CN114677548A - Neural network image classification system and method based on resistive random access memory - Google Patents
Neural network image classification system and method based on resistive random access memory Download PDFInfo
- Publication number
- CN114677548A CN114677548A CN202210579664.2A CN202210579664A CN114677548A CN 114677548 A CN114677548 A CN 114677548A CN 202210579664 A CN202210579664 A CN 202210579664A CN 114677548 A CN114677548 A CN 114677548A
- Authority
- CN
- China
- Prior art keywords
- value
- layer
- convolution
- quantized
- quantization
- Prior art date
- Legal status (The legal status 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 status listed.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 64
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 41
- 238000013139 quantization Methods 0.000 claims abstract description 145
- 238000012549 training Methods 0.000 claims abstract description 43
- 238000011176 pooling Methods 0.000 claims abstract description 42
- 238000012360 testing method Methods 0.000 claims abstract description 36
- 230000004913 activation Effects 0.000 claims abstract description 35
- 238000003062 neural network model Methods 0.000 claims abstract description 24
- 230000008447 perception Effects 0.000 claims abstract description 12
- 238000007667 floating Methods 0.000 claims description 70
- 238000013507 mapping Methods 0.000 claims description 23
- 230000015654 memory Effects 0.000 claims description 14
- 230000008569 process Effects 0.000 claims description 13
- 238000011002 quantification Methods 0.000 claims description 6
- 238000004804 winding Methods 0.000 claims description 6
- 230000003213 activating effect Effects 0.000 claims description 3
- 238000012545 processing Methods 0.000 description 9
- 238000004422 calculation algorithm Methods 0.000 description 4
- 238000004364 calculation method Methods 0.000 description 4
- 238000010586 diagram Methods 0.000 description 4
- 238000004088 simulation Methods 0.000 description 4
- 238000006073 displacement reaction Methods 0.000 description 3
- 238000003475 lamination Methods 0.000 description 3
- 238000004590 computer program Methods 0.000 description 2
- 238000013135 deep learning Methods 0.000 description 2
- 230000006870 function Effects 0.000 description 2
- 238000005096 rolling process Methods 0.000 description 2
- 238000004458 analytical method Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 239000000463 material Substances 0.000 description 1
- 239000011159 matrix material Substances 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000003058 natural language processing Methods 0.000 description 1
- 238000010606 normalization Methods 0.000 description 1
- 230000000644 propagated effect Effects 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 238000010998 test method Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/06—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons
- G06N3/063—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons using electronic means
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Health & Medical Sciences (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Molecular Biology (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Evolutionary Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Neurology (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a neural network image classification system and method based on a resistive random access memory, wherein the system comprises an input layer, a group of convolution layers and a full-connection layer which are sequentially connected, and a convolution quantization layer, a convolution inverse quantization layer, an activation layer and a pooling layer are arranged for the convolution layers in a matched mode, and the method comprises the following steps of S1: normalizing the image to be classified to obtain a normalized image; step S2: constructing a training set and a test set for the normalized image; step S3: constructing a neural network model based on a resistive random access memory; step S4: inputting the training set into a neural network model based on a resistive random access memory, and performing quantitative perception training to obtain model parameters after quantitative perception training, wherein the method comprises the following steps: step S5: and inputting the test set image into the trained neural network for forward reasoning test.
Description
Technical Field
The invention relates to the technical field of neural network image classification, in particular to a neural network image classification system and method based on a resistive random access memory.
Background
With the rapid development of deep learning, neural network technology has been widely applied in various fields such as image recognition, speech recognition, natural language processing, and the like. The application of neural networks is typically deployed at the edge device side. In a traditional chip architecture, a memory and a calculation are separated, a calculation unit reads data from the memory first, and the data is stored back to the memory after the calculation is completed. However, in the face of the high concurrency requirement of the neural network, the conventional chip architecture needs to frequently carry data, which results in huge power consumption and computation bottleneck.
The resistive random access memory (ReRAM), also known as a memristor, has the advantages of low power consumption, simple structure, high working speed, controllable and variable resistance value and the like, and meanwhile, the memristor can realize various operation forms such as logic operation, matrix multiplication and the like. The characteristic of using the memristor to store and calculate the whole body can reduce the transportation of data and reduce the storage requirement. Therefore, ReRAM has great potential to solve the problems of conventional chip architectures. In recent years, memristor-based neural network accelerators provide an effective solution for neural network reasoning.
Although ReRAM has great advantages for implementing neural network reasoning, the neural network model needs to be compressed during implementation, which results in loss of image recognition accuracy. The reasonable and effective quantization method can reduce the storage space of data and improve the calculation speed under the condition of low precision loss. In the existing mainstream deep learning platform such as the neural network quantization algorithm of tensorflow and pytorch, after the operation of a convolution operator and a full-link operator, data can exceed the quantized bit width. Therefore, the quantized bit width needs to be scaled, the algorithm needs to be multiplied by a floating point decimal number for scaling, and the hardware needs to approximate the scaling operation of the algorithm through two operations of left shift and right shift. However, the conductance range of the ReRAM device and the quantization bit width input by each layer are limited, so that the resource occupancy rate is high and the operation is complex.
Disclosure of Invention
In order to solve the defects of the prior art, based on the idea of device and algorithm collaborative design, the method combines the conductance range of a ReRAM device and the characteristic that the quantization bit width input by each layer is limited, optimizes the bit width of a quantization factor by designing a constraint condition, ensures that the quantization factor adopts the optimal power of 2, and only needs to shift a limited number of bits right in the operation process of scaling after convolution, realizes simple operation, and achieves the purposes of reducing the loss of image recognition precision, improving the speed of image recognition and reducing the resource occupancy rate, and the method adopts the following technical scheme:
a neural network image classification system based on a resistive random access memory comprises an input layer, a group of convolution layers and a full-connection layer which are sequentially connected, wherein a convolution quantization layer, a convolution inverse quantization layer, an activation layer and a pooling layer are arranged for the convolution layers in a matched mode, the input layer is used for obtaining a training set image, the convolution quantization layer quantizes an input value of the input layer and convolution of a first convolution layer to obtain a quantized input value and a convolution kernel, the convolution inverse quantization layer dequantizes the quantized input value and the convolution kernel to obtain a dequantized value of the first convolution layer, a bit on a storage device is subjected to shift operation based on a digital domain to obtain a dequantized shifted value of the first convolution layer, the dequantized shifted value is subjected to activation operation through the activation layer, the activated value is subjected to pooling operation through the pooling layer, and the pooled value is used as an input value of a next convolution quantization layer, until the pooled output corresponding to the last convolutional layer, the final pooled output is subjected to classification prediction results of training set images through a full connection layer, back propagation is carried out according to errors of the prediction results and training set true values, a neural network model based on a resistive random access memory is trained, and gradient solution cannot be carried out due to the fact that a rounding method is adopted in a quantization method, so that errors are directly transmitted back to a value before quantization by skipping a quantization layer in the back propagation process, and network parameters are optimized by updating the weight of the value before quantization, so that precision loss caused by quantization is reduced; inputting an image to be classified into a trained system, quantizing the convolution of an input value of an input layer and a first convolution layer through a convolution quantization layer, performing convolution operation on the quantized value of the input layer and the quantized value of the first convolution layer to obtain a quantized value output by the first convolution layer, mapping the quantized value of the input layer to a voltage value of a resistive random access memory, mapping the quantized value of the first convolution layer to a conductance value of the resistive random access memory, mapping the result of the convolution operation to a current value output by the resistive random access memory, converting the current value to the voltage value, performing shift operation on a bit on a storage device based on the voltage value to obtain a value output by the first convolution layer after quantization shift, performing activation operation on the quantized value through an activation layer, performing pooling operation on the activated value through a pooling layer, and using the pooled value as an input value of a next convolution layer, and obtaining the classification result of the image to be classified through the final pooled output through the full-connection layer until the pooled output corresponding to the last convolution layer.
