判时间点对应的点位置参数与所述数据位宽的调整值确定所述当前预判时间点对应的点位置参数;
    调整结果确定子模块,用于根据所述数据位宽的调整值对所述上一预判时间点对应的点位置参数的滑动平均值进行调整,获得调整结果;
    第一滑动平均值确定子模块,用于根据所述当前预判时间点对应的点位置参数、所述调整结果确定当前预判时间点对应的点位置参数的滑动平均值。
  60. 如权利要求57所述的设备,其特征在于,所述第一变化趋势值确定模块和所述第二变化趋势值确定模块均包括:
    中间结果确定子模块,用于根据上一预判时间点对应的点位置参数与上一预判时间点对应的点位置参数的滑动平均值确定当前预判时间点对应的点位置参数的滑动平均值的中间结果;
    第二滑动平均值确定子模块,用于根据当前预判时间点对应的点位置参数的滑动平均值的中间结果与所述数据位宽的调整值确定所述当前预判时间点对应的点位置参数的滑动平均值。
  61. 如权利要求54所述的设备,其特征在于,所述第二变化趋势值确定模块根据对应 所述量化误差确定数据位宽的变化趋势值。
  62. 如权利要求53~56任一项权利要求所述的设备,其特征在于,所述第一目标迭代间隔确定单元还包括:
    量化误差确定模块,用于确定对应量化误差;其中,所述量化误差对应的量化前的数据是所述预判时间点对应的权值更新迭代过程中涉及的待量化数据;
    数据位宽确定模块,用于根据对应量化误差,确定所述目标迭代间隔内量化过程中采用的数据位宽。
  63. 如权利要求62所述的设备,其特征在于,所述数据位宽确定模块具体用于:
    所述量化误差与阈值进行比较,根据比较结果,对上一目标迭代间隔内量化过程中采用的数据位宽进行调整,调整结果作为当前目标迭代间隔内量化过程中采用的数据位宽。
  64. 如权利要求44所述的设备,其特征在于,所述量化误差确定模块使用的所述量化前的数据是在目标迭代间隔内的权值更新迭代时涉及的待量化数据;其中,所述目标迭代间隔包括至少一次权值更新迭代,且同一目标迭代间隔内量化过程中采用相同的所述量化参数。
  65. 如权利要求64所述的设备,其特征在于,所述神经网络的量化参数确定设备还包括第二目标迭代间隔确定单元;其中,所述第二目标迭代间隔确定单元包括:
    第三变化趋势值确定模块,用于在预判时间点,确定权值更新迭代过程中涉及的待量化数据的点位置参数的变化趋势值;其中,所述预判时间点是用于判断是否需要对所述量化参数进行调整的时间点,所述预判时间点对应权值更新迭代完成时的时间点;
    第三目标迭代间隔模块,用于根据所述点位置参数的变化趋势值确定对应所述目标迭代间隔。
  66. 如权利要求35所述的设备,其特征在于,所述量化参数确定单元根据统计结果、和所述数据位宽确定所述点位置参数。
  67. 一种计算机可读存储介质,所述计算机可读存储介质中存储有计算机程序,其特征在于,所述计算机程序被执行时,实现如权利要求1~33中任一项所述的方法的步骤。
PCT/CN2019/106754 2019-06-12 2019-09-19 一种神经网络的量化参数确定方法及相关产品 WO2020248423A1 (zh)

Priority Applications (4)

Application Number Priority Date Filing Date Title
EP19786896.1A EP3770823A4 (en) 2019-06-12 2019-09-19 QUANTIFICATION PARAMETER DETERMINATION PROCESS FOR NEURONAL NETWORK, AND RELATED PRODUCT
US16/622,541 US11675676B2 (en) 2019-06-12 2019-09-19 Neural network quantization parameter determination method and related products
US16/720,113 US11676029B2 (en) 2019-06-12 2019-12-19 Neural network quantization parameter determination method and related products
US16/720,093 US11676028B2 (en) 2019-06-12 2019-12-19 Neural network quantization parameter determination method and related products

Applications Claiming Priority (8)

Application Number Priority Date Filing Date Title
CN201910505239.7 2019-06-12
CN201910505239 2019-06-12
CN201910515355 2019-06-14
CN201910515355.7 2019-06-14
CN201910528537 2019-06-18
CN201910528537.8 2019-06-18
CN201910570125 2019-06-27
CN201910570125.0 2019-06-27

Related Child Applications (3)

Application Number Title Priority Date Filing Date
US16/622,541 A-371-Of-International US11675676B2 (en) 2019-06-12 2019-09-19 Neural network quantization parameter determination method and related products
US16/720,093 Continuation US11676028B2 (en) 2019-06-12 2019-12-19 Neural network quantization parameter determination method and related products
US16/720,113 Continuation US11676029B2 (en) 2019-06-12 2019-12-19 Neural network quantization parameter determination method and related products

Publications (1)

Publication Number Publication Date
WO2020248423A1 true WO2020248423A1 (zh) 2020-12-17

Family

ID=69185300

Family Applications (2)

Application Number Title Priority Date Filing Date
PCT/CN2019/106801 WO2020248424A1 (zh) 2019-06-12 2019-09-19 一种神经网络的量化参数确定方法及相关产品
PCT/CN2019/106754 WO2020248423A1 (zh) 2019-06-12 2019-09-19 一种神经网络的量化参数确定方法及相关产品

Family Applications Before (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/106801 WO2020248424A1 (zh) 2019-06-12 2019-09-19 一种神经网络的量化参数确定方法及相关产品

Country Status (6)

Country Link
US (2) US11675676B2 (zh)
EP (4) EP3998554A4 (zh)
JP (3) JP2021530769A (zh)
KR (3) KR102656332B1 (zh)
CN (14) CN112085193B (zh)
WO (2) WO2020248424A1 (zh)

