EP3908981A1 - Optimisation de calculs de réseau neuronal artificiel sur la base d'une détermination automatique d'une taille de lot - Google Patents

Optimisation de calculs de réseau neuronal artificiel sur la base d'une détermination automatique d'une taille de lot

Info

Publication number
EP3908981A1
EP3908981A1 EP19702705.5A EP19702705A EP3908981A1 EP 3908981 A1 EP3908981 A1 EP 3908981A1 EP 19702705 A EP19702705 A EP 19702705A EP 3908981 A1 EP3908981 A1 EP 3908981A1
Authority
EP
European Patent Office
Prior art keywords
ann
layers
batch size
layer
batch
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.)
Pending
Application number
EP19702705.5A
Other languages
German (de)
English (en)
Inventor
Benoit Chappet De Vangel
Thomas CAGNAC
Benjamin POUMAREDE
Ludovic Larzul
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Mipsology SAS
Original Assignee
Mipsology SAS
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 Mipsology SAS filed Critical Mipsology SAS
Publication of EP3908981A1 publication Critical patent/EP3908981A1/fr
Pending legal-status Critical Current

Links

Classifications

    • 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
    • G06N3/04—Architecture, e.g. interconnection topology
    • 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

Definitions

  • the present disclosure relates generally to data processing and, more particularly, to systems and methods for optimizing artificial neural network (ANN) computations based on automatic determination of a batch size.
  • ANN artificial neural network
  • ANNs Artificial Neural Networks
  • the human brain contains 10-20 billion neurons connected through synapses. Electrical and chemical messages are passed from neurons to neurons based on input information and their resistance to passing information.
  • a neuron can be represented by a node performing a simple operation of addition coupled with a saturation function.
  • a synapse can be represented by a connection between two nodes. Each connection can be associated with an operation of a multiplication by a constant.
  • the ANNs are particularly useful for solving problems that cannot be easily solved by classical computer programs.
  • ANNs While types of the ANNs may vary, they all have the same basic elements similar to the human brain.
  • a typical ANN can be organized into layers, each of the layers may include many neurons sharing similar functionality. Inputs of a layer may come from a previous layer, multiple previous layers, any other layers or even the layer itself.
  • Major architectures of ANNs include Convolutional Neural Network (CNN), Recurrent Neural Network (RNN), and Long Term Short Memory (LTSM) network, but other architectures of ANN can be developed for specific applications. While some operations have a natural sequence, for example a layer depending on previous layers, most of the operations can be carried out in parallel within the same layer. The ANNs can then be computed in parallel on many different computing elements similar to neurons of the brain.
  • a single ANN may have hundreds of layers. Each layer may include millions of connections. Thus, a single ANN may potentially require billions of simple operations like multiplications and additions.
  • ANNs can result in a very heavy load for processing units (e.g., CPU), even the ones running at high speeds.
  • processing units e.g., CPU
  • GPUs graphics processing units
  • GPUs can be used to process large ANNs because GPUs have a much higher throughput capacity of operations in comparison to CPUs. Because this approach solves, at least partially, the throughput limitation problem, GPUs appear to be more efficient in the computations of ANNs than the CPUs.
  • GPUs are not well suited to the computations of ANNs because the GPUs have been specifically designed to compute graphical images.
  • the GPUs may provide a certain level of parallelism in computations.
  • the GPUs are constraining the computations in long pipes, which results in latency and lack of reactivity.
  • very large GPUs can be used which involve excessive power consumption, a typical issue of GPUs.
  • CPUs provide a very generic engine that can execute very few sequences of instructions with a minimum effort in terms of programming, but lack the power of computing required for ANNs.
  • GPUs are slightly more parallel and require a larger effort of programming than CPUs, which can be hidden behind libraries with some performance costs, but are not very well suitable for ANNs.
  • FPGAs Field Programmable Gate Arrays
  • the FPGAs can be configured to perform computations in parallel. Therefore, FPGAs can be well suited to compute ANNs.
  • Programming FPGAs is challenging, requiring a much larger effort than programming CPUs and GPUs. Thus, adaption of FPGAs to perform ANN computations can be more challenging than for CPUs and GPUs.
  • the inputs computed with an ANN are typically provided by an artificial intelligence (AI) framework.
  • AI artificial intelligence
  • Those programs are used by the AI community to develop new ANN or global solutions based on ANN.
  • FPGAs typically lack integration with AI frameworks.
  • input data can be bundled together and processed together as well. This is known as batching input data.
  • the batching is the ability by a user to manually define the number of pieces of data that will be bundled together in a framework. Typically, there are no tools to automate this process.
  • embodiments of the present disclosure can facilitate reduction in time required for computer systems to perform ANN computations for a batch by determining batch sizes individually for each layer of the ANN and a sequence of a computation of layers for inputs sets from the batch.
  • a system for optimizing ANN computations based on automatic determination of a batch size for an ANN may include a computation engine capable of performing computations of one or more layers of the ANN and an optimization module.
