CN109902824A - It is a kind of to generate confrontation network method with self adaptive control learning improvement - Google Patents
It is a kind of to generate confrontation network method with self adaptive control learning improvement Download PDFInfo
- Publication number
- CN109902824A CN109902824A CN201910157777.1A CN201910157777A CN109902824A CN 109902824 A CN109902824 A CN 109902824A CN 201910157777 A CN201910157777 A CN 201910157777A CN 109902824 A CN109902824 A CN 109902824A
- Authority
- CN
- China
- Prior art keywords
- training
- gan
- value
- adaptive control
- confrontation network
- 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
- 238000000034 method Methods 0.000 title claims abstract description 68
- 230000003044 adaptive effect Effects 0.000 title claims abstract description 38
- 230000006872 improvement Effects 0.000 title claims abstract description 24
- 238000012549 training Methods 0.000 claims abstract description 74
- 230000008569 process Effects 0.000 claims abstract description 21
- 238000013480 data collection Methods 0.000 claims abstract description 12
- 238000013256 Gubra-Amylin NASH model Methods 0.000 claims description 10
- 230000008901 benefit Effects 0.000 claims description 3
- 230000008859 change Effects 0.000 claims description 3
- 238000004590 computer program Methods 0.000 claims description 3
- 238000011017 operating method Methods 0.000 claims description 2
- 230000004069 differentiation Effects 0.000 claims 1
- 201000004409 schistosomiasis Diseases 0.000 claims 1
- 238000013527 convolutional neural network Methods 0.000 abstract description 5
- 230000000694 effects Effects 0.000 abstract description 4
- 238000005259 measurement Methods 0.000 abstract description 4
- 238000004458 analytical method Methods 0.000 abstract description 2
- 230000007246 mechanism Effects 0.000 abstract description 2
- 238000009826 distribution Methods 0.000 description 11
- 230000006870 function Effects 0.000 description 9
- 238000004422 calculation algorithm Methods 0.000 description 7
- 238000005516 engineering process Methods 0.000 description 5
- 230000000007 visual effect Effects 0.000 description 5
- 238000013473 artificial intelligence Methods 0.000 description 3
- 230000015572 biosynthetic process Effects 0.000 description 3
- 238000010586 diagram Methods 0.000 description 3
- 238000003786 synthesis reaction Methods 0.000 description 3
- 238000012360 testing method Methods 0.000 description 3
- 238000004364 calculation method Methods 0.000 description 2
- 230000003247 decreasing effect Effects 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 241000208340 Araliaceae Species 0.000 description 1
- 235000005035 Panax pseudoginseng ssp. pseudoginseng Nutrition 0.000 description 1
- 235000003140 Panax quinquefolius Nutrition 0.000 description 1
- 230000006978 adaptation Effects 0.000 description 1
- 230000003042 antagnostic effect Effects 0.000 description 1
- 238000013528 artificial neural network Methods 0.000 description 1
- 230000006399 behavior Effects 0.000 description 1
- 230000000052 comparative effect Effects 0.000 description 1
- 239000012141 concentrate Substances 0.000 description 1
- 238000007796 conventional method Methods 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 230000001815 facial effect Effects 0.000 description 1
- 235000008434 ginseng Nutrition 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000035772 mutation Effects 0.000 description 1
- 238000003058 natural language processing Methods 0.000 description 1
- 238000003062 neural network model Methods 0.000 description 1
- 230000008447 perception Effects 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
Abstract
The invention belongs to fight network technique field, it discloses a kind of generated with self adaptive control learning improvement and fights network method (GAN), adaptive hyper parameter learning process suitable for GAN, to improve the training stability of different data collection, to guarantee the quality of generation data (such as image, text).This is that classification and the training process of the relative complex data set of mode is instructed to realize by the well-trained learning curve obtained under classification and the relatively simple data set of mode；The present invention, which is also analyzed, fights network (Ak-GAN) model with the adaptive generation of multilayer perceptron (MLP) and depth convolutional neural networks (DCGAN) framework.The stability of general GAN training can be improved in the present invention really, and can be generalized to various improved models and data set well；For following work, planning analysis preferably generates sample measurement, this may encourage the convergence of GAN；Wish to analyze effect of the self-adaptive controlled making mechanism proposed in GAN Multimodal Learning.
Description
Technical field
The invention belongs to fight network technique field more particularly to a kind of generated with self adaptive control learning improvement to fight
Network method.
Background technique
The sample of various applications can effectively be synthesized by generating confrontation network (GAN), such as image generation, industrial design, voice
Synthesis and natural language processing.The target of GAN is alternately trained Maker model G and arbiter model D.The G and D of GAN is logical
Often selection uses multi-layer perception (MLP) (MLP) or convolutional neural networks (CNN).It is true to simulate that generator receives one group of noise priori
Real data distribution, the distribution simulated is called generation data distribution, and arbiter is trained to extract the area of truthful data
Other feature.More specifically, arbiter judges that a data are the probability from truthful data distribution；, output probability and really mark
The error of label is for instructing the parameter of arbiter and generator network to update.The process alternately, until D cannot distinguish between very
Reality and generated data.
In practice, GAN is well-known with its nonsteady behavior (unstability) and mode crash issue, especially right
When data acquisition system comprising various visual objects is trained.Two main problems to complicate the issue are: 1) training
Generator and arbiter model capability is unbalanced in journey, this hinders generator effectively learn truthful data to be distributed；2) lacking can
It calculates, interpretable convergence.
Recently, some domestic and international researchs or invention concentrate on how improving the stability of original GAN, the plan that they take
Heuristic is slightly mainly used, is distributed by simplifying training truthful data early period, gradually instructs the training of G, such as scheming
As generating field, GAN is allowed first to try to train perhaps lesser image then sharpening or the expansion gradually of unsharp image
The size of image.This heuristic learning methods can stable training to a certain extent.But it is this by obscure become gradually it is clear or
It does not depend on the equilibrium growth (competition) of G and D ability by the small process to become larger gradually, and is to rely on one and pre-defines
Asymptotic Functions.This method is extremely sensitive to different data sets and GAN model variant, it is difficult to be applied to true application scenarios.
Different from existing research or invention, the invention proposes one kind based on to data collection adaptive, dynamically adjusts hyper parameter
Come guide the method for generator and arbiter study course solve GAN training instability problem.
GAN is easy to train successfully on simple data collection, i.e. the growth or reduction that the ability of G and D can be balanced, final G reach
The purpose for generating photorealism has been arrived, and D finally cannot accurately distinguish true picture and generate the difference of image.This project mentions
The method that ginseng is adaptively adjusted based on a reference value out is established in such a observation: on simple data collection the normal GAN of training it
Two probability output value (P of middle DrAnd Pg) present with the similar trend of Fig. 6 (a) and trend, while the loss function of G and D is defeated
It is worth (L outGAnd LD) present and the similar trend of Fig. 6 (b) and trend.And Fig. 6 (c) illustrates the GAN training trend an of failure
Figure.
Based on this observation, this project is intended that the two groups of curves trained and obtained on MNIST with GAN as reference line:
Two probability output value (P of one group of displaying DmrAnd Pmg)；Loss function value (the L of another group of displaying G and DmGAnd LmD).This two groups
A reference value curve negotiating combination control algolithm instructs training process of the GAN on complicated image.
In conclusion problem of the existing technology is as follows:
Traditional GAN is unstable and mode is collapsed, and especially carries out to the data acquisition system comprising Various Complex visual object
When training；
In the prior art, two main problems to complicate the issue are: generator and arbiter in the training process
Model capability unbalanced growth, which prevent effective study of generator；Lack the convergence that can be calculated, can be explained.
Practical significance of the invention:
Adaptive GAN training method can expand the application field of GAN and promote the process of GAN application landing.Specific table
Present the following aspects: 1) promotion of model performance, model can preferably learn to truthful data to be distributed, when content creation
It is more life-like；2) adaptive, interpretable artificial intelligence technology, the deep neural network of today are the artificial intelligence skill of representative
Art lacks versatility and the adaptivity to data and environment, the adaptive GAN training to different data collection of this project concern
Method can promote the adaptation to data in a certain range and provide theory support for general artificial intelligence technology.
Summary of the invention
In view of the problems of the existing technology, confrontation is generated with self adaptive control learning improvement the present invention provides a kind of
Network method.
The invention is realized in this way a kind of generate confrontation network method with self adaptive control learning improvement, it is described with certainly
Suitable solution learning improvement generate confrontation network method the following steps are included:
Step 1: the training on relatively simple data set using any GAN model, until convergence.When record convergence
Batch number (such as 500 times)
Step 2: the penalty values or decision probability value of first number to convergence all generations of batch number interval are exported, and
These values are fitted by curve-fitting method, the value after fitting is as benchmark penalty values (LmGAnd LmD) or baseline probability
It is worth (PmrAnd Pmg)
Step 3: the training GAN on relative complex data set.In training process every c batch number (iteration coefficient) it
Relatively and calculate current output valve (L afterwardsG、LD、PrAnd PgOne of four) with the difference of a reference value, if the difference is more than pre-
The threshold value (α) being first arranged, then training G and D after being optionally adjusted to k value；If the difference is in threshold value, not to k value
It is any change, training G and D.
Step 4: operating procedure three repeatedly, until being finished compared with all a reference values.
Further, use the ratio k of the frequency of training of D and G as control variable；Dynamic adjustment k value meets different data
Constraint during collection training；K is controlled using following two inequality constraints；
Wherein PgIt is that D will generate the probability that data classification is truthful data；PmgBe on MNIST training obtain it is general
Rate instructs PgThe a reference value of value；LDIt is the penalty values of D in current training data；LmDIt is the benchmark loss of the training on MNIST
Value；α is threshold value predetermined.
Symbol definition
Further, include: based on probability/penalty values self-adaptation control method
Implement the self adaptive control learning improvement generation confrontation network another object of the present invention is to provide a kind of
The self adaptive control learning improvement of method generates confrontation network control system.
Confrontation network computer journey is generated with self adaptive control learning improvement another object of the present invention is to provide a kind of
Sequence, it is described to be generated described in confrontation network computer program implementation claim 1-3 any one with self adaptive control learning improvement
With self adaptive control learning improvement generate confrontation network method.
Another object of the present invention is to provide a kind of terminal, the terminal carrying is described to be learnt with self adaptive control
Improve the controller for generating confrontation network method.
Another object of the present invention is to provide a kind of computer readable storage mediums, including instruction, when it is in computer
When upper operation, so that computer executes described being generated with self adaptive control learning improvement and fights network method.
Due to PgOr LDStandard learning curve show identical trend in the training process, can be used above-mentioned based on Pg
Self-adaptation control method handle them.In the algorithm 1 proposed, when k be greater than 1 when, it means that D be trained to k step and
G is trained to a step.If current PgOr LDMore than the tolerance that threshold alpha defines, D is walked by (being indicated by k) based on following one
Increase or decrease 1.As a result case: overgaugeOr minus deviationIf it is the former, k is increased by 1；If it is the latter, k is reduced 1.When k is 1,
Indicate that G and D is alternately trained with a step respectively.If it is overgauge, increase D step (k=k+1) by making k be greater than 1；If it is
Minus deviation, by walking k less than 1 (k=1/ (k+1)) Lai Zengjia G.When k is less than 1, G training 1/k step, mono- step of D.If just
Deviation, then G step reduces 1；Otherwise, increase by 1.
Advantages of the present invention and good effect are as follows:
The invention proposes a kind of adaptive hyper parameter learning processes suitable for GAN, to improve the instruction of different data collection
Practice stability, to guarantee the quality of generation image.This is obtained by using under relatively easy, single data set
Well-trained learning curve (reference line) realized to instruct the training process on different complexity contextual data collection.
The invention also provides the adaptive GAN models with MLP and DCGAN framework, and have carried out a variety of confirmatory experiments, as a result table
Bright, the stability of original GAN training can be improved in this method really, and can be generalized to well various GAN mutation models and
The data set of different scenes.For following work, planning analysis preferably generates sample measurement, this may encourage GAN's
Convergence.Further, it is desirable to effect of the self-adaptive controlled making mechanism proposed in GAN Multimodal Learning is analyzed, it is available to extend it
Range.
The present invention improves traditional GAN in detail below:
Due to the Non-balanced Growth of D and G ability in training process, traditional GAN training is difficult to converge to suitable data life
At ability, the data on certain data sets (the especially more complex data set of application scenarios) is caused to generate ineffective, this hair
Performance can be improved in the bright training step or intensity by being adaptively adjusted D and G of clear proof；
Based on this observation, a kind of self adaptive control learning algorithm (generated new mould for GAN model is proposed
Type is called Ak-GAN), dynamic adjusts the ratio (k value) of the training step of generator and arbiter；
Compared with providing the performance between Ak-GAN and tradition GAN, the algorithm is shown in terms of generating picture quality
Superiority.
Symbol definition
Comparative experimental data is as follows:
The similarity generated between data distribution and truthful data distribution has used the common index Inception of industry
Score, calculation formula are as follows:
X is that generator generates data distribution (pg) sampled data, under conditions of p (y | x) is a given sampled data
Category distribution, p (y) are the data distributions of classification y.DKL(p (y | x) | | p (y)) KL distance between measurement p (y | x) and p (y)
(KL-divergence).Inception Score can quantitatively measure two performance indicators of generator: the image of generation
Clear and legible object must be contained, i.e. the entropy of p (y | x) wants low；The image that generator generates simultaneously must satisfy diversity,
I.e. the entropy of p (y) wants high.When above-mentioned two index value of an image is bigger, its Inception Score obtained is bigger,
That is the KL distance of p (y | x) and p (y) is bigger.
Following table compares the performance evaluating of the present invention with tradition GAN method, and each evaluation and test score is 50,000 image institutes
Goals for is averaged.Data set has used Anime and CelebA, and generator and arbiter use MLP framework.The present invention mentions
Method out has all surmounted conventional method on both data sets.
Traditional GAN and proposed by the present invention model (Ak-GAN) performance (Inception of the table 1. based on MLP framework
Scores it) evaluates and tests
Traditional GAN and proposed by the present invention model (Ak-GAN) performance (Inception of the table 2. based on CNN framework
Scores it) evaluates and tests
Detailed description of the invention
Fig. 1 is that being generated with self adaptive control learning improvement for implementation offer of the invention fights network method flow chart.
Fig. 2 is that the present invention implements the original GAN provided to the influence diagram of the data collection with different diversity levels.
Fig. 3 is the learning curve figure that the present invention implements the traditional GAN model provided.
Fig. 4 is that the present invention implements the comparison of the original GAN (MLP) and Ak-GAN (MLP) of offer on animation data collection
(animation face is derived from the image plate website of cartoon wallpaper, and cutting image only includes face.These image sizes are 96x
This data set forms figure by 51,223 color images).
Fig. 5 is the comparison that the present invention implements original GAN (DCGAN) and Ak-GAN (DCGAN) on the animation data collection provided
Figure.
Fig. 6 is that the present invention implements the GAN training curve figure provided；
(a), the successful GAN probability value curve graph of training: p_real is the probability that D exports true picture, p_
Generated is D to the probability graph for generating image output；(b), the successful GAN penalty values curve graph of training: D_Loss is D
Penalty values, G_Loss are the penalty values figures of G；(c), the GAN probability curve diagram of failure to train.
Fig. 7 is that the present invention implements generation image corresponding to 1 score of table provided；
Fig. 8 is that the present invention implements generation image corresponding to 2 score of table provided.
Specific embodiment
In order to make the objectives, technical solutions, and advantages of the present invention clearer, with reference to embodiments, to this hair
It is bright to be further elaborated.It should be appreciated that the specific embodiments described herein are merely illustrative of the present invention, not
For limiting the present invention.
Traditional GAN is unstable and mode is collapsed, and is especially trained to the data acquisition system comprising various visual objects
When；
In the prior art, two main problems to complicate the issue are: generator and arbiter in the training process
Model capacity unbalanced growth, which prevent effective study of generator；Lack the convergence that can be calculated, can be explained.
To solve problem of the prior art, application principle of the invention is described in detail below with reference to concrete scheme.
As shown in Figure 1, provided in an embodiment of the present invention include with self adaptive control learning improvement generation confrontation network method
Following steps:
S101, the training on relatively simple data set using any GAN model, until convergence.Batch when record convergence
Number (such as 500 times)
S102, exports the penalty values or decision probability value of first number to convergence all generations of batch number interval, and leads to
It crosses curve-fitting method to be fitted these values, the value after fitting is as benchmark penalty values (LmGAnd LmD) or baseline probability value
(PmrAnd Pmg)
S103, the training GAN on relative complex data set.In training process after every c batch number (iteration coefficient)
Relatively and calculate current output valve (LG、LD、PrAnd PgOne of four) with the difference of a reference value, if the difference is more than preparatory
The threshold value (α) of setting, then training G and D after being optionally adjusted to k value；If the difference in threshold value, is not done k value
Any change, training G and D.
S104 runs S103 repeatedly, until finishing compared with all a reference values.
It is provided by the invention to be based on P as the preferred embodiment of the present inventiongOr LDSelf-adaptation control method it is as follows:
Due to PgOr LDStandard learning curve show identical trend in the training process, use identical method come
Control them.In the algorithm 1 proposed, when k is greater than 1, it means that D is trained to k step and G is trained to a step.If worked as
Preceding PgOr LDMore than the tolerance that threshold alpha defines, D step (being indicated by k) is increased or decreased by 1 result case based on following one
Example: overgaugeOr minus deviationIf it is
The former, increases by 1 for k；If it is the latter, k is reduced 1.When k is 1, indicate that G and D is alternately trained with a step respectively.If it is
Overgauge increases D step (k=k+1) by making k be greater than 1；If it is minus deviation, by make k less than 1 (k=1 (k+1)) come
Increase G step.When k is less than 1, G training 1/k step, mono- step of D.If overgauge, G step reduces 1；Otherwise, increase by 1.
The application principle of concrete scheme to further describe the present invention is made the present invention below with reference to embodiment further
Description.
Embodiment:
1), confrontation generates network
GAN estimates generation model by antagonistic process, and wherein generator G and arbiter D plays chess.The input of D comes
From two data distributions: truthful data and generated data, the latter are generated by G.D is trained to can be complete by sample to maximize it
It is classified as true or synthesis entirely, and G can reduce the ability that D tells generated data then by training to the maximum extent.?
In original frame, training objective is defined as minimax method problem:
Wherein G is the data distribution P that input noise variable Pz (z) is mapped to generationgFunction, and D is one and will count
According to the function of space reflection to scalar value, wherein each value indicates the probability that specific sample is distributed from real data.Function G
GAN network is constituted with D, the neural network model that they are usually, and be trained to simultaneously on its objective function.G's and D
Loss function LGAnd LDIt is defined as foloows:
Wherein θDAnd θGIt is the parameter set of D and G.x(i)Indicate real data.G(z(i)) it is to receive random noise z by G(i)Afterwards
The generated data of generation.So P (x(i)；θD) it is the probability that truthful data x is classified as truthful data by D.P(G(z(i))；θD) be
D is by the data G (z of generation(i)) it is classified as the probability of truthful data.Hereinafter referred to as PrAnd Pg。
2), GAN is trained in terms of image synthesis to comprising the other image collection of opposite unitary class.These include
MNIST handwritten numeral, bedroom scene and simple animation personage, as shown in Fig. 2 (a) and Fig. 2 (b).However, for relative complex
The image collection of classification, GAN generally can not generate satisfactory generation effect.For example, the training on CIFAR-10 data set
When, the object identifiability of generation is poor, as shown in Fig. 2 (c) and Fig. 2 (d).
The convergence for balancing D and G is a sizable challenge.If one of them is trained excessively well, the training of GAN is just
It can become unstable.In practice, D is usually by over training, because D is starting Shi Tairong under the conditions of complex data collection
Easily win.
The training curve of MNIST and CIFAR-10 are tracked and have recorded, as shown in Figure 3.This diagram depicts in training process
In probability (the respectively P that is distributed by the true and generated data of D estimation from truthful datarAnd Pg) curve.Fig. 3 (a) is shown
GAN training on MNIST data set reaches convergent curve graph.Observation it is found that start when, PrAnd PgBetween exist it is very big
Gap, with trained progress, which is gradually become smaller, and finally converges on about probability value 0.5.The good stability of MNIST
It is reflected in the suitable high quality of composograph, as shown in Fig. 2 (a).However, the bad visual effect of CIFAR-10 can lead to
Unstable training curve is crossed to explain, as shown in Fig. 3 (b).
3), to the acclimatization training algorithm of GAN
A. basic ideas
The target of proposition method be instructed by using well-trained learning curve training on various data sets into
Journey, to realize compellent performance.The mode of the well-trained learning curve of description acquisition first, then proposes a kind of calculation
Method, the current value of algorithm constraint control variable is (for example, posterior probability PrAnd PgOr penalty values LDAnd LG) (instructed with a reference value
Practice the probability or penalty values of learning curve being always or usually as specified) between difference.
Because original GAN training objective function (Objective Function) is defined as minimax method double person travelling
Play problem, so determine that the convergence of G and D is very difficult because D and G minimize respective penalty values, it is such play chess
Process cannot be guaranteed to be finally reached convergence.Compared with the numerical value at a certain moment, complete learning curve is typically to identify training process
Whether benign effective means.Therefore, use above-mentioned learning curve as convergent benchmark metric.
For PgAnd LDIt proposes following inequality constraints, and uses the ratio (k) of frequency of training between D and G as only
One control variable.Traditional GAN, which is realized, sets 1 for k, and the present invention takes the method for dynamic adjustment k value to meet different data
Stability control during collection training.It is as follows to constrain inequality formula.
Wherein PgIt is generator in the data measurement (probability) true to nature currently synthesized；PmgIt is the training on MNIST data set
Obtained normal probability；LDIt is the penalty values of D in current training data；LmDIt is the standard of the D in training on MNIST data set
Penalty values；α is the threshold value for limiting current value and standard value departure degree.Note that α can be with manual setting.Abundant experimental results are aobvious
Show that it is that comparison is reasonable that α, which is set as 0.2,.
B. it is based on PgOr LDSelf adaptive control
Due to PgOr LDStandard learning curve show identical trend in the training process, use identical method come
Control them.In the algorithm 1 proposed, when k is greater than 1, it means that D is trained to k step and G is trained to a step.If worked as
Preceding PgOr LDMore than the tolerance that threshold alpha defines, D step (being indicated by k) is increased or decreased by 1 result case based on following one
Example: overgaugeOr minus deviationIf it is the former, k is increased
1；If it is the latter, k is reduced 1.When k is 1, indicate that G and D is alternately trained with a step respectively.If it is overgauge, pass through
K is set to be greater than 1 to increase D step (k=k+1)；If it is minus deviation, by walking k less than 1 (k=1/ (k+ 1)) Lai Zengjia G.Work as k
When less than 1, G training 1/k step, mono- step of D.If overgauge, G step reduces 1；Otherwise, increase by 1.
(Ak-GAN) of the invention is assessed by being applied to the GAN model with MLP and two kinds of CNN different frameworks.
In fig. 4 it is shown that using the experimental result of Ak-GAN based on MLP framework on different data sets.With tradition
The sample that GAN model generates is compared, and adaptive control technology enables GAN to capture facial recognizable feature, such as face wheel
Wide and eyes.But these samples also often generate noise.In addition, image sharpness still has improved space.Therefore, also by institute
The self adaptive control program of proposition is applied to DCGAN framework, to further increase visual quality, then by animation and CelebA
The result of data set is compared with traditional DCGAN.Fig. 5 is shown, compared with traditional DCGAN, the generation of Ak-GAN model has
Higher-quality image simultaneously slows down mode collapse phenomenon.
The foregoing is merely illustrative of the preferred embodiments of the present invention, is not intended to limit the invention, all in essence of the invention
Made any modifications, equivalent replacements, and improvements etc., should all be included in the protection scope of the present invention within mind and principle.
Claims (7)
1. a kind of generate confrontation network method with self adaptive control learning improvement, which is characterized in that described with self adaptive control
It practises improvement and generates the operation iteration ratio that confrontation network method is adaptively adjusted generation network and differentiation network, specific method packet
Include following steps:
Step 1: the training on relatively simple data set using any GAN model, until convergence, batch when record is restrained
Number；
Step 2: exporting the penalty values or decision probability value of first number to convergence all generations of batch number interval, and pass through
Curve-fitting method is fitted these values, and the value after fitting is as benchmark penalty values or baseline probability value；Benchmark penalty values
For LmGAnd LmD；Baseline probability value is PmrAnd Pmg；
Step 3: the training GAN on relative complex data set, in training process after every c batch number relatively and calculating is worked as
The difference of preceding output valve and a reference value optionally adjusts the instruction of G and D if the difference is more than pre-set threshold value (α)
Practice iteration ratio k value and training；If the difference in threshold value, does not do any change to k value, by original k value training G and D；
Step 4: operating procedure three repeatedly, until being finished compared with all a reference values.
2. generating confrontation network method with self adaptive control learning improvement as described in claim 1, which is characterized in that use D
Ratio k with the training step of G is as control variable；The value of dynamic adjustment k meets the constraint during different data collection training；Make
K is controlled with following two inequality constraints；
Wherein PgIt is that D will generate the probability that data classification is truthful data；PmgIt is the obtained probability of the training on MNIST, refers to
Lead PgThe a reference value of value；LDIt is the penalty values of D in current training data；LmDIt is the benchmark penalty values of the training on MNIST；α is
Threshold value predetermined.
3. generating confrontation network method with self adaptive control learning improvement as described in claim 1, which is characterized in that based on general
Rate/penalty values self-adaptation control method includes:
4. a kind of self adaptive control for implementing to generate confrontation network method described in claim 1 with self adaptive control learning improvement
It practises improving and generates confrontation network control system.
5. a kind of generate confrontation network computer program with self adaptive control learning improvement, which is characterized in that described with adaptive
Schistosomiasis control, which improves, to be generated described in confrontation network computer program implementation claim 1-3 any one with self adaptive control
It practises improving and generates confrontation network method.
6. a kind of terminal, which is characterized in that the terminal, which is carried, uses self adaptive control described in claims 1 to 3 any one
Learning improvement generates the controller of confrontation network method.
7. a kind of computer readable storage medium, including instruction, when run on a computer, so that computer is executed as weighed
Benefit requires to generate confrontation network method with self adaptive control learning improvement described in 1-3 any one.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910157777.1A CN109902824A (en) | 2019-03-02 | 2019-03-02 | It is a kind of to generate confrontation network method with self adaptive control learning improvement |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910157777.1A CN109902824A (en) | 2019-03-02 | 2019-03-02 | It is a kind of to generate confrontation network method with self adaptive control learning improvement |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109902824A true CN109902824A (en) | 2019-06-18 |
Family
ID=66946174
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910157777.1A Pending CN109902824A (en) | 2019-03-02 | 2019-03-02 | It is a kind of to generate confrontation network method with self adaptive control learning improvement |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109902824A (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111651642A (en) * | 2020-04-16 | 2020-09-11 | 南京邮电大学 | Improved TEXT-GAN-based flow data set generation method |
| CN111897809A (en) * | 2020-07-24 | 2020-11-06 | 中国人民解放军陆军装甲兵学院 | Command information system data generation method based on generation countermeasure network |
| CN113592553A (en) * | 2021-08-02 | 2021-11-02 | 广西大学 | Cloud energy storage double-layer optimization control method of competition condition generation type countermeasure network |
-
2019
- 2019-03-02 CN CN201910157777.1A patent/CN109902824A/en active Pending
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111651642A (en) * | 2020-04-16 | 2020-09-11 | 南京邮电大学 | Improved TEXT-GAN-based flow data set generation method |
| CN111651642B (en) * | 2020-04-16 | 2022-10-04 | 南京邮电大学 | Improved TEXT-GAN-based flow data set generation method |
| CN111897809A (en) * | 2020-07-24 | 2020-11-06 | 中国人民解放军陆军装甲兵学院 | Command information system data generation method based on generation countermeasure network |
| CN113592553A (en) * | 2021-08-02 | 2021-11-02 | 广西大学 | Cloud energy storage double-layer optimization control method of competition condition generation type countermeasure network |
| CN113592553B (en) * | 2021-08-02 | 2023-10-17 | 广西大学 | Cloud energy storage double-layer optimization control method for competitive condition generation type countermeasure network |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN105184735B (en) | A kind of portrait deformation method and device | |
| CN109902824A (en) | It is a kind of to generate confrontation network method with self adaptive control learning improvement | |
| CN112199535B (en) | Image classification method based on integrated knowledge distillation | |
| CN109214298B (en) | Asian female color value scoring model method based on deep convolutional network | |
| CN108171770A (en) | A kind of human face expression edit methods based on production confrontation network | |
| CN102567734B (en) | Specific value based retina thin blood vessel segmentation method | |
| CN106991408A (en) | The generation method and method for detecting human face of a kind of candidate frame generation network | |
| CN106778852A (en) | A kind of picture material recognition methods for correcting erroneous judgement | |
| US11068746B2 (en) | Image realism predictor | |
| TWI415011B (en) | Facial identification method and system using thereof | |
| CN103810490A (en) | Method and device for confirming attribute of face image | |
| CN105701468A (en) | Face attractiveness evaluation method based on deep learning | |
| CN110458247A (en) | The training method and device of image recognition model, image-recognizing method and device | |
| CN109773807B (en) | Motion control method and robot | |
| CN111282272B (en) | Information processing method, computer readable medium and electronic device | |
| CN110837326B (en) | Three-dimensional target selection method based on object attribute progressive expression | |
| CN110176050B (en) | Aesthetic optimization method for text generated image | |
| CN110598718A (en) | Image feature extraction method based on attention mechanism and convolutional neural network | |
| Chen | Research on college physical education model based on virtual crowd simulation and digital media | |
| CN104700416B (en) | The image segmentation threshold that view-based access control model understands determines method | |
| CN116824584A (en) | Diversified image description method based on conditional variation transducer and introspection countermeasure learning | |
| CN108734116B (en) | Face recognition method based on variable speed learning deep self-coding network | |
| CN113657466B (en) | Pre-training model generation method and device, electronic equipment and storage medium | |
| CN112419098B (en) | Power grid safety and stability simulation sample screening and expanding method based on safety information entropy | |
| CN113744129A (en) | Semantic neural rendering-based face image generation method and system |
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: 20190618 |
|
| RJ01 | Rejection of invention patent application after publication |