CN107368885A - Network model compression method and device based on more granularity beta prunings - Google Patents
Network model compression method and device based on more granularity beta prunings Download PDFInfo
- Publication number
- CN107368885A CN107368885A CN201710568710.8A CN201710568710A CN107368885A CN 107368885 A CN107368885 A CN 107368885A CN 201710568710 A CN201710568710 A CN 201710568710A CN 107368885 A CN107368885 A CN 107368885A
- Authority
- CN
- China
- Prior art keywords
- beta pruning
- pruning
- granularity level
- inessential
- network model
- 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
Abstract
The invention provides the network model compression method based on more granularity beta prunings, this method includes following one or two or three steps：Input channel granularity level beta pruning step, using inessential element beta pruning method, inessential element in the granularity level of the input channel of network model is subjected to beta pruning；Convolution kernel granularity level beta pruning step, using inessential element beta pruning method, beta pruning is carried out to inessential element in the granularity level of the convolution kernel of network model；Weight parameter granularity level beta pruning step, using inessential element beta pruning method, beta pruning is carried out to inessential element in the granularity level of the weight parameter of network model.Compared with prior art, the present invention can effectively solve network model compression problem by carrying out beta pruning to more granularity levels.
Description
Technical field
The present invention relates to image procossing, video monitoring and deep neural network, more particularly to based on more granularity beta prunings
Network model compression method and device.
Background technology
In recent years, with the fast development of artificial intelligence, deep learning network by combining low-level image feature due to forming height
Layer feature, is influenceed smaller by environmental change, breakthrough achievement is achieved in computer vision field, especially in recognition of face and
Image classification etc. has surmounted the recognition accuracy of the mankind.
However, existing high-performance deep learning network typically has millions of or even more than one hundred million individual parameters, this causes its
Storage and calculating consumption are all very huge, limit its equipment for being applied to store and computing resource is all limited.Therefore, to depth
It is a committed step for solving this problem that learning network model compression, which is compressed,.
But existing model compression technology reduces moulded dimension typically by the rarefaction of Model Weight value, but
Be can not significantly decrease operation deep learning network required for storage and computing resource.
In summary, it is necessary to propose a kind of deep learning network model compression side for reducing storage and computing resource consumption
Method.
The content of the invention
In view of this, it is a primary object of the present invention to reduce storage resource and computing resource consumption, network model is realized
Compression.
To reach above-mentioned purpose, according to the first aspect of the invention, there is provided the network model based on more granularity beta prunings
Compression method, this method include following one or two or three steps：
Input channel granularity level beta pruning step, using inessential element beta pruning method, by the input channel of network model
Inessential element carries out beta pruning in granularity level；
Convolution kernel granularity level beta pruning step, using inessential element beta pruning method, to the granularity of the convolution kernel of network model
Inessential element carries out beta pruning in level；
Weight parameter granularity level beta pruning step, using inessential element beta pruning method, to the weight parameter of network model
Inessential element carries out beta pruning in granularity level.
Further, the inessential element beta pruning method includes：
The zero setting step of inessential element, the importance of each element in current granularity level is calculated, by inessential element
Corresponding value zero setting；
Beta pruning trim step, whole network model is finely tuned according to training data；
Judgment step is lost, calculates the loss of accuracy of the network model after fine setting, if the loss of accuracy is less than essence
Exactness loses threshold value, then continues executing with the zero setting step of inessential element, otherwise terminate.