The resistive random access memory is formed into a resistive random access memory array, the quantized value of the input layer is mapped into a voltage value of the resistive random access memory and is input into a first row of resistive random access memories, the quantized value of the coiling layer is mapped into a conductance value of each resistive random access memory, and the current value output by each row of resistive random access memories is the quantized value input by the row and is convolved with the quantized value of the coiling layer.
Furthermore, the convolution layer comprises a first convolution layer and a second convolution layer, and a convolution quantization layer, a convolution inverse quantization layer, an activation layer and a pooling layer are respectively matched with the first convolution layer and the second convolution layer.
Further, the quantization process is as follows:
formula (1) represents a floating-point convolution operation;
wherein,
a floating-point value representing the input layer,
a floating-point value representing a convolution kernel of the first convolution layer,
representing a convolution operation. Respectively mapping the floating point value of the input layer and the floating point value of the convolution kernel of the first convolution layer to a fixed point value, and determining the decimal bit width of the optimal fixed point value through formulas (2), (3) and (4);
formula (2) calculating the minimum value of the floating point value mapped to the fixed point value;
where i represents the number of layers of the neural network model,
the input layer is represented by a representation of,
a first layer of the volume quantification is shown,
a first output layer of a volume is shown,
the floating-point values representing the ith layer are mapped to fractional bits wide of the fixed-point values,
which represents the bit-width of the quantization,
a minimum value representing that the floating point value of the ith layer is mapped to the fixed point value;
formula (3) calculating the maximum value of the floating point value mapped to the fixed point value;
wherein
The floating point value representing the ith layer is mapped to the maximum value of the fixed point value;
calculating the decimal bit width of the optimal fixed point value through the constraint condition of the formula (4)
The constraint of equation (4) is such that the range of fixed point values is as close as possible to the range of floating point values to reduce the loss of accuracy caused by quantization;
constraint conditions are as follows:
(4)
wherein
It is shown that the absolute value is calculated,
is shown asThe maximum value of the i-layer floating point,
represents the minimum value of the floating point of the ith layer,
，
the method comprises the steps of obtaining a maximum value and a minimum value of an ith layer floating point value through statistics;
solving the quantization value of each layer through a formula (5);
wherein
A quantization factor representing the ith layer floating point value,
a floating-point value representing the i-th layer,
represents the quantized value of the ith layer,
which represents the operation of rounding off,
representing the minimum value after quantization to the integer,
representing the maximum value after quantization to the integer,
indicating a truncation operation.
Further, respectively carrying out inverse quantization on the quantized input value and the convolution kernel, and then carrying out convolution operation on the inverse quantized input value and the inverse quantized convolution kernel through a formula (6) to obtain an inverse quantized floating point value output by the first convolution layer;
wherein,
representing the input value after the quantization,
representing the input value after the inverse quantization,
representing the quantized convolution kernel or kernels and,
representing the convolution kernel after the dequantization,
representing the quantized value of the first convolutional output layer,
representing the inverse quantized value of the first convolutional output layer.
Equation (7) can be exited by equation (6):
performing shift operation by formula (7) to obtain the quantized value output by the first convolution layer
，
Representing the minimum value after quantization to the integer,
representing the maximum value after quantization to the integer,
indicating a truncation operation.
A neural network image classification method based on a resistive random access memory comprises the following steps:
step S1: normalizing the image to be classified to obtain a normalized image;
step S2: constructing a training set and a test set for the normalized image;
step S3: constructing a neural network model based on a resistive random access memory;
step S4: inputting the training set into a neural network model based on a resistive random access memory, and performing quantitative perception training to obtain model parameters after quantitative perception training, wherein the method comprises the following steps:
step S4-1: quantizing the convolution of the input value of the input layer and the first convolution layer to obtain a quantized input value and a convolution kernel;
step S4-2: respectively carrying out inverse quantization on the quantized input value and the convolution kernel, carrying out activation operation on the inverse quantized value through an activation layer, carrying out pooling operation on the activated value through a pooling layer to obtain a first convolution layer output inverse quantized value, and carrying out shift operation on a bit on the storage equipment based on the digital domain to obtain a first convolution layer output inverse quantized shifted value; in particular in the digital domain by
Back (corresponding to shift operation on hardware storage device, shift right)
After bit) to obtain a quantized value of the convolutional layer output
. The right shift operation realizes the remaining operation of formula (7) in step S4, and finally obtains the pooled value through the activation operation and the maximum pooling operation;
step S4-3: activating the inversely quantized value through an activation layer, performing pooling operation on the activated value through a pooling layer, taking the pooled value as an input value of a next convolution quantization layer until the pooled output corresponding to the last convolution layer, obtaining a classification prediction result of a training set image through a full connection layer, performing back propagation according to an error between the prediction result and a training set true value, and training a neural network model based on a Resistive Random Access Memory (RRAM), wherein because a rounding method is adopted in a quantization method, gradient solution cannot be performed, so that in the process of back propagation, the error is directly transmitted back to the value before quantization by skipping a quantization layer, and network parameters are optimized by updating the weight of the value before quantization, thereby reducing the precision loss caused by quantization;
step S5: inputting the test set image into the trained neural network for forward reasoning test, comprising the following steps:
step S5-1: taking the test set as input, performing convolution operation on the quantized value of the input layer and the quantized value of the first coiling layer obtained in the steps S3 and S4 to obtain the quantized value output by the first coiling layer, mapping the quantized value of the input layer to be the voltage value of the resistive random access memory, mapping the quantized value of the first coiling layer to be the conductance value of the resistive random access memory, and mapping the result of the convolution operation to be the current value output by the resistive random access memory;
step S5-2: converting the current value into a voltage value, performing a shift operation based on the voltage value to obtain a value after the first convolution layer outputs the quantization shift, performing an activation operation on the quantized value through an activation layer, performing a pooling operation on the activated value through a pooling layer, taking the pooled value as an input value of the next convolution layer until the pooled output corresponding to the last convolution layer is output, and obtaining a classification result of the test set image through a full connection layer by the last pooled output.
