CN110399917A - A kind of image classification method based on hyperparameter optimization CNN - Google Patents
A kind of image classification method based on hyperparameter optimization CNN Download PDFInfo
- Publication number
- CN110399917A CN110399917A CN201910671268.0A CN201910671268A CN110399917A CN 110399917 A CN110399917 A CN 110399917A CN 201910671268 A CN201910671268 A CN 201910671268A CN 110399917 A CN110399917 A CN 110399917A
- Authority
- CN
- China
- Prior art keywords
- cnn
- particle
- hyper parameter
- image classification
- value
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
- G06F18/24147—Distances to closest patterns, e.g. nearest neighbour classification
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
- G06N20/20—Ensemble learning
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T10/00—Road transport of goods or passengers
- Y02T10/10—Internal combustion engine [ICE] based vehicles
- Y02T10/40—Engine management systems
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Software Systems (AREA)
- Medical Informatics (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Image Analysis (AREA)
Abstract
The present invention provides a kind of image classification method based on hyperparameter optimization CNN, belongs to image identification technical field.This method is hyper parameter of the invention according to the structural parameters that the design feature of CNN framework chooses convolutional layer C1 and pond layer P1, and limits the value range of hyper parameter as (Xl,Xu).Then it using the hyper parameter of two kinds of period variation PSO algorithm optimization CNN of global variation and local variations, avoids traditional PS O algorithm and local optimum is rested on to the optimization of hyper parameter, to obtain the image classification performance more more competitive than traditional PS O algorithm.It is obviously improved the efficiency and cost of deep learning CNN hyperparameter optimization, has farthest played the image classification potentiality of CNN framework, saved hardware resource when CNN carries out image classification and calculates cost, there is certain application value in engineering practice.
Description
Technical field
The present invention relates to image identification technical field more particularly to a kind of image classification sides based on hyperparameter optimization CNN
Method.
Background technique
Image Classfication Technology has developed comparative maturity, and the CNN framework suitable for different scenes classification emerges one after another, but multiple
Miscellaneous CNN structure often expends hardware resource and calculates cost.Before CNN is for image classification training, it need to set in CNN in advance
These parameters are known as hyper parameter by some parameters in portion, and choosing one group of optimal hyper parameter can be in the premise for not changing CNN structure
Under utmostly promote CNN image classification performance.Therefore, suitable hyper parameter is selected the image classification performance of CNN framework is complete
It releases entirely particularly important in engineering practice.
The hyperparameter optimization technique study of image classification has had some achievements, and the research of early stage is concentrated on machine learning
Hyperparameter optimization method be used for CNN.Hyperparameter optimization method is broadly divided into model-free optimization and the optimization based on model, the former
State-of-the-art method includes simple grid and random search, and the latter includes the heuristic value based on population, Yi Jiji
Optimize (GP) in the Bayes of Gaussian process.Heuristic value especially merits attention for CNN hyperparameter optimization, wherein grain
Swarm optimization since its simplicity and versatility have been proved to highly effective in terms of solving the multiple tasks in many regions,
And it has the great potential of large-scale parallel.Hyperparameter optimization based on particle swarm algorithm, search efficiency are much super
Other hyperparameter optimization algorithms such as grid search, random search are crossed, the search time of hyperparameter optimization is accelerated, solves tradition
The problems such as optimization efficiency of hyperparameter optimization is low, time-consuming.But particle swarm algorithm there are problems that being easily trapped into local optimum, this meeting
Hyperparameter optimization is caused to rest on local optimum, rather than the one of global optimum group of hyper parameter, this makes not searching to a certain extent
Rope is to one group of hyper parameter for being optimal CNN performance, so that CNN image classification can not be made to reach optimal result.
Summary of the invention
The present invention in view of the above shortcomings of the prior art, provides a kind of image classification method based on hyperparameter optimization CNN.