Further, the zero setting step of the inessential element includes：
Element importance sorting step, count the weight parameter vector W that beta pruning element is each treated in current granularity levele,
The importance of beta pruning element is treated described in calculatingTo needed beta pruning element according to importance EIViRisen from small to large
Sequence arranges, and obtains importance ascending order set, and calculate the importance summation of needed beta pruning element
N is the quantity of current granularity level weight parameter；
Pruning threshold calculation procedure, according to energy beta pruning rate threshold value EPR, calculate beta pruning ENERGY E P=EIVS × EPR, statistics
The cumulative distribution of importance in ascending order set, the importance corresponding to the cumulative distribution equal with beta pruning energy is chosen as beta pruning
Threshold value；
Loss function influence amount calculation procedure, one group of test data is inputted, the loss function value Loss of calculating network, is calculated
Loss function influence amount
Zero setting step, for each treating beta pruning element i in current granularity level, if treating the importance EIV of beta pruning elementi
Less than pruning threshold, and loss function influence amount ELiLess than 0, then this is treated to be worth zero setting corresponding to beta pruning element i.
According to another aspect of the present invention, there is provided the network model compression set based on more granularity beta prunings, the device
Including following one or two or three modules：
Input channel granularity level pruning module, for using inessential element pruning module, by the input of network model
Inessential element carries out beta pruning in the granularity level of passage；
Convolution kernel granularity level pruning module, for using inessential element pruning module, to the convolution kernel of network model
Granularity level in inessential element carry out beta pruning；
Weight parameter granularity level pruning module, for using inessential element pruning module, to the weight of network model
Inessential element carries out beta pruning in the granularity level of parameter.
Further, the inessential element pruning module includes：
The zero setting module of inessential element, will be inessential for calculating the importance of each element in current granularity level
Value zero setting corresponding to element；
Module is finely tuned in beta pruning, for finely tuning whole network model according to training data；
Judge module is lost, for calculating the loss of accuracy of the network model after finely tuning, if the loss of accuracy is small
In loss of accuracy's threshold value, then the zero setting module of inessential element is continued executing with, is otherwise terminated.
Further, the zero setting module of the inessential element includes：
Element importance sorting module, the weight parameter vector W of beta pruning element is each treated in current granularity level for countinge,
The importance of beta pruning element is treated described in calculatingTo needed beta pruning element according to importance EIViAscending order is carried out from small to large
Arrangement, obtains importance ascending order set, and calculate the importance summation of needed beta pruning element
N is the quantity of current granularity level weight parameter；
Pruning threshold computing module, for according to energy beta pruning rate threshold value EPR, calculating beta pruning ENERGY E P=EIVS × EPR,
The cumulative distribution of importance in ascending order set is counted, chooses the importance conduct corresponding to the cumulative distribution equal with beta pruning energy
Pruning threshold；
Loss function influence amount computing module, for input one group of test data, the loss function value Loss of calculating network,
Counting loss function influences amount
Zero setting module, for for each treating beta pruning element i in current granularity level, if treating the importance of beta pruning element
EIViLess than pruning threshold, and loss function influence amount ELiLess than 0, then this is treated to be worth zero setting corresponding to beta pruning element i.
Compared with existing network model compression method, the model compression method of the invention based on more granularity beta prunings, adopt
, not only can be with compression network moulded dimension, and due to network model with the method for one or more kinds of granularity level beta prunings
Sparse format is regular, it is possible to reduce the calculating consumption of network.
Brief description of the drawings
Fig. 1 shows the flow of the network model compression method based on more granularity beta prunings according to one embodiment of the invention
Figure.
Fig. 2 shows the structure of the network model compression set based on more granularity beta prunings according to one embodiment of the invention
Schematic diagram.
Embodiment
To enable your auditor to further appreciate that structure, feature and the other purposes of the present invention, in conjunction with appended preferably real
Apply example describe in detail it is as follows, illustrated preferred embodiment is merely to illustrate technical scheme, and the non-limiting present invention.
According to the present invention the network model compression method based on more granularity beta prunings include it is following one or two or
Three steps：
Input channel granularity level beta pruning step, using inessential element beta pruning method, by the input channel of network model
Inessential element carries out beta pruning in granularity level；
Convolution kernel granularity level beta pruning step, using inessential element beta pruning method, to the granularity of the convolution kernel of network model
Inessential element carries out beta pruning in level；
Weight parameter granularity level beta pruning step, using inessential element beta pruning method, to the weight parameter of network model
Inessential element carries out beta pruning in granularity level.
Fig. 1 gives the flow of the network model compression method based on more granularity beta prunings according to one embodiment of the invention
Figure.As shown in figure 1, include according to the network model compression method based on more granularity beta prunings of the present invention：
Input channel granularity level beta pruning step S1, using inessential element beta pruning method S10, the input of network model is led to
Inessential element carries out beta pruning in the granularity level in road；
Convolution kernel granularity level beta pruning step S2, using inessential element beta pruning method S10, to the convolution kernel of network model
Inessential element carries out beta pruning in granularity level；
Weight parameter granularity level beta pruning step S3, using inessential element beta pruning method S10, the weight of network model is joined
Inessential element carries out beta pruning in several granularity levels.
Further, the inessential element beta pruning method S10 includes：
The zero setting step S11 of inessential element, the importance of each element in current granularity level is calculated, by inessential member
Value zero setting corresponding to plain；
Beta pruning trim step S12, whole network model is finely tuned according to training data；
Judgment step S13 is lost, the loss of accuracy of the network model after fine setting is calculated, if the loss of accuracy is less than
Loss of accuracy's threshold value, then the zero setting step S11 of inessential element is continued executing with, is otherwise terminated.
Further, the zero setting step S11 of the inessential element includes：
Element importance sorting step S111, count the weight parameter vector W that beta pruning element is each treated in current granularity levele, meter
The importance of beta pruning element is treated described in calculationTo needed beta pruning element according to importance EIViAscending order row is carried out from small to large
Row, obtain importance ascending order set, and calculate the importance summation of needed beta pruning element
N is the quantity of current granularity level weight parameter；
Pruning threshold calculation procedure S112, according to energy beta pruning rate threshold value EPR, beta pruning ENERGY E P=EIVS × EPR is calculated,
The cumulative distribution of importance in ascending order set is counted, chooses the importance conduct corresponding to the cumulative distribution equal with beta pruning energy
Pruning threshold；
Loss function influence amount calculation procedure S113, one group of test data of input, the loss function value Loss of calculating network,
Counting loss function influences amount
Zero setting step S114, for each treating beta pruning element i in current granularity level, if treating the importance of beta pruning element
EIViLess than pruning threshold, and loss function influence amount ELiLess than 0, then this is treated to be worth zero setting corresponding to beta pruning element i.
Further, the span of the energy beta pruning rate threshold value EPR is 0.01~0.2.Preferably, the energy is cut
Branch rate threshold value EPR span is 0.05~0.18.
Specifically, the current grain in the inessential element beta pruning method S10 of the input channel granularity level beta pruning step S1
Spend granularity level of the level for the input channel of network model.The inessential element of the convolution kernel granularity level beta pruning step S2
Current granularity level in beta pruning method S10 is the granularity level of the convolution kernel of network model.The weight parameter granularity level is cut
Current granularity level in branch step S3 inessential element beta pruning method S10 is the granularity level of the weight parameter of network model.
Further, the beta pruning trim step S12 is finely adjusted using gradient descent method to whole network model.It is excellent
Selection of land, the beta pruning trim step S12 are finely adjusted using stochastic gradient descent method to whole network model.
Further, the span of loss of accuracy's threshold value is 0.01~0.1 in the loss judgment step S13.It is excellent
Selection of land, the span of loss of accuracy's threshold value is 0.05~0.08.
Further, the network is deep learning network.Preferably, the network include but is not limited to it is following a kind of or
The a variety of combination of person：Convolutional neural networks, depth belief network, recurrent neural network.
According to the present invention the network model compression set based on more granularity beta prunings include it is following one or two or
Three modules：
Input channel granularity level pruning module, for using inessential element pruning module, by the input of network model
Inessential element carries out beta pruning in the granularity level of passage；
Convolution kernel granularity level pruning module, for using inessential element pruning module, to the convolution kernel of network model
Granularity level in inessential element carry out beta pruning；
Weight parameter granularity level pruning module, for using inessential element pruning module, to the weight of network model
Inessential element carries out beta pruning in the granularity level of parameter.
Fig. 2 gives the structure of the network model compression set based on more granularity beta prunings according to one embodiment of the invention
Schematic diagram.As shown in Fig. 2 include according to the network model compression set based on more granularity beta prunings of the present invention：
Input channel granularity level pruning module 1, for using inessential element pruning module, by the input of network model
Inessential element carries out beta pruning in the granularity level of passage；
Convolution kernel granularity level pruning module 2, for using inessential element pruning module, to the convolution kernel of network model
Granularity level in inessential element carry out beta pruning；
Weight parameter granularity level pruning module 3, for using inessential element pruning module, to the weight of network model
Inessential element carries out beta pruning in the granularity level of parameter.
Further, the inessential element pruning module 10 includes：
The zero setting module 11 of inessential element, for calculating the importance of each element in current granularity level, it will not weigh
Want value zero setting corresponding to element；
Module 12 is finely tuned in beta pruning, for finely tuning whole network model according to training data；
Judge module 13 is lost, for calculating the loss of accuracy of the network model after finely tuning, if the loss of accuracy
Less than loss of accuracy's threshold value, then the zero setting module 11 of inessential element is continued executing with, is otherwise terminated.
Further, the zero setting module 11 of the inessential element includes：
Element importance sorting module 111, the weight parameter vector W of beta pruning element is each treated in current granularity level for countinge,
The importance of beta pruning element is treated described in calculatingTo needed beta pruning element according to importance EIViAscending order row is carried out from small to large
Row, obtain importance ascending order set, and calculate the importance summation of needed beta pruning element
N is the quantity of current granularity level weight parameter；
Pruning threshold computing module 112, for according to energy beta pruning rate threshold value EPR, calculate beta pruning ENERGY E P=EIVS ×
EPR, the cumulative distribution of importance in ascending order set is counted, choose the importance corresponding to the cumulative distribution equal with beta pruning energy
As pruning threshold；
Loss function influence amount computing module 113, for inputting one group of test data, the loss function value of calculating network
Loss, counting loss function influences amount
Zero setting module 114, for for each treating beta pruning element i in current granularity level, if treating the weight of beta pruning element
Spend EIViLess than pruning threshold, and loss function influence amount ELiLess than 0, then this is treated to be worth zero setting corresponding to beta pruning element i.
Compared with existing network model compression method, the model compression method of the invention based on more granularity beta prunings, adopt
, not only can be with compression network moulded dimension, and due to network model with the method for one or more kinds of granularity level beta prunings
Sparse format is regular, it is possible to reduce the calculating consumption of network.
The foregoing is only a preferred embodiment of the present invention, is not intended to limit the scope of the present invention, should
Understand, the present invention is not limited to implementation as described herein, and the purpose of these implementations description is to help this area
In technical staff put into practice the present invention.Any those of skill in the art are easy to do not departing from spirit and scope of the invention
In the case of be further improved and perfect, therefore the present invention is only by the content of the claims in the present invention and limiting for scope
System, its intention cover all alternatives being included in the spirit and scope of the invention being defined by the appended claims and waited
Same scheme.
Claims (10)
1. the network model compression method based on more granularity beta prunings, it is characterised in that this method include it is following one or two
Or three steps：
Input channel granularity level beta pruning step, using inessential element beta pruning method, by the granularity of the input channel of network model
Inessential element carries out beta pruning in level；
Convolution kernel granularity level beta pruning step, using inessential element beta pruning method, to the granularity level of the convolution kernel of network model
In inessential element carry out beta pruning；
Weight parameter granularity level beta pruning step, using inessential element beta pruning method, to the granularity of the weight parameter of network model
Inessential element carries out beta pruning in level.
2. the method as described in claim 1, it is characterised in that the inessential element beta pruning method includes：
The zero setting step of inessential element, the importance of each element in current granularity level is calculated, inessential element is corresponding
Value zero setting；
Beta pruning trim step, whole network model is finely tuned according to training data；
Judgment step is lost, the loss of accuracy of the network model after fine setting is calculated, if the loss of accuracy is less than accuracy
Threshold value is lost, then continues executing with the zero setting step of inessential element, otherwise terminates.
3. method as claimed in claim 2, it is characterised in that the zero setting step of the inessential element includes：Element is important
Sequence step is spent, counts the weight parameter vector W that beta pruning element is each treated in current granularity levele, beta pruning member is treated described in calculating
The importance of elementTo needed beta pruning element according to importance EIViAscending order arrangement is carried out from small to large, obtains weight
Ascending order set is spent, and calculates the importance summation of needed beta pruning elementN
For the quantity of current granularity level weight parameter；
Pruning threshold calculation procedure, according to energy beta pruning rate threshold value EPR, beta pruning ENERGY E P=EIVS × EPR is calculated, counts ascending order
The cumulative distribution of importance in set, the importance corresponding to the cumulative distribution equal with beta pruning energy is chosen as beta pruning threshold
Value；
Loss function influence amount calculation procedure, input one group of test data, the loss function value Loss of calculating network, counting loss
Function influences amount
Zero setting step, for each treating beta pruning element i in current granularity level, if treating the importance EIV of beta pruning elementiIt is less than
Pruning threshold, and loss function influence amount ELiLess than 0, then this is treated to be worth zero setting corresponding to beta pruning element i.
4. method as claimed in claim 2, the inessential element beta pruning method of the input channel granularity level beta pruning step
Current granularity level is the granularity level of the input channel of network model；The convolution kernel granularity level beta pruning step it is inessential
Current granularity level in element beta pruning method is the granularity level of the convolution kernel of network model；The weight parameter granularity level is cut
Current granularity level in the inessential element beta pruning method of branch step is the granularity level of the weight parameter of network model.
5. method as claimed in claim 3, the span of the energy beta pruning rate threshold value EPR is 0.01~0.2.
6. method as claimed in claim 2, the span of loss of accuracy's threshold value is 0.01~0.1.
7. method as claimed in claim 4, the network is deep learning network；Preferably, the network includes but unlimited
In following one or more kinds of combination：Convolutional neural networks, depth belief network, recurrent neural network.
8. the network model compression set based on more granularity beta prunings, it is characterised in that the device include it is following one or two
Or three modules：
Input channel granularity level pruning module, for using inessential element pruning module, by the input channel of network model
Granularity level in inessential element carry out beta pruning；
Convolution kernel granularity level pruning module, for using inessential element pruning module, to the grain of the convolution kernel of network model
Spend inessential element in level and carry out beta pruning；
Weight parameter granularity level pruning module, for using inessential element pruning module, to the weight parameter of network model
Granularity level in inessential element carry out beta pruning.
9. device as claimed in claim 8, it is characterised in that the inessential element pruning module includes：
The zero setting module of inessential element, for calculating the importance of each element in current granularity level, by inessential element
Corresponding value zero setting；
Module is finely tuned in beta pruning, for finely tuning whole network model according to training data；
Judge module is lost, for calculating the loss of accuracy of the network model after finely tuning, if the loss of accuracy is less than essence
Exactness loses threshold value, then continues executing with the zero setting module of inessential element, otherwise terminate.
10. device as claimed in claim 9, it is characterised in that the zero setting module of the inessential element includes：Element is important
Order module is spent, each treats the weight parameter vector W of beta pruning element in current granularity level for countinge, wait to cut described in calculating
The importance of branch elementTo needed beta pruning element according to importance EIViAscending order arrangement is carried out from small to large, is obtained
To importance ascending order set, and calculate the importance summation of needed beta pruning element
N is the quantity of current granularity level weight parameter；Pruning threshold computing module, for according to energy beta pruning rate threshold value EPR, calculating
Beta pruning ENERGY E P=EIVS × EPR, the cumulative distribution of importance in ascending order set is counted, choose equal with beta pruning energy add up
The corresponding importance of distribution is as pruning threshold；
Loss function influence amount computing module, for inputting one group of test data, the loss function value Loss of calculating network, calculate
Loss function influence amount
Zero setting module, for for each treating beta pruning element i in current granularity level, if treating the importance EIV of beta pruning elementi
Less than pruning threshold, and loss function influence amount ELiLess than 0, then this is treated to be worth zero setting corresponding to beta pruning element i.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710568710.8A CN107368885A (en) | 2017-07-13 | 2017-07-13 | Network model compression method and device based on more granularity beta prunings |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710568710.8A CN107368885A (en) | 2017-07-13 | 2017-07-13 | Network model compression method and device based on more granularity beta prunings |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN107368885A true CN107368885A (en) | 2017-11-21 |
Family
ID=60308043
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710568710.8A Pending CN107368885A (en) | 2017-07-13 | 2017-07-13 | Network model compression method and device based on more granularity beta prunings |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107368885A (en) |
Cited By (20)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108629288A (en) * | 2018-04-09 | 2018-10-09 | 华中科技大学 | A kind of gesture identification model training method, gesture identification method and system |
| CN108898168A (en) * | 2018-06-19 | 2018-11-27 | 清华大学 | The compression method and system of convolutional neural networks model for target detection |
| CN109344921A (en) * | 2019-01-03 | 2019-02-15 | 湖南极点智能科技有限公司 | A kind of image-recognizing method based on deep neural network model, device and equipment |
| CN109523017A (en) * | 2018-11-27 | 2019-03-26 | 广州市百果园信息技术有限公司 | Compression method, device, equipment and the storage medium of deep neural network |
| CN109583586A (en) * | 2018-12-05 | 2019-04-05 | 东软睿驰汽车技术（沈阳）有限公司 | A kind of convolution kernel processing method and processing device |
| CN109634401A (en) * | 2018-12-29 | 2019-04-16 | 联想(北京)有限公司 | A kind of control method and electronic equipment |
| WO2019127362A1 (en) * | 2017-12-29 | 2019-07-04 | 清华大学 | Neural network model block compression method, training method, computing device and system |
| CN110059823A (en) * | 2019-04-28 | 2019-07-26 | 中国科学技术大学 | Deep neural network model compression method and device |
| WO2019226686A3 (en) * | 2018-05-23 | 2020-02-06 | Movidius Ltd. | Deep learning system |
| CN111695375A (en) * | 2019-03-13 | 2020-09-22 | 上海云从企业发展有限公司 | Face recognition model compression algorithm based on model distillation, medium and terminal |
| CN112036563A (en) * | 2019-06-03 | 2020-12-04 | 国际商业机器公司 | Deep learning model insights using provenance data |
| CN112132062A (en) * | 2020-09-25 | 2020-12-25 | 中南大学 | Remote sensing image classification method based on pruning compression neural network |
| CN112329931A (en) * | 2021-01-04 | 2021-02-05 | 北京智源人工智能研究院 | Countermeasure sample generation method and device based on proxy model |
| KR20210015990A (en) * | 2019-05-18 | 2021-02-10 | 주식회사 디퍼아이 | Convolution neural network parameter optimization method, neural network computing method and apparatus |
| CN113408561A (en) * | 2020-03-17 | 2021-09-17 | 北京京东乾石科技有限公司 | Model generation method, target detection method, device, equipment and storage medium |
| TWI768167B (en) * | 2017-12-30 | 2022-06-21 | 大陸商中科寒武紀科技股份有限公司 | Integrated circuit chip device and related products |
| US11651202B2 (en) | 2017-12-30 | 2023-05-16 | Cambricon Technologies Corporation Limited | Integrated circuit chip device and related product |
| US11704544B2 (en) | 2017-12-30 | 2023-07-18 | Cambricon Technologies Corporation Limited | Integrated circuit chip device and related product |
| US11710031B2 (en) | 2017-12-30 | 2023-07-25 | Cambricon Technologies Corporation Limited | Parallel processing circuits for neural networks |
| US11734548B2 (en) | 2017-12-30 | 2023-08-22 | Cambricon Technologies Corporation Limited | Integrated circuit chip device and related product |
-
2017
- 2017-07-13 CN CN201710568710.8A patent/CN107368885A/en active Pending
Cited By (29)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2019127362A1 (en) * | 2017-12-29 | 2019-07-04 | 清华大学 | Neural network model block compression method, training method, computing device and system |
| US11704544B2 (en) | 2017-12-30 | 2023-07-18 | Cambricon Technologies Corporation Limited | Integrated circuit chip device and related product |
| US11710031B2 (en) | 2017-12-30 | 2023-07-25 | Cambricon Technologies Corporation Limited | Parallel processing circuits for neural networks |
| US11734548B2 (en) | 2017-12-30 | 2023-08-22 | Cambricon Technologies Corporation Limited | Integrated circuit chip device and related product |
| TWI768167B (en) * | 2017-12-30 | 2022-06-21 | 大陸商中科寒武紀科技股份有限公司 | Integrated circuit chip device and related products |
| US11651202B2 (en) | 2017-12-30 | 2023-05-16 | Cambricon Technologies Corporation Limited | Integrated circuit chip device and related product |
| CN108629288A (en) * | 2018-04-09 | 2018-10-09 | 华中科技大学 | A kind of gesture identification model training method, gesture identification method and system |
| CN108629288B (en) * | 2018-04-09 | 2020-05-19 | 华中科技大学 | Gesture recognition model training method, gesture recognition method and system |
| WO2019226686A3 (en) * | 2018-05-23 | 2020-02-06 | Movidius Ltd. | Deep learning system |
| US11900256B2 (en) | 2018-05-23 | 2024-02-13 | Intel Corporation | Deep learning system |
| CN108898168A (en) * | 2018-06-19 | 2018-11-27 | 清华大学 | The compression method and system of convolutional neural networks model for target detection |
| CN108898168B (en) * | 2018-06-19 | 2021-06-01 | 清华大学 | Compression method and system of convolutional neural network model for target detection |
| CN109523017B (en) * | 2018-11-27 | 2023-10-17 | 广州市百果园信息技术有限公司 | Gesture detection method, device, equipment and storage medium |
| CN109523017A (en) * | 2018-11-27 | 2019-03-26 | 广州市百果园信息技术有限公司 | Compression method, device, equipment and the storage medium of deep neural network |
| CN109583586A (en) * | 2018-12-05 | 2019-04-05 | 东软睿驰汽车技术（沈阳）有限公司 | A kind of convolution kernel processing method and processing device |
| CN109583586B (en) * | 2018-12-05 | 2021-03-23 | 东软睿驰汽车技术（沈阳）有限公司 | Convolution kernel processing method and device in voice recognition or image recognition |
| CN109634401A (en) * | 2018-12-29 | 2019-04-16 | 联想(北京)有限公司 | A kind of control method and electronic equipment |
| CN109634401B (en) * | 2018-12-29 | 2023-05-02 | 联想(北京)有限公司 | Control method and electronic equipment |
| CN109344921A (en) * | 2019-01-03 | 2019-02-15 | 湖南极点智能科技有限公司 | A kind of image-recognizing method based on deep neural network model, device and equipment |
| CN111695375A (en) * | 2019-03-13 | 2020-09-22 | 上海云从企业发展有限公司 | Face recognition model compression algorithm based on model distillation, medium and terminal |
| CN110059823A (en) * | 2019-04-28 | 2019-07-26 | 中国科学技术大学 | Deep neural network model compression method and device |
| KR102247896B1 (en) | 2019-05-18 | 2021-05-04 | 주식회사 디퍼아이 | Convolution neural network parameter optimization method, neural network computing method and apparatus |
| KR20210015990A (en) * | 2019-05-18 | 2021-02-10 | 주식회사 디퍼아이 | Convolution neural network parameter optimization method, neural network computing method and apparatus |
| CN112036563A (en) * | 2019-06-03 | 2020-12-04 | 国际商业机器公司 | Deep learning model insights using provenance data |
| WO2021185121A1 (en) * | 2020-03-17 | 2021-09-23 | 北京京东乾石科技有限公司 | Model generation method and apparatus, object detection method and apparatus, device, and storage medium |
| CN113408561A (en) * | 2020-03-17 | 2021-09-17 | 北京京东乾石科技有限公司 | Model generation method, target detection method, device, equipment and storage medium |
| CN112132062A (en) * | 2020-09-25 | 2020-12-25 | 中南大学 | Remote sensing image classification method based on pruning compression neural network |
| CN112329931B (en) * | 2021-01-04 | 2021-05-07 | 北京智源人工智能研究院 | Countermeasure sample generation method and device based on proxy model |
| CN112329931A (en) * | 2021-01-04 | 2021-02-05 | 北京智源人工智能研究院 | Countermeasure sample generation method and device based on proxy model |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107368885A (en) | Network model compression method and device based on more granularity beta prunings | |
| CN105512289B (en) | Image search method based on deep learning and Hash | |
| CN103810101B (en) | A kind of Software Defects Predict Methods and software defect forecasting system | |
| CN109146076A (en) | model generating method and device, data processing method and device | |
| CN107273836A (en) | A kind of pedestrian detection recognition methods, device, model and medium | |
| CN109754068A (en) | Transfer learning method and terminal device based on deep learning pre-training model | |
| CN110263863A (en) | Fine granularity mushroom phenotype recognition methods based on transfer learning Yu bilinearity InceptionResNetV2 | |
| CN108614997B (en) | Remote sensing image identification method based on improved AlexNet | |
| CN108388957B (en) | Medium and small river flood forecasting method and forecasting system based on multi-feature fusion technology | |
| CN110322453A (en) | 3D point cloud semantic segmentation method based on position attention and auxiliary network | |
| WO2020238237A1 (en) | Power exponent quantization-based neural network compression method | |
| CN105354595A (en) | Robust visual image classification method and system | |
| CN104616029B (en) | Data classification method and device | |
| CN103313047B (en) | A kind of method for video coding and device | |
| CN109993100A (en) | The implementation method of facial expression recognition based on further feature cluster | |
| CN106780737A (en) | A kind of method that utilization digital elevation model calculates Geomorphologic Instantaneous Unit Hydrograph probability | |
| CN111435463A (en) | Data processing method and related equipment and system | |
| CN110060489A (en) | A kind of traffic signal timing scheme recommended method neural network based | |
| CN109284761A (en) | A kind of image characteristic extracting method, device, equipment and readable storage medium storing program for executing | |
| CN104318241A (en) | Local density spectral clustering similarity measurement algorithm based on Self-tuning | |
| CN105631469A (en) | Bird image recognition method by multilayer sparse coding features | |
| CN109472352A (en) | A kind of deep neural network model method of cutting out based on characteristic pattern statistical nature | |
| CN110070106A (en) | Smog detection method, device and electronic equipment | |
| Cheni et al. | Reducing latency in a converted spiking video segmentation network | |
| CN103281555B (en) | Half reference assessment-based quality of experience (QoE) objective assessment method for video streaming service |
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 |
Application publication date: 20171121 |
|
| RJ01 | Rejection of invention patent application after publication |