Further, the specific quantification method of step S4-1 includes the following steps:
formula (1) represents a convolution operation of floating points;
wherein,
a floating-point value representing the input layer,
a floating-point value representing a convolution kernel of the first convolution layer,
representing a convolution operation. Respectively mapping the floating point value of the input layer and the floating point value of the convolution kernel of the first convolution layer to a fixed point value, and determining the decimal bit width of the optimal fixed point value through formulas (2), (3) and (4);
formula (2) calculating the minimum value of the floating point value mapped to the fixed point value;
where i represents the number of layers of the neural network model,
the input layer is represented by a representation of,
a first layer of the volume quantification is shown,
a first output layer of a volume is shown,
the floating-point values representing the ith layer are mapped to fractional bits wide of the fixed-point values,
which represents the bit-width of the quantization,
a minimum value representing that the floating point value of the ith layer is mapped to the fixed point value;
formula (3) calculating the maximum value of the floating point value mapped to the fixed point value;
wherein
The floating point value representing the ith layer is mapped to the maximum value of the fixed point value;
calculating the decimal bit width of the optimal fixed point value through the constraint condition of the formula (4)
The constraint of equation (4) is such that the range of fixed point values is as close as possible to the range of floating point values to reduce the loss of accuracy caused by quantization;
constraint conditions are as follows:
wherein
It is shown that the absolute value is calculated,
represents the maximum value of the floating point of the ith layer,
represents the minimum value of the floating point of the ith layer,
，
the method comprises the steps of obtaining a maximum value and a minimum value of an ith layer floating point value through statistics;
solving the quantization value of each layer through a formula (5);
wherein
A quantization factor representing the ith layer floating point value,
a floating point value representing the ith layer,
represents the quantized value of the i-th layer,
which represents the operation of rounding off,
representing the minimum value after quantization to the integer,
representing the maximum value after quantization to the integer,
indicating a truncation operation.
Further, in the step S4-2, the quantized input value and the convolution kernel are respectively dequantized, and then the dequantized input value and the dequantized convolution kernel are convolved by the formula (6), so as to obtain a dequantized floating point value output by the first convolution layer;
wherein,
representing the input value after the quantization,
representing the input value after the inverse quantization,
representing the quantized convolution kernel or kernels and,
representing the convolution kernel after the inverse quantization,
representing the quantized values of the first convolutional output layer,
representing the inverse quantized value of the first convolutional output layer.
Equation (7) can be derived from equation (6):
performing shift operation by formula (7) to obtain the quantized value output by the first convolution layer
，
Representing the minimum value after quantization to the integer,
representing quantization to unityThe maximum value after the shape is determined,
indicating a truncation operation.
In step S5-2, the resistive random access memory is configured to form a resistive random access memory array, the quantized values of the input layer are mapped to voltage values of the resistive random access memory and input to the first row of resistive random access memory, the quantized values of the rolling layer are mapped to conductance values of the resistive random access memory, and the current value output by each row of resistive random access memory is a convolution operation of the quantized values input by the row and the quantized values of the rolling layer.
A neural network image classification device based on a resistive random access memory comprises one or more processors and is used for realizing the neural network image classification method based on the resistive random access memory.
The invention has the advantages and beneficial effects that:
according to the neural network image classification system and method based on the resistive random access memory, due to the fact that the conductance range of a ReRAM device is limited, a limited bit width is needed to store a convolution kernel. Since the quantization bit width of each layer input is limited, the limited bit width is required to store the convolved output value. According to the method, the bit width of the quantization factor is optimized by designing the constraint condition, so that the quantization factor adopts the optimal power of 2, only limited digits need to be shifted to the right in the operation process of scaling after convolution, and the operation is simple. The precision loss caused by the right shift of ADC (analog-to-digital converter) is reduced. And meanwhile, quantization perception training is carried out, so that the loss of precision caused by quantization is reduced, and the reasoning speed of the model is improved.
Drawings
FIG. 1 is a flow chart of a method of an embodiment of the present invention.
Fig. 2 is a flowchart of training a neural network model based on a resistive random access memory in an embodiment of the present invention.
Fig. 3 is a flowchart of image classification prediction by a trained model according to an embodiment of the present invention.
Fig. 4 is a partial example diagram of an input image in the embodiment of the present invention.
FIG. 5 is a diagram of a ReRAM based crossbar array in an embodiment of the present invention.
FIG. 6 is a graph of a comparison of floating point models and classification accuracy of the method of the present invention for each class of the test set.
Fig. 7 is a diagram showing the structure of an apparatus according to an embodiment of the present invention.
Detailed Description
The following describes in detail embodiments of the present invention with reference to the drawings. It should be understood that the detailed description and specific examples, while indicating the present invention, are given by way of illustration and explanation only, not limitation.
As shown in fig. 1 to 3, an embodiment of the present invention is to classify a washion mnist dataset. As shown in fig. 4, the data set has a total of 50000 training samples and 10000 testing samples. Each sample is a 28 x 28 size grayscale image. The data sets are 10 types in total, namely T-shirts, trousers, blouses, dresses, coats, sandals, shirts, sports shoes, schoolbag and booties.