The technical solution used in the present invention is:
A kind of image classification method based on hyperparameter optimization CNN, includes the following steps:
Step 1: the image data set classified to needs pre-processes, and it is divided into training set T, test in proportion
Collect C and verifying collection V, wherein verifying collection V is extracted from training set T, meet | T |=9 | V |,
Step 2: CNN framework, including convolutional layer C1, pond layer P1, convolutional layer C2, pond layer P2 composition are built, by
Softmax activation terminates.It is this according to the structural parameters that the design feature of CNN framework chooses convolutional layer C1 and pond layer P1
The hyper parameter of invention determines that the value range of this group of hyper parameter is (Xl,Xu)。
Step 3: using the hyper parameter of period variation PSO algorithm optimization CNN framework on verifying collection, obtaining one group of hyper parameter
Value Xi(g)；
Step 3-1: initialization particle rapidity, fitness function value, individual optimum position Pg, overall situation optimum position Pi, setting
The number of iterations g is 0, iteration precision δ >=0, and particle search Spatial Dimension is D, particle number N；
The current location vector of each particle in group is Xi=(xi,1,xi,2,...,xi,D), i=1,2 ..., N, when
Preceding velocity vector is Vi=(vi,1,vi,2,...,vi,D), i=1,2 ..., N, individual optimum position vector are Pg=(pG, 1+pg,2
+...+pg,D), i=1,2 ..., N, global optimum position vector are Pi=(pi1+pi2+...+piD), i=1,2 ..., N；
Step 3-2: in the g times iteration, each particle updates speed and the position of oneself:
Vi(g+1)=ω Vi(g)+c1r1(Pi(g)-Xi(g))+c2r2(Pg(g)-Xi(g)) (1)
Xi(g+1)=Xi(g)+Vi(g) (2)
Wherein, ω is inertial factor, c1And c2It is constant, r for Studying factors1And r2It is in (0,1) range
Random number；
Step 3-3: global mutation operator is selected to change the position of all particles in entire group, or selection
Local variations operator changes the position of the elite particle in group:
Global mutation operator and the formula that local variations operator changes position are as follows:
Wherein, A1, A2It is customized amplitude factor, is constant, r3, r4It is 0,1) random number in range,For
Elite particle, q are the number of the new particle generated by local variations operator, f1For global variation frequency, f2For local variations frequency
Rate；
Step 3-4: checking whether particle rapidity and position cross the border, and is replaced accordingly if crossing the border with the boundary value that it exceeds
Particle value, specific judgment method are as follows:
If Vi(g)≤Vl, then Vi(g)=V1；If Vi(g)≥Vu, then Vi(g)=V1；If Xi(g)≤X1, then Xi(g)=X1；
If Xi(g)≥Xu, then Xi(g)=X1；
Wherein, (V1,Vu) be particle velocity interval, (X1,Xu) be particle position range；
Step 3-5: the i.e. required hyper parameter value X of the optimal location that execution step 3-1 to step 3-4 is obtainedi(g)。
Step 4: by hyper parameter Xi(g) be input in CNN, and the training set obtained in step 1 to the CNN after optimization into
Row training；
Step 5: the test set that step 1 is obtained inputs in trained CNN, obtains the classification results of test set C；
Step 6: judging whether iteration reaches termination condition；
Step 6-1: the fitness function value of each particle period variation PSO is calculated:
Wherein, CNN (Xi(g)) accuracy rate of the classification results obtained for step 5 in the claim 1, Xi(g) for institute
State the hyper parameter value that step 3 in claim 1 obtains；
Step 6-2: it by comparing each particle fitness function value of the obtained current iteration of step 6-1, updates respectively
Individual history optimal location Pi(g) and group's optimal location Pg(g), current iteration optimal particle X is obtainedmin(g):
Step 6-3: judge that the fitness value of optimal particle increases the threshold value for being less than and being indicated by ε, judge best in group
Particle position update be less than byThe minimum step of expression, judges whether the number of iterations g reaches maximum number of iterations gmaxIf full
One of above-mentioned termination condition of foot, then terminate iteration.
Step 7: if not reaching termination condition, executing step 3 to step 6 and continue iteration；
Step 8: if reaching termination condition, obtaining final optimal hyper parameter, be denoted as Xmin(g)；
Step 9: by final optimal hyper parameter Xmin(g) it substitutes into CNN, classifies to the image of entire data set, obtain
To classification results.
The beneficial effects of adopting the technical scheme are that provided by the invention a kind of based on hyperparameter optimization
The image classification method of CNN effectively improves the problem of particle swarm algorithm is easily trapped into local optimum, and then improves particle
Group's convergence speed of the algorithm and convergence precision；Since the search performance of hyperparameter optimization method is better, keep away to a certain extent
The problem for having exempted to select improper caused CNN nicety of grading undesirable due to hyper parameter, improves image classification accuracy, thus most
The performance of CNN processing image is played to great Cheng.