Families Citing this family (27)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US11437032B2 (en) 2017-09-29 2022-09-06 Shanghai Cambricon Information Technology Co., Ltd Image processing apparatus and method
US11169803B2 (en) 2018-02-13 2021-11-09 Shanghai Cambricon Information Technology Co., Ltd. Computing device and method
CN116991225A (zh) 2018-02-14 2023-11-03 上海寒武纪信息科技有限公司 处理器的控制装置、方法及设备
EP3798850A4 (en) 2018-06-27 2022-03-23 Shanghai Cambricon Information Technology Co., Ltd ON-CHIP CODE BREAKPOINT DEBUG METHOD, ON-CHIP PROCESSOR AND CHIP BREAKPOINT DEBUG SYSTEM
US11507823B2 (en) * 2019-01-22 2022-11-22 Black Sesame Technologies Inc. Adaptive quantization and mixed precision in a network
US11966818B2 (en) 2019-02-21 2024-04-23 Hewlett Packard Enterprise Development Lp System and method for self-healing in decentralized model building for machine learning using blockchain
US11676028B2 (en) 2019-06-12 2023-06-13 Shanghai Cambricon Information Technology Co., Ltd Neural network quantization parameter determination method and related products
CN110490309B (zh) * 2019-08-14 2022-06-07 中科寒武纪科技股份有限公司 一种用于神经网络的算子融合方法及其相关产品
JP7146952B2 (ja) 2019-08-23 2022-10-04 安徽寒武紀信息科技有限公司 データ処理方法、装置、コンピュータデバイス、及び記憶媒体
CN112487182B (zh) * 2019-09-12 2024-04-12 华为技术有限公司 文本处理模型的训练方法、文本处理方法及装置
WO2021056180A1 (en) * 2019-09-24 2021-04-01 Baidu.Com Times Technology (Beijing) Co., Ltd. Cursor-based adaptive quantization for deep neural networks
JP7354736B2 (ja) * 2019-09-30 2023-10-03 富士通株式会社 情報処理装置、情報処理方法、情報処理プログラム
US11775611B2 (en) * 2019-11-01 2023-10-03 Samsung Electronics Co., Ltd. Piecewise quantization for neural networks
JP2021111081A (ja) * 2020-01-09 2021-08-02 富士通株式会社 情報処理装置、ニューラルネットワークの演算プログラム及びニューラルネットワークの演算方法
US20210241183A1 (en) * 2020-01-31 2021-08-05 Hewlett Packard Enterprise Development Lp Adaptively synchronizing learning of multiple learning models
CN113741619B (zh) * 2020-05-27 2024-03-12 安徽寒武纪信息科技有限公司 时钟控制装置及相关产品
KR20220033713A (ko) * 2020-09-10 2022-03-17 에스케이하이닉스 주식회사 데이터 처리 시스템 및 그 동작 방법
CN112686001B (zh) * 2021-01-05 2021-12-03 中科三清科技有限公司 气象数据的变换方法、传输方法、服务器及数据传输系统
KR102507461B1 (ko) * 2021-02-16 2023-03-07 고려대학교 산학협력단 레이어-단위 양자화 신경망을 위한 인-메모리 가속기 및 이의 동작 방법
CN112990457B (zh) * 2021-03-26 2024-05-03 开放智能机器(上海)有限公司 离线量化调优方法、装置、设备、介质及程序产品
CN113220606B (zh) * 2021-05-07 2021-11-26 珠海市芯动力科技有限公司 神经网络权值存储方法、读取方法及相关设备
JP2023069780A (ja) * 2021-11-08 2023-05-18 富士通株式会社 演算プログラム、演算方法及び計算機
WO2023128024A1 (ko) * 2021-12-30 2023-07-06 한국전자기술연구원 딥러닝 네트워크 양자화 처리 방법 및 시스템
KR20230136572A (ko) * 2022-03-18 2023-09-26 인텔렉추얼디스커버리 주식회사 신경망 기반 특징 텐서 압축 방법 및 장치
CN114611697B (zh) * 2022-05-11 2022-09-09 上海登临科技有限公司 神经网络量化及部署方法、系统、电子设备及存储介质
CN117787358A (zh) * 2024-01-16 2024-03-29 之江实验室 一种基于阻变存储器的模型量化方法、装置以及设备
CN117910421A (zh) * 2024-03-15 2024-04-19 南京美辰微电子有限公司 一种基于神经网络的动态近似电路计算部署方法及系统

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US9916531B1 (en) * 2017-06-22 2018-03-13 Intel Corporation Accumulator constrained quantization of convolutional neural networks
CN109146057A (zh) * 2018-06-26 2019-01-04 杭州雄迈集成电路技术有限公司 一种基于查表计算的高精度的神经网络工程化方法
CN109214509A (zh) * 2017-07-05 2019-01-15 中国科学院沈阳自动化研究所 一种用于深度神经网络高速实时量化结构和运算实现方法
CN109389219A (zh) * 2017-08-04 2019-02-26 三星电子株式会社 用于对神经网络的参数进行量化的方法和装置
CN109472353A (zh) * 2018-11-22 2019-03-15 济南浪潮高新科技投资发展有限公司 一种卷积神经网络量化电路及量化方法
CN109800877A (zh) * 2019-02-20 2019-05-24 腾讯科技(深圳)有限公司 神经网络的参数调整方法、装置及设备