The invention provides a neural network image classification system based on a resistive random access memory, which comprises an input layer, a group of convolution layers and a full-connection layer which are sequentially connected, wherein the convolution layers are matched with a convolution quantization layer, a convolution inverse quantization layer, an activation layer and a pooling layer, the input layer is used for acquiring a training set image, the convolution quantization layer quantizes an input value of the input layer and convolution of a first volume of lamination to obtain a quantized input value and a convolution kernel, the convolution inverse quantization layer dequantizes the quantized input value and the convolution kernel to obtain a dequantized value of the first volume of lamination, the bit corresponding to a storage device is subjected to shift operation based on a digital domain to obtain a dequantized shifted value of the first volume of lamination, the dequantized shifted value is subjected to activation operation through the activation layer, the activated value is subjected to pooling operation through the pooling layer, and the pooled value is used as an input value of a next volume of the quantization layer, until the pooled output corresponding to the last convolutional layer, the final pooled output is subjected to classification prediction results of training set images through a full connection layer, back propagation is carried out according to errors of the prediction results and training set true values, a neural network model based on a resistive random access memory is trained, and gradient solution cannot be carried out due to the fact that a rounding method is adopted in a quantization method, so that errors are directly transmitted back to a value before quantization by skipping a quantization layer in the back propagation process, and network parameters are optimized by updating the weight of the value before quantization, so that precision loss caused by quantization is reduced; inputting an image to be classified into a trained system, quantizing the convolution of an input value of an input layer and a first convolution layer through a convolution quantization layer, performing convolution operation on the quantized value of the input layer and the quantized value of the first convolution layer to obtain a quantized value output by the first convolution layer, mapping the quantized value of the input layer to a voltage value of a resistive random access memory, mapping the quantized value of the first convolution layer to a conductance value of the resistive random access memory, mapping the result of the convolution operation to a current value output by the resistive random access memory, converting the current value to the voltage value, performing shift operation based on the voltage value to obtain a value after the quantized shift of the first convolution layer output, performing activation operation on the quantized value through an activation layer, performing pooling operation on the activated value through a pooling layer, taking the pooled value as an input value of a next convolution layer until the pooled output corresponding to the last convolution layer, and outputting the final pooled output through a full connection layer to obtain a classification result of the image to be classified.
The resistive random access memory forms a resistive random access memory array, the quantized value of the input layer is mapped into the voltage value of the resistive random access memory and is input into a first row of resistive random access memories, the quantized value of the winding layer is mapped into the conductance value of each resistive random access memory, and the current value output by each row of resistive random access memories is the convolution operation of the quantized value input by the row and the quantized value of the winding layer.
The convolution layer comprises a first convolution layer and a second convolution layer, and a convolution quantization layer, a convolution inverse quantization layer, an activation layer and a pooling layer are respectively matched with the first convolution layer and the second convolution layer.
The invention provides a neural network image classification method based on a resistive random access memory, which comprises the following steps:
step S1: normalizing the image to be classified to obtain a normalized image; the pixel values of the image are normalized to be between 0 and 1, and in the embodiment of the invention, after the normalization operation is performed on the pixel values of all the samples in the fast mnist data set by dividing the pixel values by 255, the range of the pixel values of the samples becomes [0,1 ].
Step S2: constructing a training set and a test set for the normalized image; and selecting the training sample in the fast mnist as a training set, and selecting the test sample in the fast mnist as a test set.
Step S3: constructing a neural network model based on a resistive random access memory;
specifically, the neural network model structure is as follows: the input layer → the first convolution quantization layer → the first convolution inverse quantization layer → the active layer → the pooling layer → the second convolution quantization layer → the second convolution inverse quantization layer → the active layer → the pooling layer → the all-connected quantization layer → the all-connected inverse quantization layer → the softmax layer. The size of each layer weight parameter is set as follows:
the input layer has a size of
；
A first convolution quantization layer with a convolution kernel parameter of size
Step length is 1;
a second convolution quantization layer with convolution kernel parameters of size
Step length is 1;
all-connected quantified layer with all-connected parameters of size
。
Step S4: inputting the training set into a neural network model based on a resistive random access memory, and performing quantitative perception training to obtain model parameters after quantitative perception training, wherein the method comprises the following steps:
in the embodiment of the present invention, the quantization bit width is 8 bits, the input is quantized to [0,255], and the weight parameter of each layer is quantized to [ -128, 127 ]. The method comprises the following specific steps:
step S4-1: quantizing the convolution of the input value of the input layer and the first convolution layer to obtain a quantized input value and a convolution kernel; the specific quantification method is as follows:
formula (1) represents a floating-point convolution operation;
wherein,
a floating-point value representing the input layer,
a floating-point value representing a convolution kernel of the first convolution layer,
representing a convolution operation. The floating point value of the input layer and the floating point value of the convolution kernel of the first convolution layer are mapped to the fixed point value, and the decimal bit width of the optimal fixed point value is determined through formulas (2), (3) and (4).
Formula (2) calculating the minimum value of the floating point value mapped to the fixed point value;
wherein
In
The input layer is represented by a representation of,
a first layer of the volume quantification is shown,
representing a first convolution outputA layer of a material selected from the group consisting of,
the floating-point values representing the i-layer are mapped to fractional bits wide of the fixed-point values,
which represents the bit-width of the quantization,
the floating-point values representing the i-layers are mapped to the minimum of the fixed-point values.
Formula (3) calculating the maximum value of the floating point value mapped to the fixed point value;
wherein
The floating point values representing the i layers are mapped to the maximum of the fixed point values. Calculating the decimal bit width of the optimal fixed point value through the constraint condition of the formula (4)
。
The constraint of equation (4) is such that the range of fixed point values is as close as possible to the range of floating point values to reduce the loss of accuracy caused by quantization;
constraint conditions are as follows:
wherein
It is shown that the absolute value is calculated,
represents the maximum value of the floating point of the i layer,
represents the minimum value of the floating point of the i layer,
， 
and counting the maximum value and the minimum value of the floating point values of the i layers.
The quantization value of each layer can be obtained through formula (5);
wherein
A quantization factor representing the i-layer floating-point values,
the floating-point values of the i-layers are represented,
represents the i-layer quantized values,
it is meant to round-off the process,
representing the minimum value after quantization to the integer,
representing the maximum value after quantization to an integer,
indicating a truncation operation.
Step S4-2: respectively carrying out inverse quantization on the quantized input value and the convolution kernel, carrying out activation operation on the inverse quantized value through an activation layer, carrying out pooling operation on the activated value through a pooling layer to obtain a first convolution layer output inverse quantized value, and carrying out shift operation on a bit on the storage equipment based on the digital domain to obtain a first convolution layer output inverse quantized shifted value;
in particular, in the digital domain by
Back (corresponding to shift operation on hardware storage device, shift right)
After bit) to obtain a quantized value of the convolutional layer output
. The right shift operation realizes the remaining operation of formula (7) in step S4, and finally obtains the pooled value through the activation operation and the maximum pooling operation;
respectively carrying out inverse quantization on the quantized input value and the convolution kernel through a formula (6), and then carrying out convolution operation on the inverse quantized input value and the inverse quantized convolution kernel to obtain an inverse quantized floating point value of a first convolution output layer;
wherein,
representing the input value after the quantization,
representing the input value after the inverse quantization,
representing the quantized convolution kernel or kernels and,
representing the convolution kernel after the inverse quantization,
representing the quantized value of the first convolutional output layer,
representing the inverse quantized value of the first convolutional output layer.
The formula (7) can be derived from the formula (6), and the quantized value of the first convolution output layer can be found by performing the shift operation using the formula (7)
；
Step S4-3: activating the inversely quantized and shifted value through an activation layer, performing pooling operation on the activated value through a pooling layer, taking the pooled value as an input value of a next convolution quantization layer until the pooled output corresponding to the last convolution layer, obtaining a classification prediction result of a training set image through a full connection layer from the pooled output, performing back propagation according to an error between the prediction result and a training set true value, and training a neural network model based on a resistive random access memory;
specifically, the floating point value after inverse quantization of the first convolution output layer is input into the next layer as the input of the next layer. By analogy, floating point values of the full connection layer can be obtained, then the output of the network is obtained through the softmax classifier, the error between the network output and the correct category of the artificial mark is solved, and the error is propagated reversely. And finally obtaining the neural network model after the quantitative perception training.
Step S5: and inputting the test set image into the trained neural network to perform forward reasoning test.
Specifically, the neural network model after the quantitative perception training is mapped to a ReRAM memristor, and a test set is input to perform a forward reasoning test. The specific steps are as shown in fig. 5, wherein V in fig. 5 represents a voltage value, G represents a conductance value, and I represents a current value.
Step S5-1: the quantized values of the input layer obtained in step S3 and step S4 are input to the test set
Quantized value of the first convolution layer
Performing convolution operation to obtain the quantized value output by the first convolution layer
Inputting the quantized value of the layer
Mapping to voltage value of the resistive random access memory, and quantizing the first convolution layer
Mapping to conductance value of resistive random access memory, and performing convolution operation
Mapping the result of (1) to a current value output by the resistive random access memory;
step S5-2: converting the current value into a voltage value, performing displacement operation on the bit corresponding to the voltage value on the storage device based on the voltage value to obtain a value after the first convolution layer outputs quantitative displacement, performing activation operation on the value after quantitative displacement through an activation layer, performing pooling operation on the activated value through a pooling layer, taking the pooled value as an input value of the next convolution layer until pooling output corresponding to the last convolution layer is reached, and obtaining a classification result of the test set image through a full connection layer by the last pooling output.
Specifically, the current value output in step S5-1 is converted into a voltage by the ADC, then converted into a numerical value, and finally divided by the numerical value in the digital domain
Back (i.e. shift operation on the hardware storage device, shift right)
After bit) to obtain a quantized value of the convolutional layer output
. The right shift operation implements the remaining operations in equation (7) of step 4, and finally the pooled value is obtained by the activation operation and the maximum pooling operation.
And by analogy, obtaining the quantized value of the full connection layer, and taking the index of the maximum value of the quantized value of the full connection layer as the category of the network prediction. Wherein the pooling layer and the full-connectivity layer are implemented in software.
The effect of the present invention is further explained by combining the simulation experiment as follows:
1. simulation conditions are as follows:
the simulation experiment of the invention is carried out under the hardware environment of NVIDIA GV100 and the software environment of Pytrch 1.5.
2. Simulation content and result analysis:
a classification problem for a washion mini dataset. In the histogram shown in fig. 6, the dark gray histogram represents the classification result of forward inference on the test set by the floating point precision model, and the light gray histogram represents the classification result of forward inference on the test set by the 8-bit quantized model of the present invention. As can be seen from the figure, the result of testing the test set by using the floating point model is compared with the test result of the quantization method based on the resistive random access memory, and the difference between the identification precision of each type of the test set and the identification precision of the test set based on the quantization method is small. Table 1 shows the average recognition accuracy of the two methods to the test set, and it can be seen that the quantization method based on the resistive random access memory of the present invention has almost no accuracy loss, but can accelerate the inference speed of the model.
Table 1: floating point model and model precision comparison table after quantization of the invention
| Test method | Average recognition accuracy |
| Reasoning test of test set by floating point model | 0.8864 |
| The invention carries out reasoning test on the test set | 0.8852 |
In summary, the invention provides a neural network model quantization method based on a resistive random access memory, the method combines the characteristics of the conductance range of a ReRAM device and the limited bit width of quantization input of each layer, and the limited bit width is required to store a convolution kernel due to the limited conductance range of the ReRAM device. The quantization bit width of each layer input is limited, and the limited bit width is needed to store the output value after convolution. According to the method, the bit width of the quantization factor is optimized by designing the constraint condition, so that the quantization factor adopts the optimal power of 2, only limited digits need to be shifted to the right in the operation process of scaling after convolution, and the operation is simple. The precision loss caused by the right shift of ADC (analog-to-digital converter) is reduced. And meanwhile, quantization perception training is carried out, so that the loss of precision caused by quantization is reduced, and the reasoning speed of the model is improved. For the fast mnist dataset classification, the 8-bit quantization precision loss is below 0.5 percentage points compared to the floating point precision.
Corresponding to the embodiment of the neural network image classification method based on the resistive random access memory, the invention also provides an embodiment of the neural network image classification device based on the resistive random access memory.
Referring to fig. 7, the neural network image classification device based on the resistive random access memory provided in the embodiment of the present invention includes one or more processors, and is configured to implement the neural network image classification method based on the resistive random access memory in the above embodiment.
The embodiment of the neural network image classification device based on the resistive random access memory can be applied to any equipment with data processing capability, and the equipment with data processing capability can be equipment or devices such as computers. The device embodiments may be implemented by software, or by hardware, or by a combination of hardware and software. The software implementation is taken as an example, and as a logical device, the device is formed by reading corresponding computer program instructions in the nonvolatile memory into the memory for running through the processor of any device with data processing capability. In terms of hardware, as shown in fig. 7, the present invention is a hardware structure diagram of any device with data processing capability in which the neural network image classification apparatus based on a resistive random access memory is located, except for the processor, the memory, the network interface, and the nonvolatile memory shown in fig. 7, in the embodiment, any device with data processing capability in which the apparatus is located may generally include other hardware according to the actual function of the any device with data processing capability, which is not described again.
The implementation process of the functions and actions of each unit in the above device is specifically described in the implementation process of the corresponding step in the above method, and is not described herein again.
For the device embodiments, since they substantially correspond to the method embodiments, reference may be made to the partial description of the method embodiments for relevant points. The above-described embodiments of the apparatus are merely illustrative, and the units described as separate parts may or may not be physically separate, and parts displayed as units may or may not be physical units, may be located in one place, or may be distributed on a plurality of network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the scheme of the invention. One of ordinary skill in the art can understand and implement it without inventive effort.
Embodiments of the present invention further provide a computer-readable storage medium, on which a program is stored, where the program, when executed by a processor, implements the neural network image classification method based on a resistance random access memory in the foregoing embodiments.
