CN108229646A - neural network model compression method, device, storage medium and electronic equipment - Google Patents
neural network model compression method, device, storage medium and electronic equipment Download PDFInfo
- Publication number
- CN108229646A CN108229646A CN201710671900.2A CN201710671900A CN108229646A CN 108229646 A CN108229646 A CN 108229646A CN 201710671900 A CN201710671900 A CN 201710671900A CN 108229646 A CN108229646 A CN 108229646A
- Authority
- CN
- China
- Prior art keywords
- network model
- network
- nerves
- nervus opticus
- layer
- 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
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
- G06N3/04—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
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Image Analysis (AREA)
Abstract
The embodiment of the present invention provides a kind of neural network model compression method, device, storage medium and electronic equipment.This method includes:Obtain first nerves network model;It keeps or increases the depth of the first nerves network model and compress an at least network parameter for an at least network layer for the first nerves network model, obtain nervus opticus network model;The nervus opticus network model is trained based on sample data set and according at least to the output of the first nerves network model.So as to train and obtain strong ability in feature extraction, performance and the comparable compressed neural network model of uncompressed neural network model, and the training method has versatility, be adapted to carry out the neural network model of any function.
Description
Technical field
The present embodiments relate to artificial intelligence technology more particularly to a kind of neural network model compression method, device, meters
Calculation machine storage medium and electronic equipment.
Background technology
In recent years, deep neural network is applied in all multitasks such as computer vision, natural language processing so that performance
The raising of making a breakthrough property.In deep neural network model, powerful ability to express is exchanged for a large amount of network weight, so as to
Obtain stronger performance.For example, the size of AlexNet models has been more than 200MB, the size of VGG-16 models has been more than 500MB.
However, due to its depth and network parameter, the network feedforward time of deep neural network is usually longer, these upper limits to a certain degree
Application of the neural network in the limited equipment of computing resource is made.
Invention content
The embodiment of the present invention proposes a kind of neural network model compression scheme.
It is according to embodiments of the present invention in a first aspect, provide a kind of neural network model compression method, including:Obtain first
Neural network model;It keeps or increases the depth of the first nerves network model and compress the first nerves network model
An at least network parameter for an at least network layer, obtains nervus opticus network model;Based on sample data set and according at least to institute
The nervus opticus network model is trained in the output for stating first nerves network model.
Optionally, an at least network parameter includes:Convolution kernel size and/or, feature port number.
Optionally, increase the depth of the first nerves network model, including:Building the nervus opticus network development process
In, it is according to the receptive field of a larger convolutional layer of convolution kernel in the first nerves network model, the convolution kernel is larger
A convolutional layer using the smaller multiple convolutional layers of convolution kernel come equivalence replacement.
Optionally, the output training described the based on sample data and according at least to the first nerves network model
Two neural network models, including:The sample data is inputted into the first nerves network model, obtains the first nerves net
The output result of at least one network layer in network model;The sample data is inputted into the nervus opticus network model, is obtained
The output result of at least one network layer in the nervus opticus network model;Determine the nervus opticus network model and described
First difference of the output result of at least one respective wire network layers of first nerves network model;According to first discrepancy adjustment
The network parameter of the nervus opticus network model.
Optionally, the output training described the based on sample data and according at least to the first nerves network model
Two neural network models, including:Last network layer of the nervus opticus network model is connect with auxiliary convolutional layer;By institute
It states sample data and inputs the first nerves network model, obtain at least one network layer in the first nerves network model
Export result;The sample data is inputted into the nervus opticus network model, is obtained in the nervus opticus network model extremely
The output result of a few network layer;Determine the nervus opticus network model and the first nerves network model extremely respectively
First difference of the output result of few respective wire network layers and the final output result of the first nerves network model and
Second difference of the output result of the auxiliary convolutional layer;According to first difference and second difference weighting adjustment
The network parameter of nervus opticus network model, and the weight of first difference is more than the weight of second difference.
Optionally, after nervus opticus network model training is completed, the auxiliary convolutional layer is removed.
Optionally, at least one respective wire network layers include:The nervus opticus network model and the first nerves
The last one respective network layer of network model.
Optionally, at least one respective wire network layers include:The nervus opticus network model and the first nerves
The identical mid-level net network layers of at least one depth of network model.
Optionally, include the identical mid-level net network layers of at least one depth at least one respective wire network layers
When, determine that first difference includes:Increase by least one fitting branch in the nervus opticus network, pass through described at least one
A fitting branch determines the defeated of the corresponding mid-level net network layers of the nervus opticus network model and the first nerves network model
Go out the first difference of result.
Optionally, after nervus opticus network model training is completed, at least one fitting branch described in removal.
Second aspect according to embodiments of the present invention also provides a kind of neural network model compression set, which is characterized in that
Including:Acquisition module, for obtaining first nerves network model;Compression module obtains for keeping or increasing the acquisition module
The depth of first nerves network model that takes and compress the first nerves network model an at least network layer an at least net
Network parameter obtains nervus opticus network model;Training module, for being based on sample data set and according at least to the first nerves
The nervus opticus network model is trained in the output of network model.
Optionally, an at least network parameter includes:Convolution kernel size and/or, feature port number.
Optionally, the compression module is used in the nervus opticus network development process is built, according to the first nerves
A larger convolutional layer of the convolution kernel is used convolution by the receptive field of a larger convolutional layer of convolution kernel in network model
The smaller multiple convolutional layers of core carry out equivalence replacement.
Optionally, the training module includes:First output result acquiring unit, for the sample data to be inputted institute
First nerves network model is stated, obtains the output result of at least one network layer in the first nerves network model;Second is defeated
Go out result acquiring unit, for the sample data to be inputted the nervus opticus network model, obtain the nervus opticus net
The output result of at least one network layer in network model;First difference acquiring unit, for determining the nervus opticus network mould
First difference of the output result of at least one respective wire network layers of type and the first nerves network model;First parameter adjustment
Unit, for the network ginseng of nervus opticus network model described in the first discrepancy adjustment for being determined according to the difference acquiring unit
Number.
Optionally, the training module includes:Auxiliary layer connection unit, for by the nervus opticus network model most
Latter network layer is connect with auxiliary convolutional layer;Third exports result acquiring unit, for by sample data input described the
One neural network model obtains the output result of at least one network layer in the first nerves network model;4th output knot
Fruit acquiring unit for the sample data to be inputted the nervus opticus network model, obtains the nervus opticus network mould
The output result of at least one network layer in type;Second difference acquiring unit, for determining the nervus opticus network mould respectively
The first difference and described first of the output result of at least one respective wire network layers of type and the first nerves network model
Second difference of the final output result of neural network model and the output result of the auxiliary convolutional layer;Second parameter adjustment list
Member, for adjusting the network parameter of the nervus opticus network model according to first difference and second difference weighting,
And the weight of first difference is more than the weight of second difference.
Optionally, the training module further includes:Auxiliary layer removal unit, for being instructed in the nervus opticus network model
Practice after completing, remove the auxiliary convolutional layer.
Optionally, at least one respective wire network layers include:The nervus opticus network model and the first nerves
The last one respective network layer of network model.
Optionally, at least one respective wire network layers include:The nervus opticus network model and the first nerves
The identical mid-level net network layers of at least one depth of network model.
Optionally, the first difference acquiring unit is divided for increasing by least one fitting in the nervus opticus network
Branch determines the phase of the nervus opticus network model and the first nerves network model by least one fitting branch
Answer the first difference of the output result of mid-level net network layers.
Optionally, the training module further includes:Branch's removal unit is fitted, in the nervus opticus network model
After training is completed, at least one fitting branch described in removal.
The third aspect according to embodiments of the present invention, also provides a kind of electronic equipment, including:Processor, memory, communication
Element and communication bus, the processor, the memory and the communication device are completed mutual by the communication bus
Communication;For the memory for storing an at least executable instruction, it is following that the executable instruction performs the processor
Operation:Obtain first nerves network model;It keeps or increases the depth of the first nerves network model and compress described first
An at least network parameter for an at least network layer for neural network model, obtains nervus opticus network model;Based on sample data
Collect and train the nervus opticus network model according at least to the output of the first nerves network model.
Optionally, an at least network parameter includes:Convolution kernel size and/or, feature port number.
Optionally, the executable instruction further makes the processor perform following operate:Building second god
Through in network development process, according to the receptive field of a larger convolutional layer of convolution kernel in the first nerves network model, will described in
A larger convolutional layer of convolution kernel is using the smaller multiple convolutional layers of convolution kernel come equivalence replacement.
Optionally, the executable instruction further makes the processor perform following operate:The sample data is defeated
Enter the first nerves network model, obtain the output result of at least one network layer in the first nerves network model;It will
The sample data inputs the nervus opticus network model, obtains at least one network layer in the nervus opticus network model
Output result;Determine at least one respective wire network layers of the nervus opticus network model and the first nerves network model
Output result the first difference;According to the network parameter of nervus opticus network model described in first discrepancy adjustment.
Optionally, the executable instruction further makes the processor perform following operate:By the nervus opticus net
Last network layer of network model is connect with auxiliary convolutional layer;The sample data is inputted into the first nerves network model,
Obtain the output result of at least one network layer in the first nerves network model;The sample data is inputted described second
Neural network model obtains the output result of at least one network layer in the nervus opticus network model;It determines respectively described
The first of the output result of at least one respective wire network layers of nervus opticus network model and the first nerves network model is poor
The final output result of different and described first nerves network model is second poor with the output result of the auxiliary convolutional layer
It is different;The network parameter of the nervus opticus network model, and institute are adjusted according to first difference and second difference weighting
The weight for stating the first difference is more than the weight of second difference.
Optionally, the executable instruction further makes the processor perform following operate:In the nervus opticus net
After network model training is completed, the auxiliary convolutional layer is removed.
Optionally, at least one respective wire network layers include:The nervus opticus network model and the first nerves
The last one respective network layer of network model.
Optionally, at least one respective wire network layers include:The nervus opticus network model and the first nerves
The identical mid-level net network layers of at least one depth of network model.
Optionally, the executable instruction further makes the processor perform following operate:In the nervus opticus net
Increase by least one fitting branch in network, the nervus opticus network model and described is determined by least one fitting branch
First difference of the output result of the corresponding mid-level net network layers of first nerves network model.
Optionally, the executable instruction further makes the processor perform following operate:In the nervus opticus net
After network model training is completed, at least one fitting branch described in removal.
Fourth aspect according to embodiments of the present invention also provides a kind of computer readable storage medium, is stored thereon with meter
Calculation machine program instruction, wherein, described program instruction realizes any nerve net provided by the embodiments of the present application when being executed by processor
The step of network model compression method.
5th aspect according to embodiments of the present invention, also provides a kind of computer program, includes computer program and refer to
It enables, described program instruction realizes any neural network model compression method provided by the embodiments of the present application when being executed by processor
Step.
The neural network model compression scheme provided according to embodiments of the present invention, passes through the neural network that will be compressed
An at least network parameter for an at least network layer for model is compressed, and obtains the neural network model through overcompression.It is compressing
In the process, the depth of original neural network model is kept or increased, is compressed so as to reduce as far as possible to the ability to express of network
It influences.Hereafter, then based on sample data set and according at least to uncompressed neural network model output training through overcompression
Neural network model, it is comparable to obtain ability in feature extraction strong, performance and uncompressed neural network model so as to training
Compressed neural network model, and the training method has versatility, is adapted to carry out the neural network mould of any function
Type.
Description of the drawings
Fig. 1 is the flow chart for the neural network model compression method for showing according to embodiments of the present invention one;
Fig. 2 is the flow chart for the neural network model compression method for showing according to embodiments of the present invention two;
Fig. 3 is the flow chart for the neural network model compression method for showing according to embodiments of the present invention three;
Fig. 4 shows the logic diagram of according to embodiments of the present invention four neural network model compression set;
Fig. 5 shows the logic diagram of according to embodiments of the present invention five neural network model compression set;
Fig. 6 shows the logic diagram of according to embodiments of the present invention six neural network model compression set;
Fig. 7 is the structure diagram for the electronic equipment for showing according to embodiments of the present invention seven.
Specific embodiment
The exemplary embodiment of the embodiment of the present invention is described in detail below in conjunction with the accompanying drawings.
Embodiment one
Fig. 1 is the flow chart for the neural network model compression method for showing according to embodiments of the present invention one.
With reference to Fig. 1, in step S110, first nerves network model is obtained.
Here first nerves network model can be trained neural network model.That is, according to this hair
The neural network model compression method of bright embodiment one is suitable for compressing any general neural network model.
The training of first nerves network model is not defined in embodiments of the present invention, it can be any using traditional net
Network training method trains first nerves network model in advance.Can according to function that first nerves network model to be realized, characteristic and
Training requirement, using based on the training the in advance such as supervised learning method, unsupervised approaches, intensified learning method or semi-supervised method
One neural network model.
For example, can be used based on supervised learning method, training is used for the first nerves network model of classification, wherein, for instruction
Practice sample and mark out desired classification value, exercised supervision with the classification value marked out to first nerves network model.Lacking foot
Enough prioris, it is difficult in the case that artificial mark classification or the artificial classification mark cost of progress are excessively high, can be used unsupervised
Method is trained for example for the first nerves network model of image characteristics extraction.For other applicable machine learning methods,
It does not give and describes one by one herein.
In step S120, keep or increase the depth of first nerves network model and compress first nerves network model extremely
An at least network parameter for a few network layer, obtains nervus opticus network model.
Network layer mentioned here may include, but be not limited to, in the neural networks such as convolutional layer, pond layer, full articulamentum
Layer.
Here, it can be joined by the one or more networks for the one or more network layers for only reducing first nerves network model
Number, to obtain the nervus opticus network model through overcompression.Alternatively, can by increasing the depth of first nerves network model, and
One or more network parameters of one or more network layers of first nerves network model are reduced, to obtain the through overcompression
Two neural network models, from there through the depth for increasing first nerves network model, to reduce the reduction pair due to network parameter
The influence of the accuracy of the nervus opticus network model of compression.
As can be seen that in aforementioned any compression processing, the depth of first nerves network model is not reduced.This be because
For the number in network parameter is comparable, and network is deeper, and the convolutional layer being superimposed before each network layer is more, god
Expressivity through network is stronger, and the receptive field of each layer neuron is bigger, and the corresponding input range of high-rise neuron is bigger, extraction
The ability of global characteristics is stronger, therefore its performance is better.Therefore, by do not reduce first nerves network model depth press
Contracting first nerves network model can not influence the ability to express of first nerves network model and the ability of global characteristics as far as possible.
In step S130, nervus opticus is trained based on sample data set and according at least to the output of first nerves network model
Network model.
Optionally, by the sample data set for the desired value/actual value for containing mark or the sample of labeled data is not contained
Data set, to train nervus opticus network model.Wherein, at least using the output of first nerves network model to nervus opticus net
The training of network model exercises supervision.
For example, aforementioned sample data set can be inputed to first nerves network model and nervus opticus network model respectively,
Carry out forward calculation.Hereafter, according to the output data counting loss value obtained from first nerves network model, further according to the loss
Value carries out backwards calculation, so as to which training obtains the nervus opticus network model with estimated performance;It alternatively, can be according to refreshing from first
The output data obtained through network model calculates first-loss value, according to the output data that is obtained from nervus opticus network model and
Corresponding desired value/actual value that sample data concentration contains calculates the second penalty values, further according to the first-loss value and second
Penalty values carry out backwards calculation, so as to which training obtains the nervus opticus network model with estimated performance.
According to embodiments of the present invention one neural network model compression method passes through the neural network mould that will be compressed
An at least network parameter for an at least network layer for type is compressed, and obtains the neural network model through overcompression.Compressed
Cheng Zhong keeps or increases the depth of original neural network model, so as to reduce shadow of the compression to the ability to express of network as far as possible
It rings.Hereafter, then based on sample data set and according at least to the output of uncompressed neural network model train through overcompression
Neural network model obtains strong ability in feature extraction, performance and the comparable pressure of uncompressed neural network model so as to training
Neural network model after contracting, and the training method has versatility, is adapted to carry out the neural network model of any function.
Embodiment two
Fig. 2 is the flow chart for the neural network model compression method for showing according to embodiments of the present invention two.
With reference to Fig. 2, in step S210, first nerves network model is obtained.The processing of the step is with abovementioned steps S110's
Handle similar, it will not be described here.
A kind of optional embodiment according to the present invention, in step S220, according to convolution kernel in first nerves network model
The receptive field of a larger convolutional layer, by a larger convolutional layer of convolution kernel using the smaller multiple convolutional layers of convolution kernel come
Equivalence replacement obtains nervus opticus network model.
For the first nerves network model for generally including convolutional layer, for the convolutional layer with larger convolution kernel,
A convolutional layer of big convolution kernel can be replaced by using multiple convolutional layers of small convolution kernel so that in suitable or identical receptive field
Under the premise of the complexity of calculating is reduced by small convolution kernel, can also increase network depth to improve the expression of neural network
Ability.
For example, 2 convolutional layers that convolution kernel size is 3*3 can be used to replace convolution kernel size in first nerves network model
For 1 convolutional layer of 5*5, the convolution kernel in first nerves network model is replaced using 3 convolutional layers that convolution kernel size is 3*3
Size is 1 convolutional layer of 7*7, etc..So that the receptive field of related convolutional layer is identical before and after replacement or phase as far as possible
Together, the complexity of calculating is reduced by small convolution kernel, can also increase network depth to improve the expression energy of neural network
Power..
According to an alternate embodiment of the invention, in step S220, compression first nerves network model is at least
An at least network parameter for one network layer, obtains nervus opticus network model.
For example, at least a network parameter may include, but it is not limited to:Convolution kernel size and/or, feature port number, etc..
By reducing the size of convolution kernel, i.e., the big convolution kernel of nuclear subsitution is accumulated using rouleau, keep first nerves network model
Depth it is constant, can reduce the computation complexity of convolutional layer, improve network feed forward velocity.For example, in first nerves network model
Two convolutional layers, use multiple 3 × 3 convolution nuclear subsitution, 5 × 5 convolution kernels or the convolution kernel of bigger so that two convolutional layers
Computation complexity and memory space reduce (1-18/25).
Number by the feature channel for reducing convolutional layer can delete the redundancy feature of convolutional layer, reduce calculation scale, and
And improve network feed forward velocity.For example, to the convolutional layer to be compressed, it can be by carrying out 1 × 1 convolution by the convolution to the convolutional layer
The number of the feature channel of layer is reduced to 256 from 512.
It can be by reducing the size of convolution kernel and reducing one of number two ways of feature channel of convolutional layer come to volume
Lamination is compressed, both modes can also be used in combination to compress any convolutional layer of first nerves network model or multiple
Convolutional layer.For example, to the convolutional layer to be compressed, it can first pass through and 1 × 1 convolution of convolutional layer progress is led to the feature of convolutional layer
The number in road reduces half;Hereafter, the convolutional layer then in this feature channel being reduced uses small convolution nuclear subsitution convolutional layer
Big convolution kernel, further to compress the convolutional layer.
Nervus opticus network model is being obtained by aforementioned compression processing, randomly nervus opticus network model can carried out
Initialization.
On the other hand, since the number of the network parameter of nervus opticus network model obtained through overcompression is less, study
Ability differs larger with original first nerves network model, is directly fitted the output data actual value of first nerves network model
It is more difficult.It is trained directly on training dataset to be susceptible to not convergent situation in the case where data volume is larger.
Further, since the predicted value of first nerves network model output is more smooth compared with the actual value of mark, learning difficulty
It is relatively low, the net of preferable nervus opticus network model can be obtained by being fitted the output of original first nerves network model
The initial value of network parameter.
For this purpose, optionally, the network parameter of nervus opticus network model is initialized as phase in first nerves network model
The network parameter answered, so as to the output of initial fitting first nerves network model.
Hereafter, the nervus opticus network model through overcompression is trained by the processing of step S230~S260.
During the neural network model more than the training number of plies, when performing reverse transfer, since network is deeper, it will go out
Existing more serious gradient disappears, and causes that the network parameter of the convolutional layer close to network inputs can not be trained, it is difficult to which training is acquired
It can preferably neural network model.For this purpose, obtaining the output of the mid-level net network layers of nervus opticus network model, make the second of compression
Neural network model is fitted the output feature of the respective layer of first nerves network model in each mid-level net network layers, and deep learning is leaned on
The network parameter of the layer (commonly referred to as low network layer) of the input terminal of nearly neural network model.
In step S230, sample data is inputted into first nerves network model, is obtained in first nerves network model at least
The output result of one network layer.
Here, forward calculation is carried out to first nerves network model by using sample data, from first nerves network mould
At least one of type network layer obtains corresponding output result.Here at least one network layer may include first nerves network
The mid-level net network layers and the last one network layer of model.
In step S240, sample data is inputted into nervus opticus network model, is obtained in nervus opticus network model at least
The output result of one network layer.
Similarly, forward calculation is carried out to nervus opticus network model by using sample data, from nervus opticus network mould
At least one of type network layer obtains corresponding output result.Here at least one network layer may include nervus opticus network
The mid-level net network layers and the last one network layer of model.
It may be noted that step S230 and S240 can be performed in any order or is performed in parallel step S230 and S240.
In step S250, at least one respective wire network layers of nervus opticus network model and first nerves network model are determined
Output result the first difference.
A kind of optional embodiment according to the present invention, at least one respective wire network layers here include:Nervus opticus net
Network model and the last one respective network layer of first nerves network model.
That is, according to the final output of first nerves network model and the final output of nervus opticus network model come
Calculate the first difference.
According to an alternate embodiment of the invention, at least one respective wire network layers here include:Nervus opticus
The network model mid-level net network layers identical at least one depth of first nerves network model.
That is, according at least one identical with depth in first nerves network model of nervus opticus network model
Between the output result of network layer determine the first difference.
For this purpose, specifically, a kind of optional embodiment according to the present invention increases by least one plan in nervus opticus network
Branch is closed, the corresponding mid-level net of nervus opticus network model and first nerves network model is determined by least one fitting branch
First difference of the output result of network layers.
For example, it is assumed that first nerves network model and nervus opticus network model have 16 layers, it can be respectively in the two god
The 6th layer of setting fitting branch 1, and the 12nd layer in the two neural network models sets plan respectively respectively through network model
Branch 2 is closed, to carry out output feature fitting in intermediate convolutional layer.Sample data is inputted to first nerves network mould respectively aforementioned
After type and nervus opticus network model, can nervus opticus network model and first nerves network mould be determined by these fitting branches
First difference of the output result of the corresponding mid-level net network layers of type.
The first difference can be calculated for example, by loss function or apart from function is calculated, so as to assess nervus opticus network model
Relative to the accuracy in detection of first nerves network model.
It is understood that above-mentioned feasible program is only one of which realization method, in practical applications, user can be with
Realization condition or design parameter are adjusted according to actual demand, the citing of above-mentioned feasible program should not be construed as unique realization side
Formula.
In step S260, according to the network parameter of the first discrepancy adjustment nervus opticus network model.
That is, in the step, make the second fitting branch that the intermediate convolutional layer of nervus opticus network model sets
The output of first fitting branch of output fitting first nerves network model intermediate convolutional layer setting accordingly, to update compression
The network parameter of nervus opticus network model.Specifically, it is multiple training samples and the first difference are refreshing to second for reverse transfer
Through network model, the network parameter of nervus opticus network model is updated for example, by gradient descent method, it is first poor to make
It is different to converge to the range allowed, reach expected nervus opticus network model so as to which training obtains performance.
Optionally, after the training of nervus opticus network model is completed, these at least one plans for network training are removed
Close branch.
According to embodiments of the present invention two neural network model compression method, by the depth for increasing neural network model
And/or the network parameter of at least one network layer is reduced, obtain the small-scale neural network model through overcompression.Compressed
Cheng Zhong keeps or increases the depth of original neural network model, so as to reduce shadow of the compression to the ability to express of network as far as possible
It rings.It is in addition, original in the fitting of the intermediate convolutional layer of original neural network model and the small-scale neural network model being trained to
Neural network model output, so as to depth learn close to neural network model input terminal network layer (commonly referred to as
Low network layer) network parameter, improve the accuracy of the neural network model through overcompression.
Embodiment three
Fig. 3 is the flow chart for the neural network model compression method for showing according to embodiments of the present invention three.
With reference to Fig. 3, in step S310, first nerves network model is obtained.The processing of the step is with abovementioned steps S110's
Handle similar, it will not be described here.
In step S320, keep or increase the depth of first nerves network model and compress first nerves network model extremely
An at least network parameter for a few network layer, obtains nervus opticus network model.
The processing of the step is similar with the processing of abovementioned steps S120 or step S220, and it will not be described here.
Since nervus opticus network model is the network that first nerves network model generates after overcompression, network parameter
It is reduced, in order to improve the feature representation ability of nervus opticus network model, according to embodiments of the present invention three neural network
It is (usual in the output terminal of nervus opticus network model during training nervus opticus network model in model compression method
Referred to as network top) auxiliary convolutional layer is added, increase the network depth of nervus opticus network model, for being fitted depth network
Output.
Correspondingly, in step S330, last network layer of nervus opticus network model is connect with auxiliary convolutional layer.It is auxiliary
The number for helping convolutional layer can be one, two or more.
In step S340, aforementioned sample data is inputted into first nerves network model, is obtained in first nerves network model
The output of at least one network layer is as a result, the output result includes the output result of aforementioned mid-level net network layers and first nerves network
The final output result of model.
In step S350, aforementioned sample data is inputted into nervus opticus network model, is obtained in nervus opticus network model
The output result of at least one network layer.
That is, by the processing of step S350, from least one mid-level net network layers of nervus opticus network model with
And auxiliary convolutional layer obtains corresponding output result respectively.
It may be noted that step S340 and S350 can be performed in any order or is performed in parallel step S340 and S350.
In step S360, at least one respective wire of nervus opticus network model and first nerves network model is determined respectively
First difference of the output result of network layers and the final output result of first nerves network model and the output of auxiliary convolutional layer
As a result the second difference.
The processing that can refer to step S250 determines the first difference, and can be for example, by loss function or apart from calculating function
The second difference is calculated, so as to assess detection of the nervus opticus network model in auxiliary convolutional layer relative to first nerves network model
Accuracy.
In step S370, the network parameter of nervus opticus network model is adjusted according to the first difference and the weighting of the second difference,
And first difference weight be more than the second difference weight.
Specifically, the first difference and the second difference are weighted respectively, obtain comprehensive differences wherein, assigned for the first difference
Give the weight more than the second difference.Since the purpose of setting auxiliary convolutional layer is to further improve the spy of nervus opticus network model
Ability to express is levied, the training of nervus opticus network model is further optimized, therefore will for the weight that the second difference assigns
Less than the weight of the first difference.Hereafter, the comprehensive differences reverse transfer multiple training samples and weighting obtained is to nervus opticus
Network model is updated the network parameter of nervus opticus network model for example, by gradient descent method, makes comprehensive differences
The range allowed is converged to, reaches expected nervus opticus network model so as to which training obtains performance.
Optionally, after the training of nervus opticus network model is completed, removal is used for the auxiliary convolutional layer of network training.
According to embodiments of the present invention three neural network model compression method on the basis of previous embodiment, is being instructed
The output terminal of neural network model experienced, through overcompression adds auxiliary convolutional layer, and passes through the output from auxiliary convolutional layer
As a result the output of uncompressed neural network model is fitted, further enhances the feature extraction for the neural network model being trained to
Ability, and optimize the performance of the neural network model.Similarly, which has versatility, is adapted to carry out any work(
The neural network model with multiple convolutional layers of energy.
Example IV
Fig. 4 shows the logic diagram of according to embodiments of the present invention four neural network model compression set.
With reference to Fig. 4, according to embodiments of the present invention four neural network model compression set includes acquisition module 410, compression
Module 420 and training module 430.
Acquisition module 410 is used to obtain first nerves network model.
Compression module 420 is used to keep or increase the depth and pressure of the first nerves network model of the acquisition of acquisition module 410
Contract the first nerves network model an at least network layer an at least network parameter, obtain nervus opticus network model.
Training module 430 is for based on sample data set and according at least to the output training of the first nerves network model
The nervus opticus network model.
The neural network model compression set of the present embodiment is used to implement corresponding neural network in preceding method embodiment
Model compression method, and the advantageous effect with corresponding embodiment of the method, details are not described herein.
Embodiment five
Fig. 5 shows the logic diagram of according to embodiments of the present invention five neural network model compression set.
According to embodiments of the present invention five, include for at least network parameter that compression module 420 compresses:Convolution kernel is big
It is small and/or, feature port number.
Optionally, compression module 420 is used in the nervus opticus network development process is built, according to the first nerves net
A larger convolutional layer of the convolution kernel is used convolution kernel by the receptive field of a larger convolutional layer of convolution kernel in network model
Smaller multiple convolutional layers carry out equivalence replacement.
Training module 430 includes:
First output result acquiring unit 431, for the sample data to be inputted the first nerves network model, is obtained
Obtain the output result of at least one network layer in the first nerves network model;
Second output result acquiring unit 432, for the sample data to be inputted the nervus opticus network model, is obtained
Obtain the output result of at least one network layer in the nervus opticus network model;
First difference acquiring unit 433, for determining the nervus opticus network model and the first nerves network mould
First difference of the output result of at least one respective wire network layers of type;
First parameter adjustment unit 434, for described in the first discrepancy adjustment for being determined according to the difference acquiring unit
The network parameter of two neural network models.
A kind of optional embodiment according to the present invention, aforementioned at least one respective wire network layers include:The nervus opticus
Network model and the last one respective network layer of the first nerves network model.
According to an alternate embodiment of the invention, aforementioned at least one respective wire network layers include:Second god
Through the network model mid-level net network layers identical at least one depth of the first nerves network model.
Optionally, the first difference acquiring unit 433 is fitted branch for increasing at least one in the nervus opticus network,
The corresponding of the nervus opticus network model and the first nerves network model is determined by least one fitting branch
First difference of the output result of mid-level net network layers.
Optionally, training module 430 further includes:Branch's removal unit (not shown) is fitted, in the nervus opticus
After network model training is completed, at least one fitting branch described in removal.
The neural network model compression set of the present embodiment is used to implement corresponding neural network in preceding method embodiment
Model compression method, and the advantageous effect with corresponding embodiment of the method, details are not described herein.
Embodiment six
Fig. 5 shows the logic diagram of according to embodiments of the present invention six neural network model compression set.
According to embodiments of the present invention six, include for at least network parameter that compression module 420 compresses:Convolution kernel is big
It is small and/or, feature port number.
Optionally, compression module 420 is used in the nervus opticus network development process is built, according to the first nerves net
A larger convolutional layer of the convolution kernel is used convolution kernel by the receptive field of a larger convolutional layer of convolution kernel in network model
Smaller multiple convolutional layers carry out equivalence replacement.
According to embodiments of the present invention six, training module 430 includes:
Auxiliary layer connection unit 435, for by last network layer of the nervus opticus network model with auxiliary convolution
Layer connection;
Third output result acquiring unit 436, for the sample data to be inputted the first nerves network model, is obtained
Obtain the output result of at least one network layer in the first nerves network model;
4th output result acquiring unit 437, for the sample data to be inputted the nervus opticus network model, is obtained
Obtain the output result of at least one network layer in the nervus opticus network model;
Second difference acquiring unit 438, for determining the nervus opticus network model and the first nerves net respectively
First difference of the output result of at least one respective wire network layers of network model and the first nerves network model it is final
Export the second difference of result and the output result of the auxiliary convolutional layer;
Second parameter adjustment unit 439, for according to first difference and second difference weighting adjustment described the
The network parameter of two neural network models, and the weight of first difference is more than the weight of second difference.
A kind of optional embodiment according to the present invention, aforementioned at least one respective wire network layers include:The nervus opticus
Network model and the last one respective network layer of the first nerves network model.
According to an alternate embodiment of the invention, aforementioned at least one respective wire network layers include:Second god
Through the network model mid-level net network layers identical at least one depth of the first nerves network model.
Optionally, training module 430 further includes:Auxiliary layer removal unit (not shown), in the nervus opticus net
After network model training is completed, the auxiliary convolutional layer is removed.
The neural network model compression set of the present embodiment is used to implement corresponding neural network in preceding method embodiment
Model compression method, and the advantageous effect with corresponding embodiment of the method, details are not described herein.
Embodiment seven
Fig. 7 is the structure diagram for the electronic equipment for showing according to embodiments of the present invention seven.
The embodiment of the present invention additionally provides a kind of electronic equipment, such as can be mobile terminal, personal computer (PC), put down
Plate computer, server etc..Below with reference to Fig. 7, it illustrates suitable for being used for realizing the terminal device of the embodiment of the present invention or service
The structure diagram of the electronic equipment 700 of device.
As shown in fig. 7, electronic equipment 700 includes one or more processors, communication device etc., one or more of places
Manage device for example:One or more central processing unit (CPU) 701 and/or one or more image processors (GPU) 713 etc.,
Processor can according to the executable instruction being stored in read-only memory (ROM) 702 or from storage section 708 be loaded into
Machine accesses the executable instruction in memory (RAM) 703 and performs various appropriate actions and processing.Communication device includes communication
Component 712 and communication interface 709.Wherein, communication component 712 may include but be not limited to network interface card, and the network interface card may include but unlimited
In IB (Infiniband) network interface card, communication interface 1409 includes the communication of the network interface card of LAN card, modem etc.
Interface, communication interface 1409 perform communication process via the network of such as internet.
Processor can communicate with read-only memory 702 and/or random access storage device 730 to perform executable instruction,
It is connected by bus 704 with communication component 712 and is communicated through communication component 712 with other target devices, thereby completing the present invention
The corresponding operation of any one method that embodiment provides, for example, obtaining first nerves network model;It keeps or increases described first
The depth of neural network model and compress the first nerves network model an at least network layer an at least network parameter, obtain
To nervus opticus network model;Based on sample data set and according at least to described in the output training of the first nerves network model
Nervus opticus network model.
In a kind of optional embodiment, an at least network parameter includes:Convolution kernel size and/or, feature is led to
Road number.
In a kind of optional embodiment, executable instruction is for so that processor further performs following operate:
It builds in the nervus opticus network development process, according to a larger convolutional layer of convolution kernel in the first nerves network model
Receptive field, by a larger convolutional layer of the convolution kernel using the smaller multiple convolutional layers of convolution kernel come equivalence replacement.
In a kind of optional embodiment, the executable instruction further makes the processor perform following operate:
The sample data is inputted into the first nerves network model, obtains at least one network in the first nerves network model
The output result of layer;The sample data is inputted into the nervus opticus network model, obtains the nervus opticus network model
In at least one network layer output result;Determine the nervus opticus network model and the first nerves network model extremely
First difference of the output result of few respective wire network layers;According to nervus opticus network model described in first discrepancy adjustment
Network parameter.
In a kind of optional embodiment, the executable instruction further makes the processor perform following operate:
Last network layer of the nervus opticus network model is connect with auxiliary convolutional layer;By sample data input described the
One neural network model obtains the output result of at least one network layer in the first nerves network model;By the sample
Data input the nervus opticus network model, obtain the output knot of at least one network layer in the nervus opticus network model
Fruit;The defeated of at least one respective wire network layers of the nervus opticus network model and the first nerves network model is determined respectively
Go out the first difference of result and the final output result of the first nerves network model and the output of the auxiliary convolutional layer
As a result the second difference;The net of the nervus opticus network model is adjusted according to first difference and second difference weighting
Network parameter, and the weight of first difference is more than the weight of second difference.
In a kind of optional embodiment, the executable instruction further makes the processor perform following operate:
After nervus opticus network model training is completed, the auxiliary convolutional layer is removed.
In a kind of optional embodiment, at least one respective wire network layers include:The nervus opticus network mould
Type and the last one respective network layer of the first nerves network model.
In another optional embodiment, at least one respective wire network layers include:The nervus opticus network
The model mid-level net network layers identical at least one depth of the first nerves network model.
In a kind of optional embodiment, the executable instruction further makes the processor perform following operate:
Increase by least one fitting branch in the nervus opticus network, second god is determined by least one fitting branch
First difference of the output result of the corresponding mid-level net network layers through network model and the first nerves network model.
In a kind of optional embodiment, the executable instruction further makes the processor perform following operate:
After nervus opticus network model training is completed, at least one fitting branch described in removal.
In addition, in RAM 703, it can also be stored with various programs and data needed for device operation.CPU701,
ROM702 and RAM703 is connected with each other by bus 704.In the case where there is RAM703, ROM702 is optional module.
RAM703 stores executable instruction or executable instruction is written into ROM702 at runtime, and executable instruction makes processor 701
Perform the corresponding operation of above-mentioned communication means.Input/output (I/O) interface 705 is also connected to bus 704.Communication component 712 can
With integrally disposed, may be set to be with multiple submodule (such as multiple IB network interface cards), and in bus link.
I/O interfaces 705 are connected to lower component:Importation 706 including keyboard, mouse etc.;It is penetrated including such as cathode
The output par, c 707 of spool (CRT), liquid crystal display (LCD) etc. and loud speaker etc.;Storage section 708 including hard disk etc.;
And the communication interface 709 of the network interface card including LAN card, modem etc..Driver 710 is also according to needing to connect
It is connected to I/O interfaces 705.Detachable media 711, such as disk, CD, magneto-optic disk, semiconductor memory etc. are pacified as needed
On driver 710, in order to be mounted into storage section 708 as needed from the computer program read thereon.
It should be noted that framework as shown in Figure 7 is only a kind of optional realization method, it, can during concrete practice
The component count amount and type of above-mentioned Fig. 7 are selected, are deleted, increased or replaced according to actual needs;In different function component
In setting, can also be used it is separately positioned or integrally disposed and other implementations, such as GPU and CPU separate setting or can be by GPU
It is integrated on CPU, the separable setting of communication component 712, can also be integrally disposed on CPU or GPU, etc..These are alternatively
Embodiment each falls within protection scope of the present invention.
Particularly, according to embodiments of the present invention, it is soft to may be implemented as computer for the process above with reference to flow chart description
Part program.For example, the embodiment of the present invention includes a kind of computer program product, including being tangibly embodied in machine readable media
On computer program, computer program included for the program code of the method shown in execution flow chart, and program code can wrap
The corresponding instruction of corresponding execution method and step provided in an embodiment of the present invention is included, for example, for obtaining first nerves network model
Executable code;For keeping or increasing the depth of the first nerves network model and compress the first nerves network mould
An at least network parameter for an at least network layer for type, obtains the executable code of nervus opticus network model;For being based on sample
Notebook data collection simultaneously exports the executable of the training nervus opticus network model according at least to the first nerves network model
Code.In such embodiments, the computer program can be downloaded and installed from network by communication device and/or
It is mounted from detachable media 711.When the computer program is performed by central processing unit (CPU) 701, it is real to perform the present invention
Apply the above-mentioned function of being limited in the method for example.
The electronic equipment that the embodiment of the present invention seven provides, by an at least net for the neural network model that will be compressed
An at least network parameter for network layers is compressed, and obtains the neural network model through overcompression.In compression process, keep or increase
The depth of big original neural network model, so as to reduce influence of the compression to the ability to express of network as far as possible.Hereafter, it then is based on
Sample data set simultaneously exports neural network model of the training through overcompression according at least to uncompressed neural network model, from
And it can train and obtain strong ability in feature extraction, performance and the comparable compressed neural network mould of uncompressed neural network model
Type, and the training method has versatility, is adapted to carry out the neural network model of any function.
Embodiment eight
According to embodiments of the present invention eight provide a kind of computer readable storage medium, are stored thereon with computer program and refer to
It enables, wherein, described program instructs the step of realizing aforementioned any neural network model compression method when being executed by processor.
The computer readable storage medium is used to implement in preceding method embodiment corresponding neural network model compression side
Method, and the advantageous effect with corresponding embodiment of the method, details are not described herein.
The computer readable storage medium is used to implement in preceding method embodiment corresponding neural network model compression side
Method, and the advantageous effect with corresponding embodiment of the method, details are not described herein.
It may be noted that according to the needs of implementation, all parts/step described in this application can be split as more multi-section
The part operation of two or more components/steps or components/steps can be also combined into new components/steps by part/step,
To realize the purpose of the embodiment of the present invention.
Methods and apparatus of the present invention, equipment may be achieved in many ways.For example, software, hardware, firmware can be passed through
Or any combinations of software, hardware, firmware realize the method and apparatus of the embodiment of the present invention, equipment.For the step of method
Merely to illustrate, the step of method of the embodiment of the present invention, is not limited to described in detail above suitable for rapid said sequence
Sequence, unless specifically stated otherwise.In addition, in some embodiments, the present invention can be also embodied as being recorded in record Jie
Program in matter, these programs include being used to implement machine readable instructions according to the method for the embodiment of the present invention.Thus, this hair
The recording medium of program of the bright also covering storage for execution according to the method for the present invention.
The description of the embodiment of the present invention in order to example and description for the sake of and provide, and be not exhaustively or will
The present invention is limited to disclosed form.Many modifications and variations are obvious for the ordinary skill in the art.Choosing
It is to more preferably illustrate the principle of the present invention and practical application to select and describe embodiment, and makes those of ordinary skill in the art
It will be appreciated that the present invention is so as to design the various embodiments with various modifications suitable for special-purpose.
Claims (10)
1. a kind of neural network model compression method, which is characterized in that including:
Obtain first nerves network model;
It keeps or increases the depth of the first nerves network model and compress an at least net for the first nerves network model
An at least network parameter for network layers, obtains nervus opticus network model;
The nervus opticus network mould is trained based on sample data set and according at least to the output of the first nerves network model
Type.
2. according to the method described in claim 1, it is characterized in that, an at least network parameter includes:Convolution kernel size,
And/or feature port number.
3. according to the method described in claim 1, it is characterized in that, increase the depth of the first nerves network model, including:
In the nervus opticus network development process is built, according to the volume that convolution kernel is larger in the first nerves network model
The receptive field of lamination replaces a larger convolutional layer of the convolution kernel using the smaller multiple convolutional layers of convolution kernel come equivalent
It changes.
4. according to any methods of claim 1-3, which is characterized in that described based on sample data and according at least to described
The nervus opticus network model is trained in the output of first nerves network model, including:
The sample data is inputted into the first nerves network model, is obtained at least one in the first nerves network model
The output result of network layer;
The sample data is inputted into the nervus opticus network model, is obtained at least one in the nervus opticus network model
The output result of network layer;
Determine the output of at least one respective wire network layers of the nervus opticus network model and the first nerves network model
As a result the first difference;
According to the network parameter of nervus opticus network model described in first discrepancy adjustment.
5. according to any methods of claim 1-3, which is characterized in that described based on sample data and according at least to described
The nervus opticus network model is trained in the output of first nerves network model, including:
Last network layer of the nervus opticus network model is connect with auxiliary convolutional layer;
The sample data is inputted into the first nerves network model, is obtained at least one in the first nerves network model
The output result of network layer;
The sample data is inputted into the nervus opticus network model, is obtained at least one in the nervus opticus network model
The output result of network layer;
At least one respective wire network layers of the nervus opticus network model and the first nerves network model are determined respectively
It exports the first difference of result and the final output result of the first nerves network model and assists the defeated of convolutional layer with described
Go out the second difference of result;
The network parameter of the nervus opticus network model, and institute are adjusted according to first difference and second difference weighting
The weight for stating the first difference is more than the weight of second difference.
6. a kind of neural network model compression set, which is characterized in that including:
Acquisition module, for obtaining first nerves network model;
Compression module, for keeping or increasing described in the depth for the first nerves network model that the acquisition module obtains and compression
An at least network parameter for an at least network layer for first nerves network model, obtains nervus opticus network model;
Training module, for based on sample data set and according at least to the output training described the of the first nerves network model
Two neural network models.
7. device according to claim 6, which is characterized in that an at least network parameter includes:Convolution kernel size,
And/or feature port number.
8. device according to claim 6, which is characterized in that the compression module is used to build the nervus opticus net
During network, according to the receptive field of a larger convolutional layer of convolution kernel in the first nerves network model, by the convolution
A larger convolutional layer of core is using the smaller multiple convolutional layers of convolution kernel come equivalence replacement.
9. a kind of electronic equipment, including:Processor, memory, communication device and communication bus, the processor, the storage
Device and the communication device complete mutual communication by the communication bus;
For the memory for storing an at least executable instruction, the executable instruction makes the processor perform following grasp
Make:
Obtain first nerves network model;
It keeps or increases the depth of the first nerves network model and compress an at least net for the first nerves network model
An at least network parameter for network layers, obtains nervus opticus network model;
The nervus opticus network mould is trained based on sample data set and according at least to the output of the first nerves network model
Type.
10. a kind of computer readable storage medium, is stored thereon with computer program instructions, wherein, described program instruction is located
Manage the step of any one of the Claims 1 to 5 neural network model compression method is realized when device performs.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710671900.2A CN108229646A (en) | 2017-08-08 | 2017-08-08 | neural network model compression method, device, storage medium and electronic equipment |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710671900.2A CN108229646A (en) | 2017-08-08 | 2017-08-08 | neural network model compression method, device, storage medium and electronic equipment |
Publications (1)
Publication Number | Publication Date |
---|---|
CN108229646A true CN108229646A (en) | 2018-06-29 |
Family
ID=62654233
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710671900.2A Pending CN108229646A (en) | 2017-08-08 | 2017-08-08 | neural network model compression method, device, storage medium and electronic equipment |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108229646A (en) |
Cited By (21)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109165720A (en) * | 2018-09-05 | 2019-01-08 | 深圳灵图慧视科技有限公司 | Neural network model compression method, device and computer equipment |
CN109165515A (en) * | 2018-08-10 | 2019-01-08 | 深圳前海微众银行股份有限公司 | Model parameter acquisition methods, system and readable storage medium storing program for executing based on federation's study |
CN109165725A (en) * | 2018-08-10 | 2019-01-08 | 深圳前海微众银行股份有限公司 | Neural network federation modeling method, equipment and storage medium based on transfer learning |
CN109214343A (en) * | 2018-09-14 | 2019-01-15 | 北京字节跳动网络技术有限公司 | Method and apparatus for generating face critical point detection model |
CN109255444A (en) * | 2018-08-10 | 2019-01-22 | 深圳前海微众银行股份有限公司 | Federal modeling method, equipment and readable storage medium storing program for executing based on transfer learning |
CN109325584A (en) * | 2018-08-10 | 2019-02-12 | 深圳前海微众银行股份有限公司 | Federation's modeling method, equipment and readable storage medium storing program for executing neural network based |
CN109634315A (en) * | 2018-12-29 | 2019-04-16 | 福建龙净环保股份有限公司 | A kind of method and device controlling plasm PH value |
CN109934300A (en) * | 2019-03-21 | 2019-06-25 | 腾讯科技(深圳)有限公司 | Model compression method, apparatus, computer equipment and storage medium |
CN110766131A (en) * | 2019-05-14 | 2020-02-07 | 北京嘀嘀无限科技发展有限公司 | Data processing device and method and electronic equipment |
CN111062396A (en) * | 2019-11-29 | 2020-04-24 | 深圳云天励飞技术有限公司 | License plate number recognition method and device, electronic equipment and storage medium |
CN111105029A (en) * | 2018-10-29 | 2020-05-05 | 北京地平线机器人技术研发有限公司 | Neural network generation method and device and electronic equipment |
CN111626409A (en) * | 2020-07-30 | 2020-09-04 | 江西高创保安服务技术有限公司 | Data generation method for image quality detection |
CN112016681A (en) * | 2019-05-31 | 2020-12-01 | 苹果公司 | Decomposition of machine learning operations |
CN112102281A (en) * | 2020-09-11 | 2020-12-18 | 哈尔滨市科佳通用机电股份有限公司 | Truck brake cylinder fault detection method based on improved Faster Rcnn |
CN112288829A (en) * | 2020-11-03 | 2021-01-29 | 中山大学 | Compression method and device for image restoration convolutional neural network |
CN112734007A (en) * | 2020-12-31 | 2021-04-30 | 青岛海尔科技有限公司 | Method and device for acquiring compression model, storage medium and electronic device |
CN112749782A (en) * | 2019-10-31 | 2021-05-04 | 上海商汤智能科技有限公司 | Data processing method and related product |
CN113326913A (en) * | 2020-02-28 | 2021-08-31 | 上海商汤智能科技有限公司 | Neural network model conversion method, model precision positioning method and device |
CN113657592A (en) * | 2021-07-29 | 2021-11-16 | 中国科学院软件研究所 | Software-defined satellite self-adaptive pruning model compression method |
CN113762510A (en) * | 2021-09-09 | 2021-12-07 | 北京百度网讯科技有限公司 | Data processing method and device for target model, electronic equipment and medium |
US11836635B2 (en) | 2019-05-31 | 2023-12-05 | Apple Inc. | Mutable parameters for machine learning models during runtime |
-
2017
- 2017-08-08 CN CN201710671900.2A patent/CN108229646A/en active Pending
Cited By (32)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2020029589A1 (en) * | 2018-08-10 | 2020-02-13 | 深圳前海微众银行股份有限公司 | Model parameter acquisition method and system based on federated learning, and readable storage medium |
CN109165515A (en) * | 2018-08-10 | 2019-01-08 | 深圳前海微众银行股份有限公司 | Model parameter acquisition methods, system and readable storage medium storing program for executing based on federation's study |
CN109165725A (en) * | 2018-08-10 | 2019-01-08 | 深圳前海微众银行股份有限公司 | Neural network federation modeling method, equipment and storage medium based on transfer learning |
CN109255444B (en) * | 2018-08-10 | 2022-03-29 | 深圳前海微众银行股份有限公司 | Federal modeling method and device based on transfer learning and readable storage medium |
CN109255444A (en) * | 2018-08-10 | 2019-01-22 | 深圳前海微众银行股份有限公司 | Federal modeling method, equipment and readable storage medium storing program for executing based on transfer learning |
CN109325584A (en) * | 2018-08-10 | 2019-02-12 | 深圳前海微众银行股份有限公司 | Federation's modeling method, equipment and readable storage medium storing program for executing neural network based |
EP3848838A4 (en) * | 2018-08-10 | 2022-06-08 | Webank Co.,Ltd | Model parameter acquisition method and system based on federated learning, and readable storage medium |
CN109165720A (en) * | 2018-09-05 | 2019-01-08 | 深圳灵图慧视科技有限公司 | Neural network model compression method, device and computer equipment |
CN109214343B (en) * | 2018-09-14 | 2021-03-09 | 北京字节跳动网络技术有限公司 | Method and device for generating face key point detection model |
CN109214343A (en) * | 2018-09-14 | 2019-01-15 | 北京字节跳动网络技术有限公司 | Method and apparatus for generating face critical point detection model |
CN111105029B (en) * | 2018-10-29 | 2024-04-16 | 北京地平线机器人技术研发有限公司 | Neural network generation method, generation device and electronic equipment |
CN111105029A (en) * | 2018-10-29 | 2020-05-05 | 北京地平线机器人技术研发有限公司 | Neural network generation method and device and electronic equipment |
CN109634315A (en) * | 2018-12-29 | 2019-04-16 | 福建龙净环保股份有限公司 | A kind of method and device controlling plasm PH value |
CN109634315B (en) * | 2018-12-29 | 2021-12-03 | 福建龙净环保股份有限公司 | Method and device for controlling pH value of slurry |
CN109934300A (en) * | 2019-03-21 | 2019-06-25 | 腾讯科技(深圳)有限公司 | Model compression method, apparatus, computer equipment and storage medium |
CN109934300B (en) * | 2019-03-21 | 2023-08-25 | 腾讯科技(深圳)有限公司 | Model compression method, device, computer equipment and storage medium |
CN110766131A (en) * | 2019-05-14 | 2020-02-07 | 北京嘀嘀无限科技发展有限公司 | Data processing device and method and electronic equipment |
CN112016681B (en) * | 2019-05-31 | 2024-04-30 | 苹果公司 | Decomposition of machine learning operations |
US11836635B2 (en) | 2019-05-31 | 2023-12-05 | Apple Inc. | Mutable parameters for machine learning models during runtime |
CN112016681A (en) * | 2019-05-31 | 2020-12-01 | 苹果公司 | Decomposition of machine learning operations |
CN112749782A (en) * | 2019-10-31 | 2021-05-04 | 上海商汤智能科技有限公司 | Data processing method and related product |
CN111062396B (en) * | 2019-11-29 | 2022-03-25 | 深圳云天励飞技术有限公司 | License plate number recognition method and device, electronic equipment and storage medium |
CN111062396A (en) * | 2019-11-29 | 2020-04-24 | 深圳云天励飞技术有限公司 | License plate number recognition method and device, electronic equipment and storage medium |
CN113326913A (en) * | 2020-02-28 | 2021-08-31 | 上海商汤智能科技有限公司 | Neural network model conversion method, model precision positioning method and device |
CN111626409A (en) * | 2020-07-30 | 2020-09-04 | 江西高创保安服务技术有限公司 | Data generation method for image quality detection |
CN112102281A (en) * | 2020-09-11 | 2020-12-18 | 哈尔滨市科佳通用机电股份有限公司 | Truck brake cylinder fault detection method based on improved Faster Rcnn |
CN112102281B (en) * | 2020-09-11 | 2021-07-06 | 哈尔滨市科佳通用机电股份有限公司 | Truck brake cylinder fault detection method based on improved Faster Rcnn |
CN112288829A (en) * | 2020-11-03 | 2021-01-29 | 中山大学 | Compression method and device for image restoration convolutional neural network |
CN112734007A (en) * | 2020-12-31 | 2021-04-30 | 青岛海尔科技有限公司 | Method and device for acquiring compression model, storage medium and electronic device |
CN113657592A (en) * | 2021-07-29 | 2021-11-16 | 中国科学院软件研究所 | Software-defined satellite self-adaptive pruning model compression method |
CN113657592B (en) * | 2021-07-29 | 2024-03-05 | 中国科学院软件研究所 | Software-defined satellite self-adaptive pruning model compression method |
CN113762510A (en) * | 2021-09-09 | 2021-12-07 | 北京百度网讯科技有限公司 | Data processing method and device for target model, electronic equipment and medium |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108229646A (en) | neural network model compression method, device, storage medium and electronic equipment | |
CN113011499B (en) | Hyperspectral remote sensing image classification method based on double-attention machine system | |
CN110532996A (en) | The method of visual classification, the method for information processing and server | |
CN107194426A (en) | A kind of image-recognizing method based on Spiking neutral nets | |
CN107944556A (en) | Deep neural network compression method based on block item tensor resolution | |
CN107609642A (en) | Computing device and method | |
CN107977704A (en) | Weighted data storage method and the neural network processor based on this method | |
CN107392255A (en) | Generation method, device, computing device and the storage medium of minority class picture sample | |
CN107578014A (en) | Information processor and method | |
CN108629326A (en) | The action behavior recognition methods of objective body and device | |
CN108573399A (en) | Method and its system are recommended by trade company based on transition probability network | |
CN111091493B (en) | Image translation model training method, image translation method and device and electronic equipment | |
CN107944358A (en) | A kind of human face generating method based on depth convolution confrontation network model | |
CN105469376A (en) | Method and device for determining picture similarity | |
CN107944545A (en) | Computational methods and computing device applied to neutral net | |
CN113012811B (en) | Traditional Chinese medicine syndrome diagnosis and health evaluation method combining deep convolutional network and graph neural network | |
CN105989849A (en) | Speech enhancement method, speech recognition method, clustering method and devices | |
Kumar et al. | MobiHisNet: a lightweight CNN in mobile edge computing for histopathological image classification | |
CN109978077A (en) | Visual identity methods, devices and systems and storage medium | |
CN108229300A (en) | Video classification methods, device, computer readable storage medium and electronic equipment | |
CN108235003A (en) | Three-dimensional video quality evaluation method based on 3D convolutional neural networks | |
CN107169574A (en) | Using nested machine learning model come the method and system of perform prediction | |
CN110322418A (en) | A kind of super-resolution image generates the training method and device of confrontation network | |
CN111382840B (en) | HTM design method based on cyclic learning unit and oriented to natural language processing | |
CN109889923A (en) | Utilize the method for combining the layering of video presentation to summarize video from attention network |
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 | ||
RJ01 | Rejection of invention patent application after publication | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20180629 |