  • the optimization module can be capable of receiving an ANN structure associated with the ANN and generating, based on the ANN structure, a configuration for the computation engine.
  • the configuration may include information concerning batch sizes of one or more layers of the ANN.
  • the optimization module may determine a batch size of a layer based on one or more of the following: a bandwidth required to read data related to the layer, a number of parameters associated with the layer, and a time the layer processes one input dataset from the batch.
  • the batch size of a layer can differ from the batch size of the ANN.
  • the ANN may include at least a first layer and a second layer, such that the batch size of the first layer differs from the batch size of the second layer.
  • the optimization module can be capable of determining the batch size for the ANN based on the batch sizes of the layers of the ANN.
  • the computation engine may include one or more processing units capable of performing operations associated with layers of the ANN.
  • the computation engine may further include a controller capable of configuring the processing units to repeat operations of a subpart of the ANN for different input datasets from a batch of input datasets.
  • the subpart may include connected layers of the ANN.
  • the controller may configure the processing units to perform, simultaneously, computations of a first layer of the ANN for a first input dataset of a batch of input datasets and computations of the first layer for a second input dataset of the batch of input datasets prior to computation of a second layer of the ANN, wherein an input dataset of the second layer includes an output dataset of the first layer.
  • the computation engine can be implemented on a FPGA and the
  • optimization module can be implemented as a software-based module.
  • optimization module can be capable of performing one or more iterations of selecting batch sizes of the layers of the ANN to optimize a performance measure.
  • the performance measure can be a function of one or more of: a desired batch size of ANN, a latency of the ANN, and a throughput of the ANN.
  • the optimization module can be capable of performing the iterations until a number of the iterations exceeds a predetermined threshold or the performance measure exceeds a pre-determined threshold.
  • the performance measure can be set based on a user input.
  • a method for optimizing ANN computations based on automatically determining a batch size for an ANN may include receiving, by an optimization module, an ANN structure associated with the ANN. The method may further include generating, based on the ANN structure, a configuration for a computation engine capable of performing computations of the layers of the ANN. The configuration can include information concerning a batch size of one or more layers of the ANN. The method may further include determining, by the optimization module and based on the batch size of the one or more layers of the ANN, the batch size for the ANN.
  • FIG. 1 is a block diagram showing an example system for optimizing ANN computations based on automatic determination of a batch size, according to some example embodiments.
  • FIG. 2 shows an ANN, neuron, and transfer function, according to an example embodiment.
  • FIG. 3 is a flow chart showing training and inference of an ANN, according to some example embodiments.
  • FIG. 4 is a block diagram showing a system for optimizing ANN
  • FIG. 5 is a block diagram showing an ANN, according to example
  • FIG. 6A is a plot showing a sequence of computations of layers in ANN computations, according to an example embodiment.
  • FIG. 6B is a plot showing a sequence of computations of layers in ANN computations, according to another example embodiment.
  • FIG. 7 is a flow chart showing steps of a method for determining a
  • FIG. 8 is a flow chart showing steps of a method for optimizing ANN computations based on automatic determination of a batch size, according to some example embodiments.
  • FIG. 9 shows a computing system that can be used to implement embodiments of the disclosed technology.
  • Embodiments of this disclosure are directed to methods and systems for optimizing ANN computations based on automatic determination of a batch size.
  • Embodiments of the present disclosure may facilitate acceleration of ANN computation for a batch of input datasets by assigning a batch size to each of the layers in ANN.
  • the batch size of the layer can be based on a latency of the layer, number of parameters associated with the layer, and bandwidth of the layer.
  • Some embodiments of the present disclosure may allow determining batch sizes of an ANN to optimize
  • module shall be construed to include a hardware device, software, or a combination of both.
  • a hardware-based module can use one or more microprocessors, FPGAs, application-specific integrated circuits (ASICs), programmable logic devices, transistor-based circuits, or various combinations thereof.
  • Software-based modules can constitute computer programs, computer program procedures, computer program functions, and the like.
  • a module of a system can be implemented by a computer or server, or by multiple computers or servers interconnected into a network.
  • a module may refer to a subpart of a computer system, a hardware device, an integrated circuit, or a computer program.
  • embodiments of the present disclosure can allow determining an optimal batch size for the ANN computations based on the structures of ANN and the architecture of the computer system configured to perform computations of the ANN.
  • FIG. 1 is a block diagram showing an example system 100 for optimizing ANN computations based on automatic determination of a batch size, according to some example embodiments.
  • the system 100 can be part of a computing system, such as a personal computer, a server, a cloud-based computing recourse, and the like.
  • the system 100 may include one or more FPGA boards 105 and a chipset 135 including at least one CPU.
  • the chipset 135 can be communicatively connected to the FPGA boards 105 via a communication interface.
  • the communication interface may include a Peripheral Component Interconnect Express (PCIE) standard 130.
  • PCIE Peripheral Component Interconnect Express
  • the communication interface may also include an Ethernet connection 131.
  • the FPGA board 105 may include an FPGA 115, a volatile memory 110, and a non-volatile memory 120.
  • the volatile memory 110 may include a double data rate synchronous dynamic random-access memory (DDR SDRAM), High Bandwidth Memory (HBM), High Bandwidth Cache (HBC), graphics DDR SDRAM, or any other type of memory.
  • the volatile memory 110 may include the host memory.
  • the non volatile memory 120 may include Electrically Erasable Programmable Read-Only Memory (EEROM), a solid-state drive (SSD), a flash memory, and so forth.
  • EEROM Electrically Erasable Programmable Read-Only Memory
  • SSD solid-state drive
  • flash memory and so forth.
  • the FPGA 115 can include blocks.
  • the blocks may include a set of elementary nodes (also referred to as gates) performing basic hardware operations, such as Boolean operations.
  • the blocks may further include registers retaining bit information, one or more memory storage of different sizes, and one or more digital signal processors (DSPs) to perform arithmetic computations, for example, additions and multiplications.
  • DSPs digital signal processors
  • Programming FPGA 115 may include configuring each of the blocks to exhibit an expected behavior and connecting the blocks by routing information between the blocks. Programming FPGA 115 can be carried out using a result from a compiler taken as an input schematic description, gate-level description, hardware languages like Verilog, System Verilog, or Very High Speed Integrated Circuit Hardware Description Language (VHDL), or any combination of thereof.
  • VHDL Very High Speed Integrated Circuit Hardware Description Language
  • the non-volatile memory 120 can be configured to store instructions in a form of bit file 125 to be executed by the FPGA 115.
  • the FPGA 115 can be configured by the instructions to perform one or more floating point operations including multiplication and addition to calculate a sum of products that can be used in neural network computations.
  • the volatile memory 110 can be configured to store weights W[i] for neurons of one or more ANNs, input values V[i] to be processed for the ANNs, and results of ANNs computation including any intermediate results of computations of layers of the ANNs.
  • FIG. 2 shows ANN 210, neuron 220, and transfer function 230, according to some example embodiments.
  • the ANN 210 may include one or more input layers 240, one or more hidden layers 250, and one or more output layers 260.
  • Each of the input layers, hidden layers, and output layers may include one or more (artificial) neurons 220. The number of neurons can be different for different layers.
  • Each of neurons 220 may be represented by a calculation of the following mathematical function:
  • V[i] are neuron input values
  • W[i] are weights assigned to input values at the neuron
  • F(X) is a transfer function.
  • the transfer function 230 F(X) is selected to be zero for X ⁇ 0 and have a limit of zero as X approaches zero.
  • the transfer function F(X) can be in the form of a sigmoid.
  • the result of the calculation of a neuron propagates as an input value of further neurons in the ANN.
  • the further neurons can belong to either the next layer, previous layer, or the same layer.
  • ANN 210 illustrated in FIG. 2 can be referred to as a feedforward neural network
  • embodiments of the present disclosure can be also used in computations of convolutional neural networks, recurrent neural networks, long short-term memory networks, and other types of ANNs.
  • FIG. 3 is a flow chart showing training 310 and inference 325 of an ANN, according to some example embodiments.
  • the training 310 (also known as learning) is a process of teaching ANN 305 to output a proper result based on a given set of training data 315.
  • the process of training may include determining weights 320 of neurons of the ANN 305 based on training data 315.
  • the training data 315 may include samples. Each sample may be represented as a pair of input values and expected output.
  • the training data 315 may include hundreds to millions of samples. While training 310 is needed to be performed only once, it may require a significant number of computations and take considerable time.
  • the ANNs can be configured to solve different tasks including, for example, image recognition, speech recognition, handwriting recognition, machine
  • the inference 325 is a process of computation of an ANN.
  • the inference 325 can use the trained ANN weights 320 and new data 330 including new sets of input values. For each new set of input values, the computation of the ANN provides a new output which answer the problem that the ANN is supposed to solve.
  • an ANN can be trained to recognize various animals in images.
  • the ANN can be trained using millions of images of animals. Submitting a new image to the ANN would provide the information for animals in the new image (this process is known as image tagging). While the inference for each image takes fewer computations than training, number of inferences can be large because new images can be received from billions of sources.
  • the inference 325 includes multiple computations of the sum of the following products:
  • V[i] are new input values and W[i] are weights associated with neurons of the ANN.
  • a computational engine (residing, for example, in a computing cloud) configured to perform ANN computations can be configured to process a batch of input datasets, for example, a sequence of images, a sequence of voice recordings, a sequence of videos, a sequence of handwritten messages, and so forth.
  • the time of processing of the batch of input datasets can be affected by the latency of ANN.
  • the latency can be defined as the time the ANN processes one input dataset from the batch.
  • an optimal batch size can be determined for each layer of the ANN to reduce latency of the ANN and determine an optimal batch size for the ANN.
  • FIG. 4 is a block diagram showing a system 400 for optimizing ANN computations based on automatic determination of a batch size, according to some example embodiments.
  • the system 400 may include a computational engine 405 configured to perform ANN computations.
  • the computational engine 405 may further include a memory controller and a plurality of memories to store data associated with the ANN and input dataset for ANN computations.
  • the system 400 may further include an optimization module 440 configured to generate a configuration 450 for the computational engine 405.
  • the controller 420 of the computational engine 405 may assign, based on the configuration 450, batch sizes to layers of ANN.
  • the controller 420 may receive, via the communication unit 410, a batch 460 of input datasets for the ANN.
  • FIG. 5 shows an example ANN 500.
  • the ANN 500 may include an input layer 505, hidden layers 510, 515, 520, 525, 530, and output layer 535.
  • the batch size of ANN is equal to 4.
  • the batch of ANN can include a sequence of input datasets A, B, C, and D.
  • the hidden layers 510, 515, and 520 can be assigned a batch size 1 and the hidden layers 525 and 530 can be assigned a batch size 2.
  • the output layer 535 can be assigned a batch size 4.
  • the batch of the hidden layer 525 may include two input datasets. Inputs of the hidden layer 525 are outputs of the hidden layer 515. Because the hidden layer 515 has batch size 1, the hidden layer 515 (and previous hidden layer 510) may be required to be executed twice to obtain two inputs for the batch of the hidden layer 525.
  • the batch of the hidden layer 530 includes two inputs. Each input of the hidden layer 530 includes an output from the hidden layer 520 and an output of the hidden layer 525. Because the hidden layer 520 has batch size 1, the hidden layer 520 (and previous hidden layers 510 and 515) is required to be executed twice to obtain two inputs for the batch of the hidden layer 530. Because the hidden layer 525 has batch size 2, the same as the batch size of the hidden layer 530, the hidden layer 525 is required to be executed once to obtain two inputs for the batch of the hidden layer 530.
  • the batch of the output layer 535 includes four inputs. Inputs for the output layer 535 are outputs of the hidden layer 530. Because the batch of the hidden layer 530 is two, the hidden layer 530 is required to be executed twice to fill the batch of the output layer 535.
  • FIG. 6A shows a plot 610 of a sequence of computation of layers of the ANN 500, according to an example embodiment.
  • the subpart of ANN 500 including the hidden layers 510 (Lo), 515 (Li), and 520 (L2) is first executed twice, once for input dataset A and once for the input dataset B.
  • the hidden layer 525 (L 3 ) can be executed once using the batch including two outputs of the hidden layer 515 (Li), where the two outputs are generated based on input dataset A and B.
  • the hidden layer 530 (L4) can be executed once using batch of two inputs, where a first input is based on of the outputs of layers 520 (L2) and 525 (L 3 ) generated based on input dataset A and a second input is based on the outputs of layers 520 (L2) and 525 (L 3 ) generated based on input dataset B.
  • the whole sequence of computations of layers Lo, Li, L2, L 3 , and L4 is further repeated one more time for the input datasets C and D.
  • the layer L4 generates four outputs which are based on the input datasets A, B, C, and D.
  • the four outputs of the layer L4 may form a batch for computation of the output layer 535 (Ls).
  • FIG. 6B shows a plot 620 of a sequence of computation of layers of the ANN 500, according to another example embodiment.
  • the layers Lo, Li, L2 are executed first to obtain outputs based on the input datasets A, B, C, and D in the batch of ANN 500.
  • the layer L3 is then executed twice for two batches, wherein the first batch includes two outputs of layer L 2 generated based on input datasets A and B, and the second batch includes two outputs of layer L 2 generated based on input datasets C and D.
  • the layer L4 is then executed twice for two batches, wherein the first batch incudes outputs of previous layers L 2 and L 3 generated based on input datasets A and B and the second batch includes outputs of previous layers L2 and L 3 generated based on input datasets C and D. Being executed twice, the layer L4 generates four outputs which are based on the input datasets A, B, C, and D. The four outputs of the layer L 4 may form a batch for computation of the output layer 535 (Ls).
  • the sequence of the computation of layers can depend on latencies and throughputs of memories storing input data of layers and other information related to the layers. Due to the latencies and throughputs, the sequence shown in FIG. 6B can be a better choice than the sequence described in FIG. 6A, even though results of the execution of the sequences can be the same.
  • the computational engine 405 can be configured to receive a request for ANN computations.
  • the request can be received from a user.
  • the request can be received from an application running on a computer system in communication with the computational engine 405.
  • the request can include a desired batch size, desired latency for ANN, and desired throughput of the ANN.
  • the computational engine 405 Prior to performing the ANN computations, the computational engine 405 can be configured by performing the following steps:
  • the batch allocation may include selection of batch sizes for layers in the ANN.
  • the memory allocation may include generating a memory configuration for the data associated with the ANN.
  • configuration may indicate which memories in computational engine 405 should be used for storing data and parameters associated with layers and input datasets of the ANN.
  • FIG. 7 is a flow chart showing a method 700 for determining a configuration of batch sizes of layers in ANN, according to some example embodiments.
  • the method 700 may be performed by optimization module 440.
  • the method 700 may commence in block 740 with selecting a batch size for each of the layers in the ANN.
  • the selection of a batch size for a layer can be based on the number of parameters of the layer, time of execution of the layer, and bandwidth required to fetch and store the layer-related data and parameters.
  • the selection of batch sizes of the layers of ANN can be based on a heuristic algorithm.
  • the method 700 may determine a performance measure.
  • the performance measure can be a function an ANN batch size, latency of ANN, and throughput of the ANN.
  • the ANN batch size, latency of the ANN, and throughput of the ANN can be estimated based on the batch sizes of layers determined in block 740.
  • the method 700 may determine whether the performance measure satisfies criteria.
  • the criteria may be based on comparison of the performance measure and the desired performance measure.
  • the desired performance measure can be a function of one or more of the desired batch size 710 for ANN, the desired latency 720 of ANN, and the desired throughput 730 of ANN.
  • the criteria can be also based on a comparison of a number of iterations of selection of the batch sizes for the layers in the ANN to a maximum number of iterations.
  • the blocks 740, 750, and 760 can be iterated to find a configuration of batch sizes of layers of the ANN corresponding to an optimal performance measure.
  • the performance measure can be tuned by a user. The user may indicate whether the performance measure is to correspond to a minimum latency of the ANN or a maximum throughput of the ANN.
  • the batch size for the ANN can be then selected as a maximum of batch sizes of the layers of the ANN.
  • the method 700 can proceed with repeating blocks 740 and 750. If the criteria are met, the method 700 may proceed to block 770. In the block 770, the method 700 may generate a configuration for the computation engine 405. The configuration may include information regarding batch sizes for the layers in the ANN, batch size for the ANN, instructions indicating an order in which the layers of the ANN should be executed.
  • FIG. 8 is a flow chart illustrating a method 800 for optimizing ANN
  • the operations may be combined, performed in parallel, or performed in a different order.
  • the method 800 may include additional or fewer operations than those illustrated.
  • the method 800 may be performed by system 400 described above with reference to in FIG. 4. [0054]
  • the method 800 may commence with receiving, an optimization module and an ANN structure associated with an ANN.
  • the method 800 may generate, based on the ANN structure, a configuration for a computation engine capable of performing computations of the layers of the ANN.
  • the configuration may include information concerning a batch size of one or more layers of the ANN and a sequence of performing computations for the layers of ANN.
  • the optimization module can include a software-based module and the computation engine may include one or more hardware-based modules implemented on FPGAs.
  • the batch size of a layer of the ANN can be determined based on a bandwidth required to read data related to the layer, a number of parameters associated with the layer, and a time the layer processes one input dataset from the batch.
  • the batch size of the layer of the ANN can differ from the batch size of the ANN.
  • the ANN may include at least a first layer and a second layer such as a batch size of the first layer differs from a batch size of the second layer.
  • Determining the batch sizes for layers of the ANN may include performing, by the optimization module, one or more iterations of selecting batch sizes of the layers of the ANN to optimize a performance measure.
  • the performance measure can be a function of one or more of: a latency of the ANN, a throughput of the ANN, and a desired size of ANN.
  • the performance measure can be set based on a user input. The iterations can be carried out until a number of the iterations exceeds a predetermined threshold or the performance measure exceeds a pre-determined threshold.
  • the method 800 may determine, by the optimization module and based on the batch sizes of the layers of the ANN, the batch size for the ANN.
  • the batch size of the ANN may be based on the maximum value of the batch sizes of the layers.
  • FIG. 9 illustrates an example computing system 900 that may be used to implement embodiments described herein.
  • the example computing system 900 of FIG. 9 may include one or more processors 910 and memory 920.
  • Memory 920 may store, in part, instructions and data for execution by the one or more processors 910.
  • Memory 920 can store the executable code when the exemplary computing system 900 is in operation.
  • the processor 910 may include internal accelerators like a graphical processing unit, a Field Programmable Gate Array, or similar accelerators that may be suitable for use with embodiments described herein.
  • the memory 920 may include internal accelerators like a graphical processing unit, a Field Programmable Gate Array, or similar accelerators that may be suitable for use with embodiments described herein.
  • the example computing system 900 of FIG. 9 may further include a mass storage 930, portable storage 940, one or more output devices 950, one or more input devices 960, a network interface 970, and one or more peripheral devices 980.
  • the components shown in FIG. 9 are depicted as being connected via a single bus 990.
  • the components may be connected through one or more data transport means.
  • the one or more processors 910 and memory 920 may be connected via a local microprocessor bus, and the mass storage 930, one or more peripheral devices 980, portable storage 940, and network interface 970 may be connected via one or more input/output buses.
  • Mass storage 930 which may be implemented with a magnetic disk drive, an optical disk drive or a solid state drive, is a non-volatile storage device for storing data and instructions for use by a magnetic disk, an optical disk drive or SSD, which in turn may be used by one or more processors 910. Mass storage 930 can store the system software for implementing embodiments described herein for purposes of loading that software into memory 920.
  • the mass storage 930 may also include internal accelerators like a graphical processing unit, a Field Programmable Gate Array, or similar accelerators that may be suitable for use with embodiments described herein.
  • Portable storage 940 may operate in conjunction with a portable non-volatile storage medium, such as a compact disk (CD) or digital video disc (DVD), to input and output data and code to and from the computing system 900 of FIG. 9.
  • a portable non-volatile storage medium such as a compact disk (CD) or digital video disc (DVD)
  • CD compact disk
  • DVD digital video disc
  • the system software for implementing embodiments described herein may be stored on such a portable medium and input to the computing system 900 via the portable storage 940.
  • One or more input devices 960 provide a portion of a user interface.
  • the one or more input devices 960 may include an alphanumeric keypad, such as a keyboard, for inputting alphanumeric and other information, or a pointing device, such as a mouse, a trackball, a stylus, or cursor direction keys.
  • the computing system 900 as shown in FIG. 9 includes one or more output devices 950. Suitable one or more output devices 950 include speakers, printers, network interfaces, and monitors.
  • Network interface 970 can be utilized to communicate with external devices, external computing devices, servers, and networked systems via one or more communications networks such as one or more wired, wireless, or optical networks including, for example, the Internet, intranet, LAN, WAN, cellular phone networks (e.g., Global System for Mobile communications network, packet switching
  • Network interface 970 may be a network interface card, such as an Ethernet card, optical transceiver, radio frequency transceiver, or any other type of device that can send and receive information.
  • network interfaces may include Bluetooth®, 3G, 4G, and WiFi® radios in mobile computing devices as well as a USB.
  • One or more peripheral devices 980 may include any type of computer support device to add additional functionality to the computing system.
  • the one or more peripheral devices 980 may include a modem or a router.
  • the example computing system 900 of FIG. 9 may also include one or more accelerator devices 985.
  • the accelerator devices 985 may include PCIe-form-f actor boards or storage-form-factor boards, or any electronic board equipped with a specific electronic component like a Graphical Processing Unit, a Neural Processing Unit, a Multi-CPU component, a Field Programmable Gate Array component, or similar accelerators electronic or photonic components, that may be suitable for use with embodiments described herein.
  • the components contained in the exemplary computing system 900 of FIG. 9 are those typically found in computing systems that may be suitable for use with embodiments described herein and are intended to represent a broad category of such computer components that are well known in the art.
  • the exemplary computing system 900 of FIG. 9 can be a personal computer, hand held computing device, telephone, mobile computing device, workstation, server, minicomputer, mainframe computer, or any other computing device.
  • the computer can also include different bus configurations, networked platforms, multi-processor platforms, and so forth.
  • Various operating systems (OS) can be used including UNIX, Linux, Windows, Macintosh OS, Palm OS, and other suitable operating systems.
  • Some of the above-described functions may be composed of instructions that are stored on storage media (e.g., computer-readable medium).
  • the instructions may be retrieved and executed by the processor.
  • Some examples of storage media are memory devices, tapes, disks, and the like.
  • the instructions are operational when executed by the processor to direct the processor to operate in accord with the example embodiments. Those skilled in the art are familiar with instructions, processor(s), and storage media.
  • Non-volatile media include, for example, optical or magnetic disks, such as a fixed disk.
  • Volatile media include dynamic memory, such as RAM.
  • Transmission media include coaxial cables, copper wire, and fiber optics, among others, including the wires that include one embodiment of a bus. Transmission media can also take the form of acoustic or light waves, such as those generated during radio frequency and infrared data
  • Common forms of computer-readable media include, for example, a floppy disk, a flexible disk, a hard disk, magnetic tape, any other magnetic medium, SSD, a CD-read-only memory (ROM) disk, DVD, any other optical medium, any other physical medium with patterns of marks or holes, a RAM, a PROM, an EPROM, an EEPROM, a FLASHEPROM, any other memory chip or cartridge, a carrier wave, or any other medium from which a computer can read.
  • Various forms of computer-readable media may be involved in carrying one or more sequences of one or more instructions to a CPU for execution.
  • a bus carries the data to system RAM, from which a CPU retrieves and executes the instructions.
  • the instructions received by system RAM can optionally be stored on a fixed disk either before or after execution by a CPU.
  • the instructions or data may not be used by the CPU but be accessed in writing or reading from the other devices without having the CPU directing them.

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)
  • General Health & Medical Sciences (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Computational Linguistics (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Artificial Intelligence (AREA)
  • Neurology (AREA)
  • Stored Programmes (AREA)

Abstract

La présente invention concerne des systèmes et des procédés pour optimiser des calculs de réseau neuronal artificiel (ANN) sur la base d'une détermination automatique d'une taille de lot. Un procédé donné à titre d'exemple peut comprendre les étapes consistant à recevoir, par un module d'optimisation, une structure d'ANN associée à l'ANN, et à générer, sur la base de la structure d'ANN, une configuration pour un moteur de calcul capable de réaliser le calcul des couches de l'ANN. La configuration peut comprendre des informations concernant une taille de lot d'une ou plusieurs couches de l'ANN. La taille de lot d'une couche peut être déterminée sur la base d'une largeur de bande requise pour lire des données relatives à la couche, d'un nombre de paramètres associés à la couche, et d'une durée de traitement, par la couche, d'un ensemble de données d'entrée à partir du lot. La taille de lot de la couche peut différer de la taille de lot de l'ANN. La taille de lot de la couche peut différer d'une taille de lot d'une autre couche de l'ANN.
EP19702705.5A 2019-01-10 2019-01-10 Optimisation de calculs de réseau neuronal artificiel sur la base d'une détermination automatique d'une taille de lot Pending EP3908981A1 (fr)

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/IB2019/050192 WO2020144493A1 (fr) 2019-01-10 2019-01-10 Optimisation de calculs de réseau neuronal artificiel sur la base d'une détermination automatique d'une taille de lot

Publications (1)

Publication Number Publication Date
EP3908981A1 true EP3908981A1 (fr) 2021-11-17

Family

ID=65269010

Family Applications (1)

Application Number Title Priority Date Filing Date
EP19702705.5A Pending EP3908981A1 (fr) 2019-01-10 2019-01-10 Optimisation de calculs de réseau neuronal artificiel sur la base d'une détermination automatique d'une taille de lot

Country Status (2)

Country Link
EP (1) EP3908981A1 (fr)
WO (1) WO2020144493A1 (fr)

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP3346425B1 (fr) * 2017-01-04 2023-12-20 STMicroelectronics S.r.l. Moteur et procédé accélérateur de matériel

Also Published As

Publication number Publication date
WO2020144493A1 (fr) 2020-07-16

Similar Documents

Publication Publication Date Title
US20200226458A1 (en) Optimizing artificial neural network computations based on automatic determination of a batch size
US11625583B2 (en) Quality monitoring and hidden quantization in artificial neural network computations
US10990525B2 (en) Caching data in artificial neural network computations
US20200311511A1 (en) Accelerating neuron computations in artificial neural networks by skipping bits
EP3924891A1 (fr) Surveillance de qualité et quantification cachée dans des calculs de réseau neuronal artificiel
US11494624B2 (en) Accelerating neuron computations in artificial neural networks with dual sparsity
US11068784B2 (en) Generic quantization of artificial neural networks
US11568255B2 (en) Fine tuning of trained artificial neural network
US10769527B2 (en) Accelerating artificial neural network computations by skipping input values
WO2020121030A1 (fr) Mise en mémoire cache de données dans des calculs de réseau neuronal artificiel
WO2020194032A1 (fr) Accélération de calculs neuronaux dans des réseaux neuronaux artificiels par saut de bits
WO2020144493A1 (fr) Optimisation de calculs de réseau neuronal artificiel sur la base d'une détermination automatique d'une taille de lot
US11126912B2 (en) Realigning streams of neuron outputs in artificial neural network computations
US20220222519A1 (en) Optimizing operations in artificial neural network
US11645510B2 (en) Accelerating neuron computations in artificial neural networks by selecting input data
US20210117800A1 (en) Multiple locally stored artificial neural network computations
US11748623B2 (en) Modifying structure of artificial neural networks by collocating parameters
EP4278304A1 (fr) Optimisation d'opérations dans un réseau de neurones artificiels
EP4049187A1 (fr) Multiples calculs stockés localement de réseau neuronal artificiel
EP3973464A1 (fr) Accélération de calculs neuronaux dans des réseaux neuronaux artificiels avec double dispersion
EP3895073A1 (fr) Réalignement de flux de sorties de neurone dans des calculs de réseau de neurones artificiels
WO2021209789A1 (fr) Structure modificatrice de réseaux neuronaux artificiels en colocalisant des paramètres
WO2022053851A1 (fr) Réglage de précision d'un réseau neuronal artificiel entraîné
EP3895071A1 (fr) Accélération de calculs de réseau neuronal artificiel par saut de valeurs d'entrée
WO2020208396A1 (fr) Accélération de calculs neuronaux dans des réseaux neuronaux artificiels par sélection des données d'entrée

Legal Events

Date Code Title Description
STAA Information on the status of an ep patent application or granted ep patent

Free format text: STATUS: UNKNOWN

STAA Information on the status of an ep patent application or granted ep patent

Free format text: STATUS: THE INTERNATIONAL PUBLICATION HAS BEEN MADE

PUAI Public reference made under article 153(3) epc to a published international application that has entered the european phase

Free format text: ORIGINAL CODE: 0009012

STAA Information on the status of an ep patent application or granted ep patent

Free format text: STATUS: REQUEST FOR EXAMINATION WAS MADE

17P Request for examination filed

Effective date: 20210810

AK Designated contracting states

Kind code of ref document: A1

Designated state(s): AL AT BE BG CH CY CZ DE DK EE ES FI FR GB GR HR HU IE IS IT LI LT LU LV MC MK MT NL NO PL PT RO RS SE SI SK SM TR

DAV Request for validation of the european patent (deleted)
DAX Request for extension of the european patent (deleted)
STAA Information on the status of an ep patent application or granted ep patent

Free format text: STATUS: EXAMINATION IS IN PROGRESS

17Q First examination report despatched

Effective date: 20230713