The computer readable storage medium may be an internal storage unit, such as a hard disk or a memory, of any data processing capability device described in any of the foregoing embodiments. The computer readable storage medium may also be any external storage device of a device with data processing capabilities, such as a plug-in hard disk, a Smart Media Card (SMC), an SD Card, a Flash memory Card (Flash Card), etc. provided on the device. Further, the computer readable storage medium may include both internal storage units and external storage devices of any data processing capable device. The computer-readable storage medium is used for storing the computer program and other programs and data required by the arbitrary data processing-capable device, and may also be used for temporarily storing data that has been output or is to be output.
The above examples are only intended to illustrate the technical solution of the present invention, but not to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical solutions described in the foregoing embodiments may still be modified, or some or all of the technical features may be equivalently replaced; and the modifications or the substitutions do not make the essence of the corresponding technical solutions depart from the scope of the technical solutions of the embodiments of the present invention.
Claims (10)
1. The utility model provides a neural network image classification system based on resistive random access memory, includes input layer, a set of convolution layer and the full connection layer that connects gradually, its characterized in that: the convolution layer is matched with a convolution layer to be provided with a convolution quantization layer, a convolution inverse quantization layer, an activation layer and a pooling layer, the input layer is used for obtaining a training set image, the convolution quantization layer quantizes an input value of the input layer and convolution of a first convolution layer to obtain a quantized input value and a convolution kernel, the convolution inverse quantization layer dequantizes the quantized input value and the convolution kernel, the dequantized input value and the convolution kernel are subjected to convolution operation to obtain a first convolution layer output dequantized value, shift operation is carried out based on a digital domain to obtain a first convolution layer output dequantized shifted value, the activation layer is used for carrying out activation operation on the dequantized shifted value, the activated value is subjected to pooling operation, the pooled value is used as an input value of a next convolution layer until the pooled output corresponding to the last convolution layer, and the final pooled output is subjected to full-connection layer to obtain a classification prediction result of the training set image, according to the error between the prediction result and the true value of the training set, carrying out back propagation, training a neural network model based on the resistive random access memory, skipping a quantization layer in the back propagation process, directly transmitting back to a value before quantization, and optimizing network parameters by updating the weight of the value before quantization; inputting an image to be classified into a trained system, quantizing the convolution of an input layer and a first volume layer through a convolution quantization layer, performing convolution operation on the quantized value of the obtained input layer and the quantized value of the first volume layer to obtain the quantized value output by the first volume layer, mapping the quantized value of the input layer to a voltage value of a resistive random access memory, mapping the quantized value of the first volume layer to a conductance value of the resistive random access memory, mapping the result of the convolution operation to a current value output by the resistive random access memory, converting the current value into the voltage value, performing shift operation based on the voltage value to obtain a value after the first volume layer outputs quantization shift, performing activation operation on the quantized value through an activation layer, performing pooling operation on the activated value through a pooling layer, using the pooled value as the input value of a next volume layer until the pooled output corresponding to the last volume layer, and outputting the final pooled output through a full connection layer to obtain a classification result of the image to be classified.
2. The neural network image classification system based on the resistive random access memory according to claim 1, wherein: the resistive random access memory forms a resistive random access memory array, the quantized value of the input layer is mapped into the voltage value of the resistive random access memory and is input into a first row of resistive random access memories, the quantized value of the winding layer is mapped into the conductance value of each resistive random access memory, and the current value output by each row of resistive random access memories is the convolution operation of the quantized value input by the row and the quantized value of the winding layer.
3. The neural network image classification system based on the resistive random access memory according to claim 1, wherein: the convolution layers comprise a first convolution layer and a second convolution layer, and a convolution quantization layer, a convolution inverse quantization layer, an activation layer and a pooling layer are respectively matched with the first convolution layer and the second convolution layer.
4. The resistive-switching-memory-based neural network image classification system according to one of claims 1 to 3, wherein: the quantization process is as follows:
formula (1) represents a floating-point convolution operation;
wherein,
a floating-point value representing the input layer,
a floating-point value representing a convolution kernel of the first convolution layer,
represents a convolution operation; respectively mapping the floating point value of the input layer and the floating point value of the convolution kernel of the first convolution layer to a fixed point value, and determining the decimal bit width of the optimal fixed point value through a formula (2), a formula (3) and a formula (4);
formula (2) calculating the minimum value of the floating point value mapped to the fixed point value;
where i represents the number of layers of the neural network model,
the floating-point values representing the ith layer are mapped to fractional bits wide of the fixed-point values,
which represents the bit-width of the quantization,
a minimum value representing that the floating point value of the ith layer is mapped to the fixed point value;
formula (3) calculating the maximum value of the floating point value mapped to the fixed point value;
wherein
The floating point value representing the ith layer is mapped to the maximum value of the fixed point value;
calculating the decimal bit width of the optimal fixed point value through the constraint condition of the formula (4)
；
Constraint conditions are as follows:
(4)
wherein
It is shown that the absolute value is calculated,
represents the maximum value of the floating point of the ith layer,
represents the minimum value of the floating point of the ith layer,
，
the method comprises the steps of obtaining a maximum value and a minimum value of an ith layer floating point value through statistics;
solving the quantization value of each layer through a formula (5);
wherein
A quantization factor representing the ith layer floating point value,
a floating-point value representing the i-th layer,
represents the quantized value of the ith layer,
which represents the operation of rounding off,
representing the minimum value after quantization to the integer,
representing the maximum value after quantization to the integer,
indicating a truncation operation.
5. The resistive-switching-memory-based neural network image classification system according to one of claims 1 to 3, wherein: respectively carrying out inverse quantization on the quantized input value and the convolution kernel, and carrying out convolution operation on the inverse quantized input value and the inverse quantized convolution kernel through a formula (6) to obtain an inverse quantized floating point value output by the first convolution layer;
wherein,
representing the input value after the quantization,
representing the input value after the inverse quantization,
representing the quantized convolution kernel or kernels and,
representing the convolution kernel after the inverse quantization,
representing the quantized values of the first convolutional output layer,
representing the inverse quantized value of the first convolution output layer;
equation (7) is derived from equation (6):
performing shift operation by formula (7) to obtain the quantized value output by the first convolution layer
，
Representing the minimum value after quantization to the integer,
representing the maximum value after quantization to the integer,
indicating a truncation operation.
6. A neural network image classification method based on a resistive random access memory is characterized by comprising the following steps:
step S1: normalizing the image to be classified to obtain a normalized image;
step S2: constructing a training set and a test set for the normalized image;
step S3: constructing a neural network model based on a resistive random access memory;
step S4: inputting the training set into a neural network model based on a resistive random access memory, and performing quantitative perception training to obtain model parameters after quantitative perception training, wherein the method comprises the following steps:
step S4-1: quantizing the input value of the input layer and the convolution kernel of the first convolution layer to obtain a quantized input value and a convolution kernel;
step S4-2: respectively carrying out inverse quantization on the quantized input value and the convolution kernel, carrying out convolution operation on the inverse quantized input value and the convolution kernel to obtain a first convolution layer output inverse quantized value, and carrying out shift operation based on a digital domain to obtain a first convolution layer output inverse quantized shifted value;
step S4-3: activating the inversely quantized and shifted value through an activation layer, performing pooling operation on the activated value through a pooling layer, taking the pooled value as an input value of a next convolution and quantization layer until the pooled output corresponding to the last convolution layer, obtaining a classification prediction result of a training set image through a full connection layer, performing back propagation according to an error between the prediction result and a training set true value, training a neural network model based on a resistive random access memory, directly returning a skipped quantization layer to a value before quantization in the back propagation process, and optimizing network parameters by updating the weight of the value before quantization;
step S5: inputting the test set image into the trained neural network for forward reasoning test, comprising the following steps:
step S5-1: taking the test set as input, performing convolution operation on the quantized value of the input layer and the quantized value of the first convolution layer obtained in the steps S3 and S4 to obtain a quantized value output by the first convolution layer, mapping the quantized value of the input layer to a voltage value of the resistive random access memory, mapping the quantized value of the first convolution layer to a conductance value of the resistive random access memory, and mapping the result of the convolution operation to a current value output by the resistive random access memory;
step S5-2: converting the current value into a voltage value, performing a shift operation based on the voltage value to obtain a value after the first convolution layer outputs the quantization shift, performing an activation operation on the quantized value through an activation layer, performing a pooling operation on the activated value through a pooling layer, taking the pooled value as an input value of the next convolution layer until the pooled output corresponding to the last convolution layer is output, and obtaining a classification result of the test set image through a full connection layer by the last pooled output.
7. The neural network image classification method based on the resistive random access memory according to claim 6, characterized in that: the specific quantification method of step S4-1 includes the following steps:
formula (1) represents a convolution operation of floating points;
wherein,
a floating-point value representing the input layer,
a floating-point value representing a convolution kernel of the first convolution layer,
represents a convolution operation; respectively mapping the floating point value of the input layer and the floating point value of the convolution kernel of the first convolution layer to a fixed point value, and determining the decimal bit width of the optimal fixed point value through a formula (2), a formula (3) and a formula (4);
formula (2) calculating the minimum value of the floating point value mapped to the fixed point value;
where i represents the number of layers of the neural network model,
mapping floating point values representing the ith layer to fixed pointsThe fractional bit width of the point values,
which represents the bit-width of the quantization,
a minimum value representing that the floating point value of the ith layer is mapped to the fixed point value;
formula (3) calculating the maximum value of the floating point value mapped to the fixed point value;
wherein
The floating point value representing the ith layer is mapped to the maximum value of the fixed point value;
calculating the decimal bit width of the optimal fixed point value through the constraint condition of the formula (4)
；
Constraint conditions are as follows:
(4)
wherein
It is shown that the absolute value is calculated,
represents the maximum value of the floating point of the ith layer,
represents the minimum value of the floating point of the ith layer,
，
the method comprises the steps of obtaining a maximum value and a minimum value of an ith layer floating point value through statistics;
solving the quantization value of each layer through a formula (5);
wherein
A quantization factor representing the ith layer floating point value,
a floating point value representing the ith layer,
represents the quantized value of the ith layer,
which represents the operation of rounding off the object,
representing the minimum value after quantization to the integer,
representing the maximum value after quantization to the integer,
indicating a truncation operation.
8. The neural network image classification method based on the resistive random access memory according to claim 6, characterized in that: in the step S4-2, performing inverse quantization on the quantized input value and the convolution kernel, and then performing convolution operation on the inverse-quantized input value and the inverse-quantized convolution kernel through a formula (6) to obtain an inverse-quantized floating point value output by the first convolution layer;
wherein,
representing the value of the input after the quantization,
representing the input value after the inverse quantization,
representing the quantized convolution kernel or kernels and,
representing the convolution kernel after the inverse quantization,
representing the quantized values of the first convolutional output layer,
representing the inverse quantized value of the first convolution output layer;
equation (7) is derived from equation (6):
performing shift operation by formula (7) to obtain the quantized value output by the first convolution layer
，
Representing the minimum value after quantization to an integer,
representing the maximum value after quantization to the integer,
indicating a truncation operation.
9. The neural network image classification method based on the resistive random access memory according to claim 6, characterized in that: in step S5-2, a resistive random access memory is constructed to form a resistive random access memory array, the quantized values of the input layer are mapped to voltage values of the resistive random access memory and input to the first row of resistive random access memory, the quantized values of the winding layer are mapped to conductance values of the resistive random access memory, and the current value output by each row of resistive random access memory is a convolution operation between the quantized values input by the row and the quantized values of the winding layer.
10. A neural network image classification device based on a resistive random access memory, which is characterized by comprising one or more processors and is used for realizing the neural network image classification method based on the resistive random access memory as claimed in any one of claims 6 to 9.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210579664.2A CN114677548B (en) | 2022-05-26 | 2022-05-26 | Neural network image classification system and method based on resistive random access memory |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210579664.2A CN114677548B (en) | 2022-05-26 | 2022-05-26 | Neural network image classification system and method based on resistive random access memory |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN114677548A true CN114677548A (en) | 2022-06-28 |
| CN114677548B CN114677548B (en) | 2022-10-14 |
Family
ID=82080811
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210579664.2A Active CN114677548B (en) | 2022-05-26 | 2022-05-26 | Neural network image classification system and method based on resistive random access memory |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114677548B (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115311506A (en) * | 2022-10-11 | 2022-11-08 | 之江实验室 | Image classification method and device based on quantization factor optimization of resistive random access memory |
| CN115905546A (en) * | 2023-01-06 | 2023-04-04 | 之江实验室 | Graph convolution network document identification device and method based on resistive random access memory |
| CN116561050A (en) * | 2023-04-07 | 2023-08-08 | 清华大学 | Fine granularity mapping method and device for RRAM (remote radio access memory) integrated chip |
Citations (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| GB201704751D0 (en) * | 2017-03-24 | 2017-05-10 | Imagination Tech Ltd | Floating point to fixed point conversion |
| US20190042948A1 (en) * | 2017-08-04 | 2019-02-07 | Samsung Electronics Co., Ltd. | Method and apparatus for generating fixed-point quantized neural network |
| CN110363281A (en) * | 2019-06-06 | 2019-10-22 | 上海交通大学 | A kind of convolutional neural networks quantization method, device, computer and storage medium |
| CN111260048A (en) * | 2020-01-14 | 2020-06-09 | 上海交通大学 | Method for realizing activation function in neural network accelerator based on memristor |
| CN111382788A (en) * | 2020-03-06 | 2020-07-07 | 西安电子科技大学 | Hyperspectral image classification method based on binary quantization network |
| CN111612147A (en) * | 2020-06-30 | 2020-09-01 | 上海富瀚微电子股份有限公司 | Quantization method of deep convolutional network |
| CN111695671A (en) * | 2019-03-12 | 2020-09-22 | 北京地平线机器人技术研发有限公司 | Method and device for training neural network and electronic equipment |
| CN112561049A (en) * | 2020-12-23 | 2021-03-26 | 首都师范大学 | Resource allocation method and device of DNN accelerator based on memristor |
| CN112884133A (en) * | 2021-03-24 | 2021-06-01 | 苏州科达科技股份有限公司 | Convolutional neural network quantization method, system, device and storage medium |
| WO2021179587A1 (en) * | 2020-03-10 | 2021-09-16 | 北京迈格威科技有限公司 | Neural network model quantification method and apparatus, electronic device and computer-readable storage medium |
| CN114330694A (en) * | 2021-12-31 | 2022-04-12 | 上海集成电路装备材料产业创新中心有限公司 | Circuit and method for realizing convolution operation |
| CN114418057A (en) * | 2020-10-28 | 2022-04-29 | 华为技术有限公司 | Operation method of convolutional neural network and related equipment |
-
2022
- 2022-05-26 CN CN202210579664.2A patent/CN114677548B/en active Active
Patent Citations (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| GB201704751D0 (en) * | 2017-03-24 | 2017-05-10 | Imagination Tech Ltd | Floating point to fixed point conversion |
| US20190042948A1 (en) * | 2017-08-04 | 2019-02-07 | Samsung Electronics Co., Ltd. | Method and apparatus for generating fixed-point quantized neural network |
| CN111695671A (en) * | 2019-03-12 | 2020-09-22 | 北京地平线机器人技术研发有限公司 | Method and device for training neural network and electronic equipment |
| CN110363281A (en) * | 2019-06-06 | 2019-10-22 | 上海交通大学 | A kind of convolutional neural networks quantization method, device, computer and storage medium |
| CN111260048A (en) * | 2020-01-14 | 2020-06-09 | 上海交通大学 | Method for realizing activation function in neural network accelerator based on memristor |
| CN111382788A (en) * | 2020-03-06 | 2020-07-07 | 西安电子科技大学 | Hyperspectral image classification method based on binary quantization network |
| WO2021179587A1 (en) * | 2020-03-10 | 2021-09-16 | 北京迈格威科技有限公司 | Neural network model quantification method and apparatus, electronic device and computer-readable storage medium |
| CN111612147A (en) * | 2020-06-30 | 2020-09-01 | 上海富瀚微电子股份有限公司 | Quantization method of deep convolutional network |
| CN114418057A (en) * | 2020-10-28 | 2022-04-29 | 华为技术有限公司 | Operation method of convolutional neural network and related equipment |
| CN112561049A (en) * | 2020-12-23 | 2021-03-26 | 首都师范大学 | Resource allocation method and device of DNN accelerator based on memristor |
| CN112884133A (en) * | 2021-03-24 | 2021-06-01 | 苏州科达科技股份有限公司 | Convolutional neural network quantization method, system, device and storage medium |
| CN114330694A (en) * | 2021-12-31 | 2022-04-12 | 上海集成电路装备材料产业创新中心有限公司 | Circuit and method for realizing convolution operation |
Non-Patent Citations (2)
| Title |
|---|
| DARRYL D.LIN 等，: "Fixed point quantization of deep convolutional networks", 《PROCEEDINGS OF THE 33RD INTERNATIONAL CONFERENCE ON MACHINE LEARNING,PMLR4 8》 * |
| 孙磊 等，: "改进的基于嵌入式SoC卷积神经网络识别模型", 《计算机应用与软件》 * |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115311506A (en) * | 2022-10-11 | 2022-11-08 | 之江实验室 | Image classification method and device based on quantization factor optimization of resistive random access memory |
| CN115905546A (en) * | 2023-01-06 | 2023-04-04 | 之江实验室 | Graph convolution network document identification device and method based on resistive random access memory |
| CN116561050A (en) * | 2023-04-07 | 2023-08-08 | 清华大学 | Fine granularity mapping method and device for RRAM (remote radio access memory) integrated chip |
Also Published As
| Publication number | Publication date |
|---|---|
| CN114677548B (en) | 2022-10-14 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108701250B (en) | Data fixed-point method and device | |
| CN108345939B (en) | Neural network based on fixed-point operation | |
| CN114677548B (en) | Neural network image classification system and method based on resistive random access memory | |
| CN108510067B (en) | Convolutional neural network quantification method based on engineering realization | |
| US11593658B2 (en) | Processing method and device | |
| CN110413255B (en) | Artificial neural network adjusting method and device | |
| CN110880038B (en) | System for accelerating convolution calculation based on FPGA and convolution neural network | |
| CN109002889B (en) | Adaptive iterative convolution neural network model compression method | |
| WO2019238029A1 (en) | Convolutional neural network system, and method for quantifying convolutional neural network | |
| CN115311506B (en) | Image classification method and device based on quantization factor optimization of resistive random access memory | |
| CN110363297A (en) | Neural metwork training and image processing method, device, equipment and medium | |
| WO2020001401A1 (en) | Operation method and apparatus for network layer in deep neural network | |
| CN110874627B (en) | Data processing method, data processing device and computer readable medium | |
| CN113011532A (en) | Classification model training method and device, computing equipment and storage medium | |
| CN115564987A (en) | Training method and application of image classification model based on meta-learning | |
| CN116188878A (en) | Image classification method, device and storage medium based on neural network structure fine adjustment | |
| CN110503182A (en) | Network layer operation method and device in deep neural network | |
| CN113240090B (en) | Image processing model generation method, image processing device and electronic equipment | |
| CN114239799A (en) | Efficient target detection method, device, medium and system | |
| CN115905546B (en) | Graph convolution network literature identification device and method based on resistive random access memory | |
| CN116956997A (en) | LSTM model quantization retraining method, system and equipment for time sequence data processing | |
| US20220253709A1 (en) | Compressing a Set of Coefficients for Subsequent Use in a Neural Network | |
| CN116384471A (en) | Model pruning method, device, computer equipment, storage medium and program product | |
| Goel et al. | CompactNet: High accuracy deep neural network optimized for on-chip implementation | |
| CN116306879A (en) | Data processing method, device, electronic equipment and storage medium |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |