CN110020718A - The layer-by-layer neural networks pruning method and system inferred based on variation - Google Patents
The layer-by-layer neural networks pruning method and system inferred based on variation Download PDFInfo
- Publication number
- CN110020718A CN110020718A CN201910195272.4A CN201910195272A CN110020718A CN 110020718 A CN110020718 A CN 110020718A CN 201910195272 A CN201910195272 A CN 201910195272A CN 110020718 A CN110020718 A CN 110020718A
- Authority
- CN
- China
- Prior art keywords
- neural network
- noise
- layer
- training
- indicate
- 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
- G06N3/082—Learning methods modifying the architecture, e.g. adding, deleting or silencing nodes or connections
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 Processing (AREA)
Abstract
The present invention provides a kind of layer-by-layer neural networks pruning method and systems inferred based on variation to obtain the neural network for making an uproar this method comprises: infusing noise in neural network by sample mode;It is trained according to weight of the preset objective function to the neural network for making an uproar, the neural network weight after being trained and the neural network after training;The variation lower bound inferred according to variation, the corresponding noise parameter of multiplying property Gaussian noise of training injection, the noise parameter after being trained;Based on the noise parameter after the training, the neural network weight after training, corresponding neuron or convolution kernel in the neural network after the training is successively deleted by threshold function table.In method of the invention, the noise injected has incorporated the hierarchical relationship of neural network in the training process, so that fully considering the dependence of interlayer during beta pruning, then ensure that the output result of the neural network under height beta pruning still has robustness.
Description
Technical field
The present invention relates to depth learning technology fields, and in particular, to the layer-by-layer neural networks pruning inferred based on variation
Method and system.
Background technique
In deep learning field, the powerful Function Fitting ability of neural network enables it ideally to complete such as image point
The tasks such as class, object detection, speech recognition.However when the weight quantity of neural network is excessively huge, generally require to expend huge
Big storage and computing cost, therefore be difficult to apply in the equipment that the computing resources such as mobile terminal, embedded board are limited.For reality
The compression and acceleration of existing neural network, technical staff propose the schemes such as beta pruning, quantization, low-rank decomposition.
It is the mainstream scheme for carrying out neural network compression and acceleration currently, carrying out beta pruning processing to neural network, it is main
Thought is to delete in neural network to influence minimum neuron or convolution kernel to output result.Based on this all kinds of algorithms, usually
It is to be positioned using regular terms, Hessian matrix etc. and delete neuron or convolution kernel.But guaranteeing neural network precision feelings
Under condition, realize that large-scale beta pruning remains the difficult point of research.Recently, some pruning algorithms have been embodied in precision and compression ratio
Between it is well balanced, but mostly for didactic design, lack theories integration.As Bayesian frame is in deep learning field
Gradually develop, variation is inferred as a kind of approximate Bayesian inference, starts to attract attention.Its good behaviour under particular task
Behind, be complete mathematical theory basis.To assign this theoretical basis to pruning algorithms, newest some researchs are begun trying
By injecting optimizable sparse noise to neural network, to realize the height beta pruning under low loss of significance.
But the interlayer relation in neural network is often had ignored, is then influenced to reduce difficulty in computation in the above method
The compression ratio and precision of neural network.
Summary of the invention
For the defects in the prior art, the object of the present invention is to provide a kind of layer-by-layer neural networks inferred based on variation
Pruning method and system.
In a first aspect, the embodiment of the present invention provides a kind of layer-by-layer neural networks pruning method inferred based on variation, comprising:
Noise is infused in neural network by sample mode, obtains the neural network for making an uproar;
It is trained according to weight of the preset objective function to the neural network for making an uproar, the neural network after being trained
Neural network after weight and training;
The variation lower bound inferred according to variation, the corresponding noise parameter of multiplying property Gaussian noise of training injection, obtains
Noise parameter after training;
Based on the noise parameter after the training, the neural network weight after training, institute is successively deleted by threshold function table
Corresponding neuron or convolution kernel in neural network after stating training.
Optionally, noise is injected in neural network by sample mode, obtains the neural network for making an uproar, comprising:
Input in l layers of jth dimension of neural network is denoted as:It then enables defeated in l layers of jth dimension of neural network
EnterMultiplied by noiseWherein,Indicate the noise injected in l layers of jth dimension,Follow following Gaussian Profile:
∈ follows standard gaussian distribution:
Wherein:ForCorresponding noise parameter.It is mutually indepedent for the noise of different dimension injections, it is optional
Ground, if input is the multi-channel feature figure from convolutional layer,It can be defined as the characteristic pattern in jth channel.
Optionally, it is trained according to weight of the preset objective function to the neural network for making an uproar, after being trained
Neural network after neural network weight and training, comprising:
Assuming that the weight of neural network is W, the noise of l layers of neural network injection is θl, then the power of the neural network
Weight optimal solution W*Calculation formula it is as follows:
Wherein: p (y | x, W, θl) indicate input data be x, weight W, injection noise be θlWhen, neural network output
Value is the probability of y;D indicates the data set of training, and wherein x is input data, and y is label;Indicate that total L layers of neural network exists
Output valve on y-dimension;Indicate output valve of total L layers of the neural network in the y ' dimension, the value range of y ' is total
All dimensions of L layers of neural network.
Optionally, the variation lower bound inferred according to variation, the corresponding noise ginseng of the multiplying property Gaussian noise of training injection
Number, the noise parameter after being trained, comprising:
The calculation formula for maximizing variation lower bound is as follows:
Wherein:It indicates to maximize variation lower bound, rlIndicate l layers of noise parameter, JlIndicate l layers
Dimension is inputted,It indicatesIt arrivesKullback-Leibler divergence;Indicate noiseProbability density function,Indicate priori of l layers of the noise in jth dimension;It indicates to inject in l layers of jth dimension
Noise;
The Gaussian prior distribution that induction sparsity is arranged is as follows:
Wherein:Indicate that set of real numbers, σ indicate standard deviation;
Noise parameter after being trained is as follows:
Wherein:Indicate rlOptimal solution,Indicate rlOptimal solution first dimension element,Indicate rlIt is optimal
The last one-dimensional element of solution, arg max indicate maximizing operation.
After the completion of trainingThe as approximate Posterior distrbutionp of noise, to guarantee that sparsity is successfully induced into the distribution,
Selection uses minimum variance, such as σ on designed prior distribution2=0.01.
For the calculating convenient for variation lower bound, Kullback-Leibler divergence can be done to following abbreviation:
Optionally, layer-by-layer by threshold function table based on the noise parameter after the training, the neural network weight after training
Corresponding neuron or convolution kernel in neural network after deleting the training, comprising:
After the completion of l layers of weight and the noise parameter training of neural network, pass through the modification injection of following threshold function table
Noise:
Wherein,Indicate the noise injected in l layers of jth dimension,Indicate the jth dimension element of l layers of noise parameter,
Beta pruning processing is carried out to l layers of neuron or convolution kernel by modified injection noise, calculation formula is as follows:
Wherein: on the left of equationFor l layers after rarefaction of j-th of neuron or convolution kernel;On the right side of equationTable
Show j-th of neuron or convolution kernel before rarefaction.
By the operation of above-mentioned rarefaction, the fractional weight of neural network will be set to 0, then corresponding input will be lost
It abandons, is no longer participate in the operation of neural network, delete neuron or convolution kernel to realize, i.e. the effect of beta pruning.By from mind
Start through network first tier, to the last terminates for one layer, the neural network that final output beta pruning is completed.Wherein, because it is designed
Layer-by-layer training and beta pruning form, the noise injected incorporated the hierarchical relationship of neural network in the training process so that
The dependence of interlayer is fully considered during beta pruning, then ensure that the output result of the neural network under height beta pruning still has
Standby robustness.
Second aspect, the embodiment of the present invention provide a kind of layer-by-layer neural networks pruning system inferred based on variation, comprising:
Processor and memory are stored with computer program in the memory, when the computer program is called by the processor,
The processor executes the layer-by-layer neural networks pruning method inferred based on variation as described in any one of first aspect.
Compared with prior art, the present invention have it is following the utility model has the advantages that
The layer-by-layer neural networks pruning method and system provided by the invention inferred based on variation is trained and is cut using layer-by-layer
The mode of branch, combining between neural network middle layer has the characteristics that dependence, can be by each interlayer relation in training completion
It is dissolved into noise parameter, to more precisely position the neuron or convolution kernel of redundancy, ensure that refreshing under height beta pruning
Output result through network still has robustness.
Detailed description of the invention
Upon reading the detailed description of non-limiting embodiments with reference to the following drawings, other feature of the invention,
Objects and advantages will become more apparent upon:
Fig. 1 is the schematic illustration of the layer-by-layer neural networks pruning method provided by the invention inferred based on variation;
Fig. 2 is the structural schematic diagram of the layer-by-layer neural networks pruning device provided by the invention inferred based on variation.
Specific embodiment
The present invention is described in detail combined with specific embodiments below.Following embodiment will be helpful to the technology of this field
Personnel further understand the present invention, but the invention is not limited in any way.It should be pointed out that the ordinary skill of this field
For personnel, without departing from the inventive concept of the premise, several changes and improvements can also be made.These belong to the present invention
Protection scope.
As shown in Figure 1, for the flow chart for the layer-by-layer neural networks pruning embodiment of the method inferred the present invention is based on variation, it should
Designed Gaussian noise is injected neural network by injection noise step by method, and is walked respectively in neural network weight training
Suddenly it and in noise parameter training step, for particular task training network weight and noise parameter, is finally utilized in beta pruning step
The noise parameter that training is completed realizes beta pruning by threshold function table.Above-mentioned process is simultaneously successively repeated since neural network first layer
It carries out, to the last terminates for one layer.
The present invention can realize the height beta pruning of neural network under low loss of significance.Because being inferred based on variation, this hair
It is bright to absorb the complete theoretical basis of Bayesian frame, simultaneously because the mode of designed layer-by-layer training and beta pruning, in theory
On the basis of combine between neural network middle layer and have the characteristics that dependence, this interlayer relation can be dissolved into when training is completed
In noise parameter, the neuron or convolution kernel of redundancy are more precisely positioned.
Specifically, referring to Fig.1, described method includes following steps:
Injection noise step: designing a kind of Gaussian noise of multiplying property, injects neural network by sampling;
Neural network weight training step: according to the specific objective function training neural network weight of task;
Noise parameter training step: the parameter of injected noise is trained according to the variation lower bound that variation is inferred;
Beta pruning step: the noise parameter completed according to training deletes phase in trained neural network by threshold function table
The neuron or convolution kernel answered.
Corresponding to the above method, referring to Fig. 2, a kind of layer-by-layer neural networks pruning device inferred based on variation, comprising:
Injection noise module: designing a kind of Gaussian noise of multiplying property, injects neural network by sampling;
Neural network weight training module: according to the specific objective function training neural network weight of task;
Noise parameter training module: the parameter of injected noise is trained according to the variation lower bound that variation is inferred;
Pruning module: the noise parameter completed according to training deletes phase in trained neural network by threshold function table
The neuron or convolution kernel answered.
The specific implementation of above-mentioned each step and module is described in detail below, to understand the technology of the present invention side
Case.
In section Example of the present invention, the injection noise step, in which: for the random layer of neural network, design
A kind of multiplying property Gaussian noise inputted with this layer with dimension, when neural network is calculated, which will be by sampling with phase
The mode multiplied injects the input of this layer.For that designed noise can be converted into row equivalent parametrization convenient for sampling.
In section Example of the present invention, the neural network weight training step, in which: be directed to any training data
Collection, the training objective of neural network weight are the probability for maximizing known input data and exporting corresponding label.
In section Example of the present invention, the noise parameter training step, in which: inferred using variation and change is calculated
Divide lower bound, the training objective of noise parameter is to maximize the variation lower bound.The noise that training is completed is the approximation of its Posterior distrbutionp
Solution.
In section Example of the present invention, the beta pruning step is specific as follows: the noise inputs threshold value that training is completed
Function obtains new noise, then realizes beta pruning with the neural network weight of new noise rarefaction respective layer in the form of multiplication
Effect.The input of this layer will no longer be required to multiplied by noise after rarefaction operation.
As shown in Fig. 2, beta pruning system is by injection noise module, neural network weight training module, noise parameter training mould
Block, pruning module composition, whole system frame can be trained end-to-endly.
In the system framework of embodiment as shown in Figure 2, for l layers of neural network, Gaussian noiseThrough over-sampling
It is multiplied by j-th of neuron or input channel of this layer afterwards, noise is distributed as
Wherein,For noise parameter, and it is mutually indepedent for different j.
In the system framework of embodiment as shown in Figure 2, the objective function for training noise parameter is
Wherein, D is training dataset, and x is input data, and y is label, and W is neural network weight, θlForSet, Jl
For l layers of input dimension or port number, Kullback-Leibler divergence calculation formula is
Then the training objective of noise parameter is
In the system framework of embodiment shown in Fig. 2, the objective function of training neural network weight is training data
Maximum likelihood, calculation formula are
Wherein,The output valve signed for the L layer neural network trained in y category.The then instruction of neural network weight
Practicing target is
The training of noise parameter starts from neural network first layer, works as objective functionWhen convergence, terminate to current
The update of the noise parameter of layer, and it is passed through into following threshold function table:
Further, due toIt will be no longer stochastic variable by threshold function table assignment as above, so do not need past
Through over-sampling in training process afterwards.Particularly, forThe case where, noise is directly replicated to 0, then corresponding
Neuron or channel be considered as redundancy and explicitly leave out, wherein generating the convolution kernel in these channels can also directly leave out, with
This realizes the beta pruning effect to this layer of neural network;ForThe case where, for the expense for reducing storage parameter, it is also convenient for
Neural computing, the input of this layer will no longer be multiplied by noise, and the weight of this layer will be directly by noise rarefaction, calculation
It is as follows:
Wherein,For the jth row of l layers of weight of neural network.
In the system framework of embodiment shown in Fig. 2, terminate the noise parameter training of this layer, the network weight of current round
After retraining and beta pruning, entire training and beta pruning process will be repeatedly applied on next layer of neural network, until neural network
The last layer.So form of successively training and beta pruning makes injected noise incorporate neural network in the training process
Hierarchical relationship.
To sum up, the present invention is using injected noise as the index of redundancy, after inferring that approximation acquires noise using variation
It tests distribution and positions the neuron or convolution kernel of redundancy by threshold function table later, and carry out beta pruning.Simultaneously as trained and beta pruning
It is successively to carry out, the noise injected has incorporated the hierarchical relationship of neural network in the training process, so that during beta pruning
It fully considers the dependence of interlayer, then ensure that the output result of the neural network under height beta pruning still has robustness.
It should be noted that the step in the layer-by-layer neural networks pruning method inferred based on variation provided by the invention
Suddenly, it can use corresponding module, device, unit etc. in the layer-by-layer neural networks pruning system inferred based on variation to give
It realizes, the technical solution that those skilled in the art are referred to the system realizes the step process of the method, that is, the system
Embodiment in system can be regarded as realizing the preference of the method, and it will not be described here.
One skilled in the art will appreciate that in addition to realizing system provided by the invention in a manner of pure computer readable program code
And its other than each device, completely can by by method and step carry out programming in logic come so that system provided by the invention and its
Each device is in the form of logic gate, switch, specific integrated circuit, programmable logic controller (PLC) and embedded microcontroller etc.
To realize identical function.So system provided by the invention and its every device are considered a kind of hardware component, and it is right
The device for realizing various functions for including in it can also be considered as the structure in hardware component;It can also will be for realizing each
The device of kind function is considered as either the software module of implementation method can be the structure in hardware component again.
Specific embodiments of the present invention are described above.It is to be appreciated that the invention is not limited to above-mentioned
Particular implementation, those skilled in the art can make a variety of changes or modify within the scope of the claims, this not shadow
Ring substantive content of the invention.In the absence of conflict, the feature in embodiments herein and embodiment can any phase
Mutually combination.
Claims (6)
1. a kind of layer-by-layer neural networks pruning method inferred based on variation characterized by comprising
Noise is infused in neural network by sample mode, obtains the neural network for making an uproar;
It is trained according to weight of the preset objective function to the neural network for making an uproar, the neural network weight after being trained
With the neural network after training;
The variation lower bound inferred according to variation, the corresponding noise parameter of multiplying property Gaussian noise of training injection, is trained
Noise parameter afterwards;
Based on the noise parameter after the training, the neural network weight after training, the instruction is successively deleted by threshold function table
Corresponding neuron or convolution kernel in neural network after white silk.
2. the layer-by-layer neural networks pruning method according to claim 1 inferred based on variation, which is characterized in that by adopting
Sample loading mode injects noise in neural network, obtains the neural network for making an uproar, comprising:
Input in l layers of jth dimension of neural network is denoted as:Then enable the input in l layers of jth dimension of neural network
Multiplied by noiseWherein,Indicate the noise injected in l layers of jth dimension,Follow following Gaussian Profile:
∈ follows standard gaussian distribution:
Wherein:ForCorresponding noise parameter.
3. the layer-by-layer neural networks pruning method according to claim 1 inferred based on variation, which is characterized in that according to pre-
If objective function the weight of the neural network for making an uproar is trained, the neural network weight after being trained and training after
Neural network, comprising:
Assuming that the weight of neural network is W, the noise of l layers of neural network injection is θl, then the weight of the neural network is optimal
Solve W*Calculation formula it is as follows:
Wherein: p (y | x, W, θl) indicate input data be x, weight W, injection noise be θlWhen, neural network output valve is y
Probability;D indicates the data set of training, and wherein x is input data, and y is label;Indicate that total L layers of neural network is tieed up in y
Output valve on degree;Indicate that output valve of total L layers of the neural network in the y ' dimension, the value range of y ' are total L layers
All dimensions of neural network.
4. the layer-by-layer neural networks pruning method according to claim 3 inferred based on variation, which is characterized in that according to change
Divide the variation lower bound inferred and obtained, the corresponding noise parameter of multiplying property Gaussian noise of training injection, the noise ginseng after being trained
Number, comprising:
The calculation formula for maximizing variation lower bound is as follows:
Wherein:It indicates to maximize variation lower bound, rlIndicate l layers of noise parameter, JlIndicate l layers of input
Dimension,It indicatesIt arrivesKullback-Leibler divergence;Indicate noise's
Probability density function,Indicate priori of l layers of the noise in jth dimension;Indicate that injects in l layers of jth dimension makes an uproar
Sound;
The Gaussian prior distribution that induction sparsity is arranged is as follows:
Wherein:Indicate that set of real numbers, σ indicate standard deviation;
Noise parameter after being trained is as follows:
Wherein:Indicate rlOptimal solution,Indicate rlOptimal solution first dimension element,Indicate rlOptimal solution
Last one-dimensional element, argmax indicate maximizing operation.
5. the layer-by-layer neural networks pruning method inferred described in any one of -4 based on variation according to claim 1, feature
It is, based on the noise parameter after the training, the neural network weight after training, the instruction is successively deleted by threshold function table
Corresponding neuron or convolution kernel in neural network after white silk, comprising:
After the completion of l layers of weight and the noise parameter training of neural network, pass through the noise of following threshold function table modification injection:
Wherein,Indicate the noise injected in l layers of jth dimension,Indicate the jth dimension element of l layers of noise parameter,
Beta pruning processing is carried out to l layers of neuron or convolution kernel by modified injection noise, calculation formula is as follows:
Wherein: on the left of equationFor l layers after rarefaction of j-th of neuron or convolution kernel;On the right side of equationIndicate dilute
Preceding j-th of the neuron of thinization or convolution kernel.
6. a kind of layer-by-layer neural networks pruning system inferred based on variation characterized by comprising processor and memory,
It is stored with computer program in the memory, when the computer program is called by the processor, the processor is executed
The layer-by-layer neural networks pruning method according to any one of claims 1 to 5 inferred based on variation.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910195272.4A CN110020718A (en) | 2019-03-14 | 2019-03-14 | The layer-by-layer neural networks pruning method and system inferred based on variation |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910195272.4A CN110020718A (en) | 2019-03-14 | 2019-03-14 | The layer-by-layer neural networks pruning method and system inferred based on variation |
Publications (1)
Publication Number | Publication Date |
---|---|
CN110020718A true CN110020718A (en) | 2019-07-16 |
Family
ID=67189619
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910195272.4A Pending CN110020718A (en) | 2019-03-14 | 2019-03-14 | The layer-by-layer neural networks pruning method and system inferred based on variation |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110020718A (en) |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110766044A (en) * | 2019-09-11 | 2020-02-07 | 浙江大学 | Neural network training method based on Gaussian process prior guidance |
CN110826713A (en) * | 2019-10-25 | 2020-02-21 | 广州思德医疗科技有限公司 | Method and device for acquiring special convolution kernel |
CN111242287A (en) * | 2020-01-15 | 2020-06-05 | 东南大学 | Neural network compression method based on channel L1 norm pruning |
CN111582446A (en) * | 2020-04-28 | 2020-08-25 | 北京达佳互联信息技术有限公司 | System for neural network pruning and neural network pruning processing method |
CN112215353A (en) * | 2020-09-29 | 2021-01-12 | 电子科技大学 | Channel pruning method based on variational structure optimization network |
WO2021114859A1 (en) * | 2019-12-09 | 2021-06-17 | 清华大学 | Method and device for implementing bayesian neural network by using memristor intrinsic noise |
-
2019
- 2019-03-14 CN CN201910195272.4A patent/CN110020718A/en active Pending
Cited By (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110766044A (en) * | 2019-09-11 | 2020-02-07 | 浙江大学 | Neural network training method based on Gaussian process prior guidance |
CN110766044B (en) * | 2019-09-11 | 2021-10-26 | 浙江大学 | Neural network training method based on Gaussian process prior guidance |
CN110826713A (en) * | 2019-10-25 | 2020-02-21 | 广州思德医疗科技有限公司 | Method and device for acquiring special convolution kernel |
CN110826713B (en) * | 2019-10-25 | 2022-06-10 | 广州思德医疗科技有限公司 | Method and device for acquiring special convolution kernel |
WO2021114859A1 (en) * | 2019-12-09 | 2021-06-17 | 清华大学 | Method and device for implementing bayesian neural network by using memristor intrinsic noise |
CN111242287A (en) * | 2020-01-15 | 2020-06-05 | 东南大学 | Neural network compression method based on channel L1 norm pruning |
CN111582446A (en) * | 2020-04-28 | 2020-08-25 | 北京达佳互联信息技术有限公司 | System for neural network pruning and neural network pruning processing method |
CN111582446B (en) * | 2020-04-28 | 2022-12-06 | 北京达佳互联信息技术有限公司 | System for neural network pruning and neural network pruning processing method |
CN112215353A (en) * | 2020-09-29 | 2021-01-12 | 电子科技大学 | Channel pruning method based on variational structure optimization network |
CN112215353B (en) * | 2020-09-29 | 2023-09-01 | 电子科技大学 | Channel pruning method based on variational structure optimization network |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110020718A (en) | The layer-by-layer neural networks pruning method and system inferred based on variation | |
US20190354895A1 (en) | Learning data augmentation policies | |
US11854248B2 (en) | Image classification method, apparatus and training method, apparatus thereof, device and medium | |
CN106709565A (en) | Optimization method and device for neural network | |
CN108009594B (en) | A kind of image-recognizing method based on change grouping convolution | |
CN112163601B (en) | Image classification method, system, computer device and storage medium | |
CN110458084B (en) | Face age estimation method based on inverted residual error network | |
CN109086768A (en) | The semantic image dividing method of convolutional neural networks | |
CN111079899A (en) | Neural network model compression method, system, device and medium | |
DE112020002042T5 (en) | GENERATION OF AN INTENT RECOGNITION MODEL BASED ON RANDOMIZED INTENT VECTOR APPROACHES | |
CN109754359A (en) | A kind of method and system that the pondization applied to convolutional neural networks is handled | |
CN110516537A (en) | A kind of face age estimation method based on from step study | |
CN106991999A (en) | Audio recognition method and device | |
CN106803092B (en) | Method and device for determining standard problem data | |
CN109460813A (en) | Accelerated method, device, equipment and the storage medium that convolutional neural networks calculate | |
CN106297778A (en) | The neutral net acoustic model method of cutting out based on singular value decomposition of data-driven | |
CN116775843A (en) | Question-answer pair evaluation data generation method, question-answer pair evaluation data generation device, computer equipment and storage medium | |
CN106709566A (en) | Deep learning-based data missing value refilling method | |
CN111079691A (en) | Pruning method based on double-flow network | |
CN117350373B (en) | Personalized federal aggregation algorithm based on local self-attention mechanism | |
CN110222817A (en) | Convolutional neural networks compression method, system and medium based on learning automaton | |
CN109359542A (en) | The determination method and terminal device of vehicle damage rank neural network based | |
CN116384471A (en) | Model pruning method, device, computer equipment, storage medium and program product | |
CN115936108A (en) | Knowledge distillation-based neural network compression method for multivariate time series prediction graph | |
CN115983366A (en) | Model pruning method and system for federal learning |
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: 20190716 |