Family Cites Families (243)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2604712B2 (ja) * 1985-10-18 1997-04-30 ソニー株式会社 テレビジヨン信号の高能率符号化/復号装置
JPS63111768A (ja) * 1986-10-30 1988-05-17 Nec Corp 画像デ−タ量子化装置
JPH0375860A (ja) 1989-08-18 1991-03-29 Hitachi Ltd パーソナライズド端末
US5052043A (en) 1990-05-07 1991-09-24 Eastman Kodak Company Neural network with back propagation controlled through an output confidence measure
EP0509576B1 (en) * 1991-04-18 1998-01-28 Ampex Systems Corporation Method and apparatus for determining a quantizing factor for processes involving multiple compression/decompression of data
US6144977A (en) 1995-07-10 2000-11-07 Motorola, Inc. Circuit and method of converting a floating point number to a programmable fixed point number
GB9602701D0 (en) 1996-02-09 1996-04-10 Canon Kk Image manipulation
JPH10233691A (ja) * 1998-03-30 1998-09-02 Nec Corp 符号化方式および復号方式
US7242414B1 (en) 1999-07-30 2007-07-10 Mips Technologies, Inc. Processor having a compare extension of an instruction set architecture
JP2000293371A (ja) 1999-04-09 2000-10-20 Hitachi Ltd マイクロプログラム制御方法及び装置
US6671796B1 (en) 2000-02-25 2003-12-30 Sun Microsystems, Inc. Converting an arbitrary fixed point value to a floating point value
US6931639B1 (en) 2000-08-24 2005-08-16 International Business Machines Corporation Method for implementing a variable-partitioned queue for simultaneous multithreaded processors
ATE354688T1 (de) 2000-09-07 2007-03-15 Nippon Steel Corp Sechswertiges chrom-freies oberflächenbehandlungsmittel für auf sn- oder al- basis beschichtetes stahlblech, und oberflächenbehandltes stahlblech
US7062445B2 (en) * 2001-01-26 2006-06-13 Microsoft Corporation Quantization loop with heuristic approach
US20020138714A1 (en) 2001-03-22 2002-09-26 Sun Microsystems, Inc. Scoreboard for scheduling of instructions in a microprocessor that provides out of order execution
WO2002086817A1 (en) 2001-04-19 2002-10-31 Telefonaktiebolaget Lm Ericsson (Publ) Adaptive memory allocation
US20030167460A1 (en) 2002-02-26 2003-09-04 Desai Vipul Anil Processor instruction set simulation power estimation method
JP4148356B2 (ja) * 2002-11-18 2008-09-10 学校法人東海大学 量子化ステップパラメータ決定装置と量子化ステップパラメータ決定方法と量子化ステップパラメータ決定プログラム、ならびに非線形量子化方法と非線形量子化装置と非線形量子化プログラム
US7236995B2 (en) 2002-12-27 2007-06-26 Arm Limited Data processing apparatus and method for converting a number between fixed-point and floating-point representations
DE10316381A1 (de) 2003-04-10 2004-10-28 Bayer Technology Services Gmbh Verfahren zum Training von neuronalen Netzen
JP3889738B2 (ja) * 2003-09-26 2007-03-07 三洋電機株式会社 逆量子化装置、オーディオ復号化装置、画像復号化装置、逆量子化方法および逆量子化プログラム
JP4202244B2 (ja) 2003-12-22 2008-12-24 Necエレクトロニクス株式会社 Vliw型dsp,及びその動作方法
US20060161375A1 (en) 2004-12-30 2006-07-20 Allen Duberstein Optimizing processing speed based on measured temperatures
KR100762591B1 (ko) * 2005-09-29 2007-10-01 엘지전자 주식회사 비디오 코덱의 양자화 파라미터 결정방법
US7721128B2 (en) 2005-11-29 2010-05-18 International Business Machines Corporation Implementation of thermal throttling logic
WO2007116551A1 (ja) * 2006-03-30 2007-10-18 Kabushiki Kaisha Toshiba 画像符号化装置及び画像符号化方法並びに画像復号化装置及び画像復号化方法
CN1851668A (zh) 2006-06-01 2006-10-25 北京天碁科技有限公司 片上系统芯片、片上系统芯片的跟踪调试系统及方法
JP5224666B2 (ja) * 2006-09-08 2013-07-03 株式会社東芝 オーディオ符号化装置
DE102006059156B4 (de) 2006-12-14 2008-11-06 Advanced Micro Devices, Inc., Sunnyvale Verfahren zum Testen eines integrierten Schaltkreischips mit zumindest zwei Schaltungskernen sowie integrierter Schaltkreischip und Testsystem
US20110060587A1 (en) 2007-03-07 2011-03-10 Phillips Michael S Command and control utilizing ancillary information in a mobile voice-to-speech application
US8560591B2 (en) 2007-04-25 2013-10-15 International Business Machines Corporation Detection of potential need to use a larger data format in performing floating point operations
US8051118B2 (en) 2007-04-26 2011-11-01 International Business Machines Corporation Composition of decimal floating point data
US8051117B2 (en) 2007-04-26 2011-11-01 International Business Machines Corporation Shift significand of decimal floating point data
US8190664B2 (en) 2007-04-26 2012-05-29 International Business Machines Corporation Employing a mask field of an instruction to encode a sign of a result of the instruction
JP5184824B2 (ja) 2007-06-15 2013-04-17 キヤノン株式会社 演算処理装置及び方法
JP2009110353A (ja) 2007-10-31 2009-05-21 Hitachi Ltd マイクロコントローラ及び制御システム
US7904287B2 (en) 2007-11-13 2011-03-08 International Business Machines Corporation Method and system for real-time prediction of power usage for a change to another performance state
JP4998794B2 (ja) 2007-11-29 2012-08-15 Nkワークス株式会社 画像補正方法と画像補正装置
KR101518237B1 (ko) * 2008-09-01 2015-05-15 삼성전자주식회사 영상의 역양자화 방법 및 장치, 복호화 방법 및 장치
US20100073068A1 (en) 2008-09-22 2010-03-25 Hanwoo Cho Functional block level thermal control
CN101754490B (zh) * 2008-12-17 2012-11-07 电信科学技术研究院 一种数据传输的方法、系统和装置
CN101572829B (zh) 2009-06-10 2011-02-02 中国联合网络通信集团有限公司 Iptv视频质量监测方法、装置和系统
EP2336882A1 (en) 2009-12-18 2011-06-22 Telefonaktiebolaget L M Ericsson (PUBL) Technique for run-time provision of executable code using off-device services
US8478507B2 (en) 2010-04-21 2013-07-02 Toyota Jidosha Kabushiki Kaisha Control device for internal combustion engine
JP2011253374A (ja) 2010-06-02 2011-12-15 Sony Corp 情報処理装置、および情報処理方法、並びにプログラム
US8452463B2 (en) 2010-06-04 2013-05-28 Apple Inc. Adjusting the thermal behavior of a computing system using indirect information about ambient temperature
US8694572B2 (en) 2010-07-06 2014-04-08 Silminds, Llc, Egypt Decimal floating-point fused multiply-add unit
CN102622207B (zh) * 2011-01-30 2015-07-22 中兴通讯股份有限公司 定点化处理方法及装置
US8924455B1 (en) 2011-02-25 2014-12-30 Xilinx, Inc. Multiplication of matrices using systolic arrays
WO2012117955A1 (ja) * 2011-02-28 2012-09-07 株式会社メガチップス 画像符号化装置
CN102761509B (zh) 2011-04-27 2016-01-06 联芯科技有限公司 Ofdm系统的接收系统及降低接收系统内存的方法
CN103534664B (zh) 2011-05-12 2016-08-31 苹果公司 存在感测
CN102789413B (zh) 2011-05-23 2016-02-17 同济大学 一种并行程序的调试系统及方法
US8594982B2 (en) 2011-06-09 2013-11-26 Pulsar Informatics, Inc. Systems and methods for distributed calculation of fatigue-risk prediction and optimization
CN102291773B (zh) * 2011-07-18 2014-12-10 电信科学技术研究院 一种数据压缩方法和设备
CN102404673B (zh) 2011-11-24 2013-12-18 苏州上声电子有限公司 数字化扬声器系统通道均衡与声场控制方法和装置
CN103152673B (zh) 2011-12-07 2015-07-08 中国科学院声学研究所 基于四元码动态失配整形的数字扬声器驱动方法和装置
CN102684701B (zh) 2012-04-27 2014-07-09 苏州上声电子有限公司 基于编码转换的数字扬声器驱动方法和装置
DE102012009502A1 (de) 2012-05-14 2013-11-14 Kisters Ag Verfahren zum Trainieren eines künstlichen neuronalen Netzes
US9417891B2 (en) 2012-06-11 2016-08-16 Vmware, Inc. Unified storage/VDI provisioning methodology
US9224089B2 (en) * 2012-08-07 2015-12-29 Qualcomm Incorporated Method and apparatus for adaptive bit-allocation in neural systems
US9063731B2 (en) 2012-08-27 2015-06-23 Samsung Electronics Co., Ltd. Ultra low power apparatus and method to wake up a main processor
CN102903089B (zh) 2012-09-07 2014-12-17 山东大学 一种Linux环境下生成遥感图像快视图的方法
US9412366B2 (en) 2012-09-18 2016-08-09 Adobe Systems Incorporated Natural language image spatial and tonal localization
JP5913059B2 (ja) * 2012-11-13 2016-04-27 日本電信電話株式会社 分散型無線通信基地局システム、信号処理装置、無線装置、及び分散型無線通信基地局システムの動作方法
CN102981854A (zh) 2012-11-16 2013-03-20 天津市天祥世联网络科技有限公司 基于浮点数运算内联函数库的神经网络优化方法
US20150369963A1 (en) 2012-11-22 2015-12-24 Keio University Acrylic copolymer, optical film, polarizing plate and liquid crystal display device
US9851977B2 (en) 2012-12-06 2017-12-26 Kalray Apparatus and method for combining thread warps with compatible execution masks for simultaneous execution and increased lane utilization
US9720732B1 (en) 2013-02-11 2017-08-01 Amazon Technologies, Inc. Parameter selection for optimization of task execution based on execution history for prior tasks
JP2014170295A (ja) 2013-03-01 2014-09-18 Honda Motor Co Ltd 物体認識システム及び物体認識方法
US20190138372A1 (en) 2013-04-29 2019-05-09 Moogsoft, Inc. System for managing an instructure with security
US20150063461A1 (en) * 2013-08-27 2015-03-05 Magnum Semiconductor, Inc. Methods and apparatuses for adjusting macroblock quantization parameters to improve visual quality for lossy video encoding
JP6184891B2 (ja) 2014-03-12 2017-08-23 東芝メモリ株式会社 情報処理装置、半導体チップ、情報処理方法およびプログラム
CN105100810B (zh) * 2014-05-16 2018-02-13 中国科学院声学研究所 一种成像声纳实时处理系统中的图像压缩解压方法及系统
US9507405B2 (en) 2014-06-18 2016-11-29 Oracle International Corporation System and method for managing power in a chip multiprocessor using a proportional feedback mechanism
US10318882B2 (en) * 2014-09-11 2019-06-11 Amazon Technologies, Inc. Optimized training of linear machine learning models
US9575537B2 (en) 2014-07-25 2017-02-21 Intel Corporation Adaptive algorithm for thermal throttling of multi-core processors with non-homogeneous performance states
US10282100B2 (en) 2014-08-19 2019-05-07 Samsung Electronics Co., Ltd. Data management scheme in virtualized hyperscale environments
GB2524126B (en) 2014-08-28 2016-07-27 Imagination Tech Ltd Combining paths
US9916130B2 (en) 2014-11-03 2018-03-13 Arm Limited Apparatus and method for vector processing
FR3030077B1 (fr) 2014-12-10 2016-12-02 Arnault Ioualalen Procede d'ajustement de la precision d'un programme d'ordinateur manipulant au moins un nombre a virgule.
EP3035204B1 (en) 2014-12-19 2018-08-15 Intel Corporation Storage device and method for performing convolution operations
US20170061279A1 (en) 2015-01-14 2017-03-02 Intel Corporation Updating an artificial neural network using flexible fixed point representation
US10373050B2 (en) * 2015-05-08 2019-08-06 Qualcomm Incorporated Fixed point neural network based on floating point neural network quantization
US20160328645A1 (en) 2015-05-08 2016-11-10 Qualcomm Incorporated Reduced computational complexity for fixed point neural network
US10262259B2 (en) * 2015-05-08 2019-04-16 Qualcomm Incorporated Bit width selection for fixed point neural networks
US10083395B2 (en) 2015-05-21 2018-09-25 Google Llc Batch processing in a neural network processor
CN104899641B (zh) 2015-05-25 2018-07-13 杭州朗和科技有限公司 深度神经网络学习方法、处理器和深度神经网络学习系统
CN115100017A (zh) 2015-06-10 2022-09-23 无比视视觉技术有限公司 用于处理图像的图像处理器和方法
CN104978303B (zh) 2015-06-19 2019-06-04 上海兆芯集成电路有限公司 单芯片整合的传感器集线器和多传感器管理方法
CN106469291A (zh) 2015-08-19 2017-03-01 中兴通讯股份有限公司 图像处理方法及终端
US10970617B2 (en) * 2015-08-21 2021-04-06 Institute Of Automation Chinese Academy Of Sciences Deep convolutional neural network acceleration and compression method based on parameter quantification
US10031765B2 (en) 2015-09-24 2018-07-24 Intel Corporation Instruction and logic for programmable fabric hierarchy and cache
EP3353745A1 (en) 2015-09-30 2018-08-01 Piksel, Inc. Improved video stream delivery via adaptive quality enhancement using error correction models
US11061672B2 (en) 2015-10-02 2021-07-13 Via Alliance Semiconductor Co., Ltd. Chained split execution of fused compound arithmetic operations
CN106570559A (zh) 2015-10-09 2017-04-19 阿里巴巴集团控股有限公司 一种基于神经网络的数据处理方法和装置
CN105488565A (zh) * 2015-11-17 2016-04-13 中国科学院计算技术研究所 加速深度神经网络算法的加速芯片的运算装置及方法
US9930248B2 (en) 2015-11-17 2018-03-27 Eman Bayani Digital image capturing device system and method
CN106814639A (zh) 2015-11-27 2017-06-09 富泰华工业(深圳)有限公司 语音控制系统及方法
CN105893419A (zh) 2015-11-30 2016-08-24 乐视致新电子科技(天津)有限公司 一种多媒体照片生成方法、装置、设备及手机
US10699186B2 (en) 2015-12-02 2020-06-30 Google Llc Determining orders of execution of a neural network
CN106991478B (zh) 2016-01-20 2020-05-08 中科寒武纪科技股份有限公司 用于执行人工神经网络反向训练的装置和方法
CN106991477B (zh) * 2016-01-20 2020-08-14 中科寒武纪科技股份有限公司 一种人工神经网络压缩编码装置和方法
CN106997236B (zh) 2016-01-25 2018-07-13 亮风台(上海)信息科技有限公司 基于多模态输入进行交互的方法和设备
US10664766B2 (en) 2016-01-27 2020-05-26 Bonsai AI, Inc. Graphical user interface to an artificial intelligence engine utilized to generate one or more trained artificial intelligence models
US10497089B2 (en) 2016-01-29 2019-12-03 Fotonation Limited Convolutional neural network
US10103714B2 (en) 2016-03-01 2018-10-16 Qualcomm Incorporated Adjust voltage for thermal mitigation
JP2017156511A (ja) 2016-03-01 2017-09-07 ソニー株式会社 情報処理装置、情報処理方法、およびプログラム
US10019779B2 (en) 2016-03-08 2018-07-10 Amazon Technologies, Inc. Browsing interface for item counterparts having different scales and lengths
US20190116693A1 (en) * 2016-03-31 2019-04-18 Clear Px Technologies Ltd Temperature controlling device and system having static cooling capacity
US10552119B2 (en) 2016-04-29 2020-02-04 Intel Corporation Dynamic management of numerical representation in a distributed matrix processor architecture
CN107330515A (zh) 2016-04-29 2017-11-07 北京中科寒武纪科技有限公司 一种用于执行人工神经网络正向运算的装置和方法
US10187568B1 (en) 2016-05-02 2019-01-22 Bao Tran Video smart phone
US11055063B2 (en) 2016-05-02 2021-07-06 Marvell Asia Pte, Ltd. Systems and methods for deep learning processor
GB201607713D0 (en) 2016-05-03 2016-06-15 Imagination Tech Ltd Convolutional neural network
CN105978611B (zh) 2016-05-12 2019-09-17 京信通信系统(中国)有限公司 一种频域信号压缩方法及装置
AU2016203619A1 (en) 2016-05-31 2017-12-14 Canon Kabushiki Kaisha Layer-based operations scheduling to optimise memory for CNN applications
EP3252949B1 (en) 2016-06-01 2020-03-18 Intel IP Corporation Methods and devices for predistortion of signals
US20170357910A1 (en) 2016-06-10 2017-12-14 Apple Inc. System for iteratively training an artificial intelligence using cloud-based metrics
CN107545889B (zh) 2016-06-23 2020-10-23 华为终端有限公司 适用于模式识别的模型的优化方法、装置及终端设备
CN106156310A (zh) 2016-06-30 2016-11-23 努比亚技术有限公司 一种图片处理装置和方法
US20180005111A1 (en) * 2016-06-30 2018-01-04 International Business Machines Corporation Generalized Sigmoids and Activation Function Learning
US10372588B2 (en) 2016-07-08 2019-08-06 International Business Machines Corporation Providing debug information on production containers using debug containers
DE102016214786A1 (de) 2016-08-09 2018-02-15 Fujitsu Limited Anwendungsprofiling-Jobmanagement-System, -Programm und -Verfahren
CN107657316B (zh) 2016-08-12 2020-04-07 北京深鉴智能科技有限公司 通用处理器与神经网络处理器的协同系统设计
US20180046903A1 (en) 2016-08-12 2018-02-15 DeePhi Technology Co., Ltd. Deep processing unit (dpu) for implementing an artificial neural network (ann)
CN106354568A (zh) 2016-08-23 2017-01-25 京信通信技术(广州)有限公司 一种不同进程间的通信方法及通信装置
CN107797913A (zh) 2016-09-07 2018-03-13 大陆汽车电子(连云港)有限公司 一种实时系统的软件分析系统与方法
US20180075347A1 (en) * 2016-09-15 2018-03-15 Microsoft Technology Licensing, Llc Efficient training of neural networks
US11907760B2 (en) 2016-09-23 2024-02-20 Apple Inc. Systems and methods of memory allocation for neural networks
CN106650922B (zh) 2016-09-29 2019-05-03 清华大学 硬件神经网络转换方法、计算装置、软硬件协作系统
US20180096243A1 (en) 2016-09-30 2018-04-05 General Electric Company Deep learning for data driven feature representation and anomaly detection
US10726330B2 (en) 2016-10-11 2020-07-28 The Research Foundation For The State University Of New York System, method, and accelerator to process convolutional neural network layers
US11321609B2 (en) * 2016-10-19 2022-05-03 Samsung Electronics Co., Ltd Method and apparatus for neural network quantization
CN106485316B (zh) 2016-10-31 2019-04-02 北京百度网讯科技有限公司 神经网络模型压缩方法以及装置
CN106502626A (zh) 2016-11-03 2017-03-15 北京百度网讯科技有限公司 数据处理方法和装置
US10216479B2 (en) 2016-12-06 2019-02-26 Arm Limited Apparatus and method for performing arithmetic operations to accumulate floating-point numbers
CN106815551B (zh) * 2016-12-08 2019-09-10 新疆农业大学 一种森林资源管理的变异函数参数拟合的优化方法
CN106600070A (zh) * 2016-12-20 2017-04-26 郭建峰 基于ipso‑bp神经网络的短期股价预测算法
US10997492B2 (en) 2017-01-20 2021-05-04 Nvidia Corporation Automated methods for conversions to a lower precision data format
CN108345939B (zh) 2017-01-25 2022-05-24 微软技术许可有限责任公司 基于定点运算的神经网络
JP7004503B2 (ja) * 2017-01-27 2022-01-21 ラピスセミコンダクタ株式会社 自動利得制御回路(agc)、逆拡散回路及び受信データの再生方法
CN106951587A (zh) 2017-02-15 2017-07-14 芯启源(南京)半导体科技有限公司 Fpga调试系统及方法
CN106951962B (zh) 2017-03-22 2020-09-01 南京地平线机器人技术有限公司 用于神经网络的复合运算单元、方法和电子设备
US10332302B2 (en) 2017-04-17 2019-06-25 Intel Corporation Scatter gather engine
US10402932B2 (en) 2017-04-17 2019-09-03 Intel Corporation Power-based and target-based graphics quality adjustment
WO2018192500A1 (zh) * 2017-04-19 2018-10-25 上海寒武纪信息科技有限公司 处理装置和处理方法
CN108734287A (zh) * 2017-04-21 2018-11-02 展讯通信(上海)有限公司 深度神经网络模型的压缩方法及装置、终端、存储介质
CN107025629B (zh) 2017-04-27 2021-03-26 维沃移动通信有限公司 一种图像处理方法及移动终端
KR102034661B1 (ko) * 2017-04-28 2019-10-21 서울대학교산학협력단 뉴럴네트워크를 위한 데이터 양자화 방법 및 장치
US11842280B2 (en) * 2017-05-05 2023-12-12 Nvidia Corporation Loss-scaling for deep neural network training with reduced precision
US10019668B1 (en) 2017-05-19 2018-07-10 Google Llc Scheduling neural network processing
KR102526650B1 (ko) * 2017-05-25 2023-04-27 삼성전자주식회사 뉴럴 네트워크에서 데이터를 양자화하는 방법 및 장치
CN107239826A (zh) * 2017-06-06 2017-10-10 上海兆芯集成电路有限公司 在卷积神经网络中的计算方法及装置
CN107256422A (zh) * 2017-06-06 2017-10-17 上海兆芯集成电路有限公司 数据量化方法及装置
TWI675335B (zh) 2017-06-09 2019-10-21 宏達國際電子股份有限公司 訓練任務優化系統、訓練任務優化方法及其非暫態電腦可讀媒體
US10944902B2 (en) 2017-06-20 2021-03-09 Adobe Inc. Digital image generation using capture support data
EP3646164A4 (en) 2017-06-30 2021-01-20 INTEL Corporation HETEROGENIC MULTIPLIER
CN107451654B (zh) 2017-07-05 2021-05-18 深圳市自行科技有限公司 卷积神经网络的加速运算方法、服务器及存储介质
US10427306B1 (en) 2017-07-06 2019-10-01 X Development Llc Multimodal object identification
CN107992329B (zh) 2017-07-20 2021-05-11 上海寒武纪信息科技有限公司 一种计算方法及相关产品
CN107451658B (zh) 2017-07-24 2020-12-15 杭州菲数科技有限公司 浮点运算定点化方法及系统
CN107480770B (zh) * 2017-07-27 2020-07-28 中国科学院自动化研究所 可调节量化位宽的神经网络量化与压缩的方法及装置
CN107688849B (zh) 2017-07-28 2021-04-13 赛灵思电子科技(北京)有限公司 一种动态策略定点化训练方法及装置
CN107679618B (zh) * 2017-07-28 2021-06-11 赛灵思电子科技(北京)有限公司 一种静态策略定点化训练方法及装置
US11481218B2 (en) 2017-08-02 2022-10-25 Intel Corporation System and method enabling one-hot neural networks on a machine learning compute platform
CN109388779A (zh) * 2017-08-03 2019-02-26 珠海全志科技股份有限公司 一种神经网络权重量化方法和神经网络权重量化装置
US11593644B2 (en) 2017-08-08 2023-02-28 Samsung Electronics Co., Ltd. Method and apparatus for determining memory requirement in a network
US20190050710A1 (en) * 2017-08-14 2019-02-14 Midea Group Co., Ltd. Adaptive bit-width reduction for neural networks
EP3679524A4 (en) * 2017-09-05 2020-10-28 Panasonic Intellectual Property Corporation of America EXECUTION METHOD, EXECUTION DEVICE, LEARNING METHOD, LEARNING DEVICE AND PROGRAM FOR A DEEP NEURONAL NETWORK
CN107644254A (zh) * 2017-09-09 2018-01-30 复旦大学 一种卷积神经网络权重参数量化训练方法及系统
KR20190034985A (ko) * 2017-09-25 2019-04-03 삼성전자주식회사 인공 신경망의 양자화 방법 및 장치
KR102380494B1 (ko) 2017-09-29 2022-03-29 상하이 캠브리콘 인포메이션 테크놀로지 컴퍼니 리미티드 화상처리장치 및 방법
US10224954B1 (en) 2017-09-29 2019-03-05 Intel Corporation Floating point to fixed point conversion
CN107679490B (zh) * 2017-09-29 2019-06-28 百度在线网络技术(北京)有限公司 用于检测图像质量的方法和装置
US10223114B1 (en) 2017-09-29 2019-03-05 Intel Corporation Fixed point to floating point conversion
US11437032B2 (en) 2017-09-29 2022-09-06 Shanghai Cambricon Information Technology Co., Ltd Image processing apparatus and method
US11450319B2 (en) 2017-09-29 2022-09-20 Cambricon (Xi'an) Semiconductor Co., Ltd. Image processing apparatus and method
JP6540770B2 (ja) * 2017-10-17 2019-07-10 富士通株式会社 演算処理回路、演算処理回路を含む演算処理装置、演算処理装置を含む情報処理装置、および方法
KR102564456B1 (ko) * 2017-10-19 2023-08-07 삼성전자주식회사 뉴럴 네트워크의 파라미터를 양자화하는 장치 및 방법
US10410121B2 (en) 2017-10-25 2019-09-10 SparkCognition, Inc. Adjusting automated neural network generation based on evaluation of candidate neural networks
US20210061028A1 (en) 2017-10-26 2021-03-04 Applied Mechatronic Products Apparatus and method for vehicular monitoring, analysis, and control
KR20190054454A (ko) * 2017-11-13 2019-05-22 삼성전자주식회사 인공 신경망의 양자화 방법 및 장치
US10783634B2 (en) 2017-11-22 2020-09-22 General Electric Company Systems and methods to deliver point of care alerts for radiological findings
US10803379B2 (en) 2017-12-12 2020-10-13 Amazon Technologies, Inc. Multi-memory on-chip computational network
CN108053028B (zh) 2017-12-21 2021-09-14 深圳励飞科技有限公司 数据定点化处理方法、装置、电子设备及计算机存储介质
CN108229681A (zh) * 2017-12-28 2018-06-29 郑州云海信息技术有限公司 一种神经网络模型压缩方法、系统、装置及可读存储介质
US11636327B2 (en) 2017-12-29 2023-04-25 Intel Corporation Machine learning sparse computation mechanism for arbitrary neural networks, arithmetic compute microarchitecture, and sparsity for training mechanism
US11373088B2 (en) 2017-12-30 2022-06-28 Intel Corporation Machine learning accelerator mechanism
CN108229663A (zh) * 2018-01-29 2018-06-29 百度在线网络技术(北京)有限公司 用于生成卷积神经网络的方法和装置
CN108288089A (zh) * 2018-01-29 2018-07-17 百度在线网络技术(北京)有限公司 用于生成卷积神经网络的方法和装置
US20190251429A1 (en) 2018-02-12 2019-08-15 Kneron, Inc. Convolution operation device and method of scaling convolution input for convolution neural network
US11169803B2 (en) 2018-02-13 2021-11-09 Shanghai Cambricon Information Technology Co., Ltd. Computing device and method
US11630666B2 (en) 2018-02-13 2023-04-18 Shanghai Cambricon Information Technology Co., Ltd Computing device and method
US11106598B2 (en) 2018-02-13 2021-08-31 Shanghai Cambricon Information Technology Co., Ltd. Computing device and method
CN116991225A (zh) 2018-02-14 2023-11-03 上海寒武纪信息科技有限公司 处理器的控制装置、方法及设备
JP7056225B2 (ja) 2018-02-26 2022-04-19 富士通株式会社 演算処理装置、情報処理装置、情報処理方法、およびプログラム
US10628275B2 (en) 2018-03-07 2020-04-21 Nxp B.V. Runtime software-based self-test with mutual inter-core checking
US11475306B2 (en) 2018-03-22 2022-10-18 Amazon Technologies, Inc. Processing for multiple input data sets
CN108631727B (zh) * 2018-03-26 2019-08-09 河北工业大学 一种基于卷积神经网络的太阳能电池板缺陷识别方法
CN108491928B (zh) * 2018-03-29 2019-10-25 腾讯科技(深圳)有限公司 模型参数发送方法、装置、服务器及存储介质
CN108509627B (zh) * 2018-04-08 2021-08-31 腾讯科技(深圳)有限公司 数据离散化模型训练方法和装置、数据离散方法
CN108510067B (zh) 2018-04-11 2021-11-09 西安电子科技大学 基于工程化实现的卷积神经网络量化方法
US11562213B2 (en) 2018-04-17 2023-01-24 Intel Corporation Methods and arrangements to manage memory in cascaded neural networks
CN108596328B (zh) * 2018-04-26 2021-02-02 北京市商汤科技开发有限公司 一种定点化方法及装置、计算机设备
US10691413B2 (en) 2018-05-04 2020-06-23 Microsoft Technology Licensing, Llc Block floating point computations using reduced bit-width vectors
EP3624020A4 (en) 2018-05-18 2021-05-05 Shanghai Cambricon Information Technology Co., Ltd CALCULATION PROCEDURES AND RELATED PRODUCTS
CN108717570A (zh) 2018-05-23 2018-10-30 电子科技大学 一种脉冲神经网络参数量化方法
CN115268082A (zh) 2018-05-31 2022-11-01 中强光电股份有限公司 头戴式显示装置
US10360304B1 (en) 2018-06-04 2019-07-23 Imageous, Inc. Natural language processing interface-enabled building conditions control system
CN109062540B (zh) 2018-06-06 2022-11-25 北京理工大学 一种基于cordic算法的可重构浮点运算装置
CN109063820A (zh) 2018-06-07 2018-12-21 中国科学技术大学 利用时频联合长时循环神经网络的数据处理方法
CN108830331A (zh) * 2018-06-22 2018-11-16 西安交通大学 一种基于全卷积网络的探地雷达目标检测方法
CN109102064B (zh) * 2018-06-26 2020-11-13 杭州雄迈集成电路技术股份有限公司 一种高精度的神经网络量化压缩方法
EP3798850A4 (en) 2018-06-27 2022-03-23 Shanghai Cambricon Information Technology Co., Ltd ON-CHIP CODE BREAKPOINT DEBUG METHOD, ON-CHIP PROCESSOR AND CHIP BREAKPOINT DEBUG SYSTEM
CN110728364A (zh) 2018-07-17 2020-01-24 上海寒武纪信息科技有限公司 一种运算装置和运算方法
CN109002889B (zh) * 2018-07-03 2021-12-17 华南理工大学 自适应迭代式卷积神经网络模型压缩方法
CN109214504B (zh) * 2018-08-24 2020-09-04 北京邮电大学深圳研究院 一种基于fpga的yolo网络前向推理加速器设计方法
EP3757896B1 (en) 2018-08-28 2023-01-11 Cambricon Technologies Corporation Limited Method and device for pre-processing data in a neural network
WO2020062392A1 (zh) 2018-09-28 2020-04-02 上海寒武纪信息科技有限公司 信号处理装置、信号处理方法及相关产品
CN109598331A (zh) * 2018-12-04 2019-04-09 北京芯盾时代科技有限公司 一种欺诈识别模型训练方法、欺诈识别方法及装置
CN109685202B (zh) 2018-12-17 2023-03-21 腾讯科技(深圳)有限公司 数据处理方法及装置、存储介质和电子装置
CN109754074A (zh) * 2018-12-29 2019-05-14 北京中科寒武纪科技有限公司 一种神经网络量化方法、装置以及相关产品
GB2580171B (en) * 2018-12-21 2021-02-17 Imagination Tech Ltd Methods and systems for selecting quantisation parameters for deep neural networks using back-propagation
CN111385462A (zh) 2018-12-28 2020-07-07 上海寒武纪信息科技有限公司 信号处理装置、信号处理方法及相关产品
CN109740739B (zh) * 2018-12-29 2020-04-24 中科寒武纪科技股份有限公司 神经网络计算装置、神经网络计算方法及相关产品
CN109740754B (zh) * 2018-12-29 2020-04-14 中科寒武纪科技股份有限公司 神经网络计算装置、神经网络计算方法及相关产品
CN109800865B (zh) * 2019-01-24 2021-03-23 北京市商汤科技开发有限公司 神经网络生成及图像处理方法和装置、平台、电子设备
US20190164057A1 (en) * 2019-01-30 2019-05-30 Intel Corporation Mapping and quantification of influence of neural network features for explainable artificial intelligence
CN109859135B (zh) * 2019-01-31 2021-05-07 北京邮电大学 一种应用于关联成像的图像增强处理方法
CN109902745A (zh) 2019-03-01 2019-06-18 成都康乔电子有限责任公司 一种基于cnn的低精度训练与8位整型量化推理方法
CN110059733A (zh) 2019-04-01 2019-07-26 苏州科达科技股份有限公司 卷积神经网络的优化及快速目标检测方法、装置
CN109993296B (zh) 2019-04-01 2020-12-29 安徽寒武纪信息科技有限公司 量化实现方法及相关产品
US20200334522A1 (en) 2019-04-18 2020-10-22 Cambricon Technologies Corporation Limited Data processing method and related products
CN111832737B (zh) 2019-04-18 2024-01-09 中科寒武纪科技股份有限公司 一种数据处理方法及相关产品
US20200364552A1 (en) * 2019-05-13 2020-11-19 Baidu Usa Llc Quantization method of improving the model inference accuracy
US11531893B2 (en) * 2019-06-03 2022-12-20 Samsung Electronics Co., Ltd. Method and apparatus with neural network parameter quantization
US11676028B2 (en) * 2019-06-12 2023-06-13 Shanghai Cambricon Information Technology Co., Ltd Neural network quantization parameter determination method and related products
US20210374510A1 (en) 2019-08-23 2021-12-02 Anhui Cambricon Information Technology Co., Ltd. Data processing method, device, computer equipment and storage medium
JP7146952B2 (ja) 2019-08-23 2022-10-04 安徽寒武紀信息科技有限公司 データ処理方法、装置、コンピュータデバイス、及び記憶媒体
EP4020329A4 (en) 2019-08-23 2023-07-05 Anhui Cambricon Information Technology Co., Ltd. DATA PROCESSING METHOD AND APPARATUS, COMPUTER EQUIPMENT, AND STORAGE MEDIA
JP7146953B2 (ja) 2019-08-27 2022-10-04 安徽寒武紀信息科技有限公司 データ処理方法、装置、コンピュータデバイス、及び記憶媒体
CN110780845B (zh) 2019-10-17 2021-11-30 浙江大学 一种用于量化卷积神经网络的可配置近似乘法器及其实现方法

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US9916531B1 (en) * 2017-06-22 2018-03-13 Intel Corporation Accumulator constrained quantization of convolutional neural networks
CN109214509A (zh) * 2017-07-05 2019-01-15 中国科学院沈阳自动化研究所 一种用于深度神经网络高速实时量化结构和运算实现方法
CN109389219A (zh) * 2017-08-04 2019-02-26 三星电子株式会社 用于对神经网络的参数进行量化的方法和装置
CN109146057A (zh) * 2018-06-26 2019-01-04 杭州雄迈集成电路技术有限公司 一种基于查表计算的高精度的神经网络工程化方法
CN109472353A (zh) * 2018-11-22 2019-03-15 济南浪潮高新科技投资发展有限公司 一种卷积神经网络量化电路及量化方法
CN109800877A (zh) * 2019-02-20 2019-05-24 腾讯科技(深圳)有限公司 神经网络的参数调整方法、装置及设备

Also Published As

Publication number Publication date
JP2021179966A (ja) 2021-11-18
EP3772023A1 (en) 2021-02-03
CN112085181B (zh) 2024-03-29
CN111652367A (zh) 2020-09-11
JP7167405B2 (ja) 2022-11-09
CN112085183B (zh) 2024-04-02
CN112085186B (zh) 2024-03-05
CN112085185A (zh) 2020-12-15
KR20210011461A (ko) 2021-02-01
CN112085185B (zh) 2024-04-02
CN112400176A (zh) 2021-02-23
CN112085190B (zh) 2024-04-02
CN112085192B (zh) 2024-03-29
CN111652367B (zh) 2024-04-09
CN112085193B (zh) 2024-03-29
KR20210018352A (ko) 2021-02-17
CN112085189B (zh) 2024-03-29
KR102609719B1 (ko) 2023-12-04
CN112085191B (zh) 2024-04-02
CN112085190A (zh) 2020-12-15
JP7166704B2 (ja) 2022-11-08
KR102656332B1 (ko) 2024-04-09
CN111652368B (zh) 2024-03-29
CN112085183A (zh) 2020-12-15
US11675676B2 (en) 2023-06-13
EP3998554A4 (en) 2023-11-15
JP2021177369A (ja) 2021-11-11
CN112085188A (zh) 2020-12-15
CN112085184B (zh) 2024-03-29
CN111652368A (zh) 2020-09-11
WO2020248424A1 (zh) 2020-12-17
US20220261634A1 (en) 2022-08-18
CN112085181A (zh) 2020-12-15
JP2021530769A (ja) 2021-11-11
CN112085191A (zh) 2020-12-15
US20210286688A1 (en) 2021-09-16
EP3770823A1 (en) 2021-01-27
CN112085184A (zh) 2020-12-15
CN112085188B (zh) 2024-04-02
KR20210011462A (ko) 2021-02-01
EP3772022A1 (en) 2021-02-03
EP3998554A1 (en) 2022-05-18
CN112085193A (zh) 2020-12-15
CN112085189A (zh) 2020-12-15
CN112085186A (zh) 2020-12-15
CN112085192A (zh) 2020-12-15
EP3770823A4 (en) 2021-01-27

Similar Documents

Publication Publication Date Title
WO2020248423A1 (zh) 一种神经网络的量化参数确定方法及相关产品
US11676028B2 (en) Neural network quantization parameter determination method and related products
WO2021036905A1 (zh) 数据处理方法、装置、计算机设备和存储介质
WO2021036908A1 (zh) 数据处理方法、装置、计算机设备和存储介质
WO2021036904A1 (zh) 数据处理方法、装置、计算机设备和存储介质
JPWO2020248424A5 (zh)
WO2022012233A1 (zh) 一种量化校准方法、计算装置和计算机可读存储介质
WO2021082653A1 (zh) 数据处理方法、装置、计算机设备和存储介质
US20220222041A1 (en) Method and apparatus for processing data, and related product

Legal Events

Date Code Title Description
ENP Entry into the national phase

Ref document number: 2019786896

Country of ref document: EP

Effective date: 20191024

NENP Non-entry into the national phase

Ref country code: DE