Detailed description of the invention
Fig. 1 is a kind of image classification method flow chart based on hyperparameter optimization CNN of the present invention；
Fig. 2 is the segment in first embodiment of the invention for the Handwritten Digit Recognition MNIST data set of image classification；
Fig. 3 is that hyperparameter optimization method improves front and back MNIST data set classification performance with repeatedly in first embodiment of the invention
Generation number variation diagram；
Fig. 4 is the segment in second embodiment of the invention for the object identification cifar-10 data set of image classification；
Fig. 5 be second embodiment of the invention in hyperparameter optimization method improve front and back cifar-10 data set classification performance with
The number of iterations variation diagram.
Specific embodiment
With reference to the accompanying drawings and examples, specific embodiments of the present invention will be described in further detail.Implement below
Example is not intended to limit the scope of the invention for illustrating the present invention.
Embodiment 1
As shown in Figure 1, the method for the present embodiment is as described below.
Step 1: selection benchmark dataset Handwritten Digit Recognition MNIST data set is data set to be sorted, the data set
Segment as shown in Fig. 2, the data set has 70000 gray level images, each image is 28 × 28 pixels, wherein including 10 classes
Not, each classification has 7000 images.60000 images in the data set are randomly selected as training set, remaining 10000
A image picks out 6000 images as verifying collection as test set at random from training set.
Step 2: CNN framework, including convolutional layer C1, pond layer P1, convolutional layer C2, pond layer P2 composition are built, by
Softmax activation terminates.It is this according to the structural parameters that the design feature of CNN framework chooses convolutional layer C1 and pond layer P1
The hyper parameter of invention determines that the value range of this group of hyper parameter is (Xl,Xu), as shown in table 1.
The range of the parameter and its permission of 1 convolutional layer of table and maximum pond layer
Step 3: using the hyper parameter of period variation PSO algorithm optimization CNN framework on verifying collection, obtaining one group of hyper parameter
Value Xi(g)；
Step 3-1: initialization particle rapidity, fitness function value, individual optimum position Pg, overall situation optimum position Pi, setting
The number of iterations g is 0, iteration precision δ >=0, particle search Spatial Dimension D=4, particle number N=10；
The current location vector of each particle in group is Xi=(xi,1,xi,2,...,xi,4), i=1,2 ..., 10,
Current velocity vector is Vi=(vi,1,vi,2,...,vi,4), i=1,2 ..., 10, individual optimum position vector is Pg=(pG, 1+
pg,2+...+pg,4), i=1,2 ..., 10, global optimum position vector is Pi=(pi1+pi2+...+pi4), i=1,2 ...,
10；
Step 3-2: in the g times iteration, each particle updates speed and the position of oneself:
Vi(g+1)=ω Vi(g)+c1r1(Pi(g)-Xi(g))+c2r2(Pg(g)-Xi(g)) (1)
Xi(g+1)=Xi(g)+Vi(g) (2)
Wherein, ω=c1=c2=0.5, r1And r2It is the random number in (0,1) range；
Step 3-3: global mutation operator is selected to change the position of all particles in entire group, or selection
Local variations operator changes the position of the elite particle in group:
Global mutation operator and the formula that local variations operator changes position are as follows:
Wherein, A1=A2=1, r3、r4For the random number in (0,1) range,For elite particle, calculated by local variations
The number q=3, global variation frequency f for the new particle that son generates1=10, local variations frequency f2=2；
Step 3-4: checking whether particle rapidity and position cross the border, and is replaced accordingly if crossing the border with the boundary value that it exceeds
Particle value, specific judgment method are as follows:
If Vi(g)≤Vl, then Vi(g)=V1；If Vi(g)≥Vu, then Vi(g)=V1；If Xi(g)≤X1, then Xi(g)=X1；
If Xi(g)≥Xu, then Xi(g)=X1；
Wherein, (V1,Vu) be particle velocity interval value be (- 2,2), (X1,Xu) be particle position range numerical value such as
Described in table 1；
Step 3-5: the i.e. required hyper parameter value X of the optimal location that execution step 3-1 to step 3-4 is obtainedi(g)。
Step 4: by hyper parameter Xi(g) be input in CNN, and the training set obtained in step 1 to the CNN after optimization into
Row training；
Step 5: the test set that step 1 is obtained inputs in trained CNN, obtains the classification results of image；
Step 6: judging whether iteration reaches termination condition；
Step 6-1: the fitness function value of each particle period variation PSO is calculated, fitness function curve is drawn, such as schemes
Shown in 3:
Wherein, CNN (Xi(g)) accuracy rate of the classification results obtained for step 5 in the claim 1, Xi(g) for institute
State the hyper parameter value that step 3 in claim 1 obtains；
Step 6-2: it by comparing each particle fitness function value of the obtained current iteration of step 6-1, updates respectively
Individual history optimal location Pi(g) and group's optimal location Pg(g), current iteration optimal particle X is obtainedmin(g):
Step 6-3: judge that the increase of optimal particle fitness value five generations successively is less than threshold epsilon=0.0001, judge group
In best particle position continuous 5 generation update be less than minimum stepJudge whether the number of iterations g reaches maximum
The number of iterations 20 terminates iteration if meeting one of above-mentioned termination condition.
Step 7: if not reaching termination condition, executing step 3 to step 6；
Step 8: if reaching termination condition, obtaining final optimal hyper parameter, be denoted as Xmin(g)；
Step 9: by final optimal hyper parameter Xmin(g) it substitutes into CNN, classifies to the image of entire data set, obtain
To classification results.
Embodiment 2
As shown in Figure 1, the method for the present embodiment is as described below.
Step 1: selection benchmark dataset object identification cifar-10 data set is data set to be sorted, the data set
Segment is as shown in figure 4, the data set has 60000 32 × 32 pixel color images, wherein including 10 classifications, each classification has
6000 images.50000 images in the data set are randomly selected as training set, remaining 10000 images are as survey
Examination collection picks out 5000 images as verifying collection at random from training set.
Step 2, with embodiment 1, obtains hyperparameter optimization method and improves front and back cifar-10 data set classification to step 9
It can change with the number of iterations as shown in Figure 5.
Hyperparameter optimization method improves front and back Handwritten Digit Recognition MNIST data set and object identification cifar-10 data set
The comparison of image classification accuracy rate is as shown in table 2:
For the accuracy rate of different data collection CNN image classification before and after 2 hyperparameter optimization of table
The result shows that the method for the present invention has centainly image classification accuracy rate under the premise of not changing classification CNN framework
Degree is promoted, and has farthest played the image classification potentiality of CNN framework, has saved hardware when CNN carries out image classification
Resource and calculating cost, there is certain application value in engineering practice.
Claims (5)
1. a kind of image classification method based on hyperparameter optimization CNN, it is characterised in that include the following steps:
Step 1: the image data set classified to needs pre-processes, and it is divided into training set T, test set C in proportion
Collect V with verifying；
Step 2: building CNN framework, hyper parameter and its value range are chosen according to the design feature of CNN framework；
Step 3: using the hyper parameter of period variation PSO algorithm optimization CNN framework on verifying collection, obtaining one group of hyper parameter value Xi
(g)；
Step 4: by hyper parameter Xi(g) it is input in CNN, and the training set obtained in step 1 instructs the CNN after optimization
Practice；
Step 5: the test set C that step 1 is obtained is inputted in trained CNN, obtains the classification results of test set C；
Step 6: judging whether iteration reaches termination condition；
Step 7: if not reaching termination condition, executing step 3 to step 6 and continue iteration；
Step 8: if reaching termination condition, obtaining final optimal hyper parameter, be denoted as Xmin(g)；
Step 9: by final optimal hyper parameter Xmin(g) it substitutes into CNN, classifies to the image of entire data set, divided
Class result.
2. a kind of image classification method based on hyperparameter optimization CNN according to claim 1, it is characterised in that the step
Verifying collection V is extracted from training set T in rapid 1, is met | T |=9 | and V |,
3. a kind of image classification method based on hyperparameter optimization CNN according to claim 1, it is characterised in that the step
The CNN framework built in rapid 2 includes convolutional layer C1, pond layer P1, convolutional layer C2, pond layer P2 composition, is activated eventually by Softmax
Only, it is hyper parameter of the invention according to the structural parameters that the design feature of CNN framework chooses convolutional layer C1 and pond layer P1, determines
The value range of this group of hyper parameter is (Xl,Xu)。
4. a kind of image classification method based on hyperparameter optimization CNN according to claim 1, it is characterised in that the step
Process in rapid 3 on verifying collection using the hyper parameter of period variation PSO algorithm optimization CNN framework is as follows:
Step 3-1: initialization particle rapidity, fitness function value, individual optimum position Pg, overall situation optimum position Pi, iteration is set
Number g is 0, iteration precision δ >=0, and particle search Spatial Dimension is D, particle number N；
The current location vector of each particle in group is Xi=(xi,1,xi,2,...,xi,D), i=1,2 ..., N, current speed
Degree vector is Vi=(vi,1,vi,2,...,vi,D), i=1,2 ..., N, individual optimum position vector are Pg=(pG, 1+pg,2+...+
pg,D), i=1,2 ..., N, global optimum position vector are Pi=(pi1+pi2+...+piD), i=1,2 ..., N；
Step 3-2: in the g times iteration, each particle updates speed and the position of oneself:
Vi(g+1)=ω Vi(g)+c1r1(Pi(g)-Xi(g))+c2r2(Pg(g)-Xi(g)) (1)
Xi(g+1)=Xi(g)+Vi(g) (2)
Wherein, ω is inertial factor, c1And c2It is constant, r for Studying factors1And r2It is random in (0,1) range
Number；
Step 3-3: selecting global mutation operator to change the position of all particles in entire group, or selection part
Mutation operator changes the position of the elite particle in group:
Global mutation operator and the formula that local variations operator changes position are as follows:
Xi(g)=Xi(g)[1+A1(0.5-r3)δ]
I=1,2 ..., N (3)
Wherein, A1, A2It is customized amplitude factor, is constant, r3, r4For the random number in (0,1) range,For elite
Particle, q are the number of the new particle generated by local variations operator, f1For global variation frequency, f2For local variations frequency；
Step 3-4: checking whether particle rapidity and position cross the border, and replaces corresponding particle if crossing the border with the boundary value that it exceeds
Value, specific judgment method are as follows:
If Vi(g)≤Vl, then Vi(g)=V1；If Vi(g)≥Vu, then Vi(g)=V1；If Xi(g)≤X1, then Xi(g)=X1；If Xi
(g)≥Xu, then Xi(g)=X1；
Wherein, (V1,Vu) be particle velocity interval, (X1,Xu) be particle position range；
Step 3-5: the i.e. required hyper parameter value X of the optimal location that execution step 3-1 to step 3-4 is obtainedi(g)。
5. a kind of image classification method based on hyperparameter optimization CNN according to claim 1, it is characterised in that the step
Judge in rapid 6 iteration whether reach termination condition process it is as follows:
Step 6-1: the fitness function value of each particle period variation PSO is calculated:
Wherein, CNN (Xi(g)) accuracy rate of the classification results obtained for step 5 in the claim 1, XiIt (g) is the power
Benefit requires the hyper parameter value that step 3 obtains in 1；
Step 6-2: by comparing each particle fitness function value of the obtained current iteration of step 6-1, difference more new individual
History optimal location Pi(g) and group's optimal location Pg(g), current iteration optimal particle X is obtainedmin(g):
Step 6-3: judge that the fitness value of optimal particle increases the threshold value for being less than and being indicated by ε, judge the best particle in group
Location updating be less than byThe minimum step of expression, judges whether the number of iterations g reaches maximum number of iterations gmaxIf on meeting
One of termination condition is stated, then terminates iteration.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910671268.0A CN110399917B (en) | 2019-07-24 | 2019-07-24 | Image classification method based on hyper-parameter optimization CNN |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910671268.0A CN110399917B (en) | 2019-07-24 | 2019-07-24 | Image classification method based on hyper-parameter optimization CNN |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110399917A true CN110399917A (en) | 2019-11-01 |
| CN110399917B CN110399917B (en) | 2023-04-18 |
Family
ID=68324921
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910671268.0A Active CN110399917B (en) | 2019-07-24 | 2019-07-24 | Image classification method based on hyper-parameter optimization CNN |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110399917B (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110942090A (en) * | 2019-11-11 | 2020-03-31 | 北京迈格威科技有限公司 | Model training method, image processing method, device, electronic equipment and storage medium |
| CN111160459A (en) * | 2019-12-30 | 2020-05-15 | 上海依图网络科技有限公司 | Device and method for optimizing hyper-parameters |
| CN112197876A (en) * | 2020-09-27 | 2021-01-08 | 中国科学院光电技术研究所 | Single far-field type depth learning wavefront restoration method based on four-quadrant discrete phase modulation |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108334949A (en) * | 2018-02-11 | 2018-07-27 | 浙江工业大学 | A kind of tachytelic evolution method of optimization depth convolutional neural networks structure |
| CN109085469A (en) * | 2018-07-31 | 2018-12-25 | 中国电力科学研究院有限公司 | A kind of method and system of the signal type of the signal of cable local discharge for identification |
| US20190180188A1 (en) * | 2017-12-13 | 2019-06-13 | Cognizant Technology Solutions U.S. Corporation | Evolution of Architectures For Multitask Neural Networks |
| CN109919202A (en) * | 2019-02-18 | 2019-06-21 | 新华三技术有限公司合肥分公司 | Disaggregated model training method and device |
-
2019
- 2019-07-24 CN CN201910671268.0A patent/CN110399917B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20190180188A1 (en) * | 2017-12-13 | 2019-06-13 | Cognizant Technology Solutions U.S. Corporation | Evolution of Architectures For Multitask Neural Networks |
| CN108334949A (en) * | 2018-02-11 | 2018-07-27 | 浙江工业大学 | A kind of tachytelic evolution method of optimization depth convolutional neural networks structure |
| CN109085469A (en) * | 2018-07-31 | 2018-12-25 | 中国电力科学研究院有限公司 | A kind of method and system of the signal type of the signal of cable local discharge for identification |
| CN109919202A (en) * | 2019-02-18 | 2019-06-21 | 新华三技术有限公司合肥分公司 | Disaggregated model training method and device |
Non-Patent Citations (1)
| Title |
|---|
| 张进等: "改进的基于粒子群优化的支持向量机特征选择和参数联合优化算法", 《计算机应用》 * |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110942090A (en) * | 2019-11-11 | 2020-03-31 | 北京迈格威科技有限公司 | Model training method, image processing method, device, electronic equipment and storage medium |
| CN110942090B (en) * | 2019-11-11 | 2024-03-29 | 北京迈格威科技有限公司 | Model training method, image processing device, electronic equipment and storage medium |
| CN111160459A (en) * | 2019-12-30 | 2020-05-15 | 上海依图网络科技有限公司 | Device and method for optimizing hyper-parameters |
| CN112197876A (en) * | 2020-09-27 | 2021-01-08 | 中国科学院光电技术研究所 | Single far-field type depth learning wavefront restoration method based on four-quadrant discrete phase modulation |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110399917B (en) | 2023-04-18 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Ren et al. | A scatter learning particle swarm optimization algorithm for multimodal problems | |
| Wang et al. | Stud krill herd algorithm | |
| CN110399917A (en) | A kind of image classification method based on hyperparameter optimization CNN | |
| Jovanovic et al. | Ant colony optimization algorithm with pheromone correction strategy for the minimum connected dominating set problem | |
| Luo et al. | A clonal selection algorithm for dynamic multimodal function optimization | |
| CN110363344A (en) | Probability integral parameter prediction method based on MIV-GP algorithm optimization BP neural network | |
| CN106502092B (en) | A kind of thermal process model parameter identification method using improvement Hybrid Particle Swarm | |
| Hoseini et al. | Efficient contrast enhancement of images using hybrid ant colony optimisation, genetic algorithm, and simulated annealing | |
| CN107392919B (en) | Adaptive genetic algorithm-based gray threshold acquisition method and image segmentation method | |
| Lan et al. | A two-phase learning-based swarm optimizer for large-scale optimization | |
| CN107506865B (en) | Load prediction method and system based on LSSVM optimization | |
| CN108399450A (en) | Improvement particle cluster algorithm based on biological evolution principle | |
| CN105868775A (en) | Imbalance sample classification method based on PSO (Particle Swarm Optimization) algorithm | |
| CN105844628B (en) | It is a kind of based on the table ore zoning map of krill optimization algorithm as split plot design | |
| CN108564592A (en) | Based on a variety of image partition methods for being clustered to differential evolution algorithm of dynamic | |
| Li et al. | Dynamic community detection algorithm based on incremental identification | |
| CN113011076A (en) | Efficient particle swarm optimization method based on RBF proxy model | |
| CN110097176A (en) | A kind of neural network structure searching method applied to air quality big data abnormality detection | |
| CN107292381A (en) | A kind of method that mixed biologic symbiosis for single object optimization is searched for | |
| CN110287985A (en) | A kind of deep neural network image-recognizing method based on the primary topology with Mutation Particle Swarm Optimizer | |
| CN112733458A (en) | Engineering structure signal processing method based on self-adaptive variational modal decomposition | |
| CN108985323A (en) | A kind of short term prediction method of photovoltaic power | |
| CN106485030B (en) | A kind of symmetrical border processing method for SPH algorithm | |
| CN109034479B (en) | Multi-target scheduling method and device based on differential evolution algorithm | |
| CN105205534B (en) | A kind of three value FPRM circuit areas and power consumption optimum polarity search method |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |