CN110598843A - Generation countermeasure network organization structure based on discriminator sharing and training method thereof - Google Patents
Generation countermeasure network organization structure based on discriminator sharing and training method thereof Download PDFInfo
- Publication number
- CN110598843A CN110598843A CN201910668780.XA CN201910668780A CN110598843A CN 110598843 A CN110598843 A CN 110598843A CN 201910668780 A CN201910668780 A CN 201910668780A CN 110598843 A CN110598843 A CN 110598843A
- Authority
- CN
- China
- Prior art keywords
- discriminator
- generator
- training
- parameters
- data
- 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
- 238000000034 method Methods 0.000 title claims abstract description 112
- 238000012549 training Methods 0.000 title claims abstract description 106
- 230000008520 organization Effects 0.000 title abstract description 13
- 230000006870 function Effects 0.000 claims description 35
- 230000008569 process Effects 0.000 claims description 35
- 238000004364 calculation method Methods 0.000 claims description 28
- 238000005070 sampling Methods 0.000 claims description 21
- 238000011478 gradient descent method Methods 0.000 claims description 17
- 230000001360 synchronised effect Effects 0.000 claims description 17
- 238000005457 optimization Methods 0.000 claims description 14
- 230000003042 antagnostic effect Effects 0.000 claims description 5
- 239000013589 supplement Substances 0.000 abstract description 2
- 238000013256 Gubra-Amylin NASH model Methods 0.000 description 16
- 238000013528 artificial neural network Methods 0.000 description 11
- 241001465754 Metazoa Species 0.000 description 5
- 238000013135 deep learning Methods 0.000 description 5
- 238000005516 engineering process Methods 0.000 description 5
- 230000006872 improvement Effects 0.000 description 5
- 238000006243 chemical reaction Methods 0.000 description 4
- 238000011160 research Methods 0.000 description 4
- 241000282326 Felis catus Species 0.000 description 3
- 238000013473 artificial intelligence Methods 0.000 description 3
- 230000008901 benefit Effects 0.000 description 3
- 230000000694 effects Effects 0.000 description 3
- 238000006467 substitution reaction Methods 0.000 description 3
- 210000004556 brain Anatomy 0.000 description 2
- 238000013527 convolutional neural network Methods 0.000 description 2
- 230000007123 defense Effects 0.000 description 2
- 238000013461 design Methods 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 238000007620 mathematical function Methods 0.000 description 2
- 230000002787 reinforcement Effects 0.000 description 2
- NAWXUBYGYWOOIX-SFHVURJKSA-N (2s)-2-[[4-[2-(2,4-diaminoquinazolin-6-yl)ethyl]benzoyl]amino]-4-methylidenepentanedioic acid Chemical compound C1=CC2=NC(N)=NC(N)=C2C=C1CCC1=CC=C(C(=O)N[C@@H](CC(=C)C(O)=O)C(O)=O)C=C1 NAWXUBYGYWOOIX-SFHVURJKSA-N 0.000 description 1
- 230000004913 activation Effects 0.000 description 1
- 230000003044 adaptive effect Effects 0.000 description 1
- 230000004075 alteration Effects 0.000 description 1
- 230000008034 disappearance Effects 0.000 description 1
- 238000005315 distribution function Methods 0.000 description 1
- 230000009977 dual effect Effects 0.000 description 1
- 238000003384 imaging method Methods 0.000 description 1
- 238000013178 mathematical model Methods 0.000 description 1
- 230000005012 migration Effects 0.000 description 1
- 238000013508 migration Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000008447 perception Effects 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 238000012552 review Methods 0.000 description 1
- 238000004088 simulation Methods 0.000 description 1
- 230000009466 transformation Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/044—Recurrent networks, e.g. Hopfield networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention provides a generation confrontation network organization structure based on discriminator sharing and a training method thereof. The invention fully understands the principle of GAN network learning promotion and supplements and strengthens the distribution integrity of generated data by introducing generators with the same generation target from other angles. The generated data characteristics of a plurality of generators are absorbed by the discriminator, so that the discrimination capability of the discriminator is more comprehensively improved; therefore, the generator learns more discrimination principles, obtains knowledge of other generators indirectly from feedback of the discriminator, and better promotes the reality of generated data. Therefore, compared with the traditional GAN, the method can better promote the reality and diversity of the generated data.
Description
Technical Field
The invention relates to the field of deep learning, in particular to a generation countermeasure network organization structure based on discriminator sharing and a training method thereof, which are used for countermeasure sample generation and defense application in the field of deep learning and are also suitable for combination of deep learning data generation application with the same function of other discriminators.
Background
With the appearance of big data and the rapid improvement of computing power, the artificial intelligence technology represented by deep learning is rapidly developed, so that the machine has sensing capability, can recognize images and can understand human languages. However, the artificial intelligence pursues not only the perception ability of the machine but also the creativity of the machine, so that the academic community rapidly develops the research on generating models, and the machine has imagination and learns the creativity. According to the probability statistics theory, the generation model is a model capable of randomly generating observation data under the condition of given implicit parameters, and a joint probability distribution is assigned to an observation value and a labeled data sequence (x, y). In the deep learning field, the generation model is realized in the form of a neural network, and can be used for modeling observation data, learning and mastering the characteristics of the observation data according to the observation data for training, and generating output data with specific distribution (simulating the observation data) according to the learned characteristics. The generative models can be classified into a direct type and an indirect type according to whether a distribution function of data can be completely expressed, and GAN (generative countermeasure network) concerned by the patent belongs to the indirect type generative model.
Although the generative model is only used for simulating observation data to generate similar simulated data, the generative model has wide application in the scientific field and the industrial field. The generated model can be used for high-dimensional probability distribution processing, ultrahigh-resolution imaging, artistic creation, image style migration, text-to-image conversion, future deduction in reinforcement learning, training data reinforcement and the like, is suitable for the industrial and academic fields, and can generate value in the consumer market, so that the method has important research significance.
The traditional generation model adopts a single neural network (simple human brain simulation mathematical model) to generate data by simulating human brain thinking. However, the traditional data generation is not rational, the structure and parameters of the neural network need to be optimized continuously, a large amount of training data sets are needed, and the final optimization result is not necessarily satisfactory. Therefore, the learner Ian Goodfellow generates a new idea, and the learner Ian Goodfellow adopts two neural networks to enable the two networks to quickly and efficiently learn the target task through the game and the confrontation of the two networks, so that a generation model (called a GAN model for short) based on a GAN structure is provided. The GAN model becomes one of hot artificial intelligence technologies due to the advantages of parallelism, less function constraint, good data generation effect and the like, and is evaluated as a ten-major breakthrough technology in 2018 in the world by the maja province science and technology review.
The core idea of the GAN model is the game of two neural networks, both of which can be represented as parameter-controlled differentiable functions, one neural network called the producer, which attempts to produce a data set that is co-distributed with the target data set (assuming that the target data is distributed as p)data) The data samples of (a); another neural network, called a discriminator, is responsible for determining whether the input data is true or false, i.e., whether the data is from the target data set or the generator. The discriminator adopts the traditional supervised learning technology to learn and distinguish the true and false of the input, and the generator deceives the discriminator to distinguish the true and false through the feedback learning of the discriminator. When both the games obtain enough strong ability through learning, the games of both the games reach Nash equilibrium (Nash equilibrium), and at this time, the discriminator cannot distinguish whether the data is true or false.
Let it be assumed that the mathematical function of the generator neural network is denoted G (its functional form is represented by its network parameter θ)(G)Decision), the mathematical function of the neural network of the discriminator is expressed as D (the functional form of which is represented by its network parameter θ)(D)Decision), x represents an observed variable, i.e. a target data set (x obeys p)dataDistribution), z represents a simple form of latent variable corresponding to the target data set (typically z is a random number that follows an even or normal distribution). The input of the generator is z, and the output G (z) represents the generationThe generated data generated by the discriminator has the input of x or G (z) and the output of 0 (indicating that the input is not the original target data) or 1 (indicating that the input is the original target data). The training goal of the generator is by adjusting θ(G)Such that G (z) is approximately equal to x, the training goal of the arbiter is by adjusting θ(D)Such that D (x) is approximately equal to 1 and D (G (z)) is approximately equal to 0. The ultimate goal of the GAN model is to find the optimal θ(D)Optimum theta under parametric conditions(G)In order to achieve the goal, a cost function needs to be set according to the training goal, and the parameter corresponding to the minimum value of the cost function is the optimal parameter.
For the discriminator, its cost function J(D)As shown in equation (1):
the cost function means the inverse of the sum of expectations of the discriminators for correctly identifying the target data and the likelihood of generating the data, where E represents the calculation expectation from the corresponding distribution of the subscript variable. Thus making the cost function J(D)The minimum parameter is the optimal parameter of the discriminator, as shown in equation (2).
For the generator, its cost function J(G)As in equation (3):
the meaning of the cost function is the expected inverse number of the possibility that the discriminator erroneously regards the generated data as the target data, and the stronger the capability of the discriminator is, the better the verisimilitude of the generated data is, and thus the optimal parameter θ is*(D)Under the condition that the cost function J is enabled(G)The minimum parameter is the optimal parameter of the generator, as shown in equation (4).
The design principle of GAN is described above, however, in practical application, the implementation method of GAN can be adjusted by means of approximation and the like, taking the actual situation and operability into consideration, and the specific adjustment is as follows: firstly, the extreme value calculation of the cost function generally adopts methods such as random gradient descent, momentum-based optimization, root mean square back propagation (RMSProp) and the like to realize the solution of the formula (4) in an iterative mode, and the invention takes the random gradient descent method as an example for description; secondly, according to the principle of the GAN, an optimal generator should be found through the countermeasure with the optimal discriminator, however, in the actual process, the optimal discriminator can cause the gradient disappearance problem, so that under the condition of adopting the gradient descent method, the feedback result of the optimal discriminator can not guide the parameter adjusting direction of the generator, therefore, in the actual operation, the GAN usually adopts the iterative training method of alternately updating the discriminator and the generator, the generator is updated for 1 time every time the discriminator updates k times, so that the performance of the generator and the discriminator is gradually and alternately improved, the process is repeated until the GAN meets the requirement or reaches the target training iteration times, and the k value is usually 1; furthermore the desired calculation in the cost function is in practice often replaced by a randomly sampled average. The pseudo code of the specific implementation method of GAN is shown in table 1. Wherein k represents the updating times of the discriminator in each training iteration process, and m represents the sample number of approximate expected calculation by mean value calculation.
TABLE 1 GAN implementation method pseudo code
With the progress of research, the GAN model has also formed many forms of improvement. The Deep Convolutional GAN (DCGAN) is a combination of a Deep Convolutional neural network and GAN, and 2 Deep Convolutional neural networks are used for respectively realizing a generator and a discriminator, the training process of the network is relatively stable, high-quality picture generation and other related applications can be effectively realized, and most of subsequent researches on GAN are based on the improvement of the structure.
Wasserstein GAN (WGAN) is another optimized form of GAN, which is only a four-point improvement over the original GAN: abandoning the activation function of the last layer of the discriminator neural network; the cost function in equation (3) no longer uses log calculations; after each time of updating the parameters of the discriminator, limiting the absolute value of the discriminator within a certain fixed constant range; momentum-based parameter update algorithms are not employed. Through the improvement of the 4 points, the problem of unstable GAN training is well solved, and the training degree of a generator and a discriminator does not need to be carefully balanced.
The classification label of the data is introduced into the data generation and discrimination process based on the classification optimized GAN (adaptive Classifier GAN, ACGAN), so that the data generation of the generator and the discrimination of the discriminator are more targeted. For example, the target data set is pictures of various animals, the traditional GAN working process generates animal pictures and judges whether the animal pictures are real or not, and then the animal pictures are updated iteratively; the ACGAN generates a picture of an animal of a corresponding type (generates a picture of a cat) by a generator according to an additionally input type tag (such as a cat) and judges whether the picture is a picture of a corresponding type in the target data set (judges whether the picture is a picture of a cat in the target data set).
Dual Generator generated countermeasure Network (G) for image transformation tasks2GAN) uses two generators to process image conversion and image restoration tasks, the two independent generators can realize different structures and parameters, and the discriminator is used to discriminate real images and converted pictures. However, in the actual training process, the discriminator only has feedback with the image conversion generator, the image restoration generator does not affect the discriminator, and only interacts with the image conversion generator, so the structure is essentially a GAN network assisted by an additional generation model (image restoration generator).
Although the training stability, pertinence and application flexibility of the GAN are improved, the adopted dual-network confrontation form only performs confrontation learning from a single angle, and cannot support multi-angle confrontation training. For example, in the field of countermeasure samples, in order to generate countermeasure samples, a generator in a conventional GAN can perform countermeasure learning based on one attack method and a discriminator, but cannot simultaneously consider multiple attack methods, let alone simultaneously consider attack and defense means (such as input reconstruction), so that the learning capability of the discriminator is more comprehensive and the theoretically optimal discriminator is difficult to achieve.
Disclosure of Invention
Aiming at the problem that the dual-network GAN cannot give consideration to multi-angle and multi-aspect antagonistic learning, the invention provides a novel GAN organization structure which can give consideration to the characteristics of fusing a plurality of angles and aspects of the problem aiming at a specific problem, thereby improving the performance of a GAN model. The new GAN organization structure is based on the shared GAN organization structure of the arbiter, and not only the traditional GAN model can be directly applied, but also the improved models of ACGAN and WGAN, etc. mentioned in the background section, which are directed to the traditional GAN model can be directly applied.
The technical scheme of the invention is as follows:
the generation countermeasure network organization structure based on discriminator sharing is characterized in that: the system comprises a discriminator and a plurality of generators, wherein each generator and the discriminator form a generation countermeasure network structure;
for each generator, its input is a random latent variable z, its output is generation data g (z), and g (z) has the same data structure as the target data set x; when updating the parameters, the generator receives a feedback result output by the discriminator and solves the feedback result to obtain the updated parameters of the generator;
the input In to the arbiter comprises the target data set x and the output of each generator, the output of the arbiter being a real number value representing the likelihood that the data In is from the target data set x; when updating parameters, the discriminator solves and obtains the updating parameters of the discriminator according to the data labels and the feedback results output by the discriminator; the data tag indicates that the data source is the target data set x or the output data of the generator.
In a further preferred aspect, the generator countermeasure network organization structure based on arbiter sharing is characterized in that: when the generator updates the parameters, the adopted solving method is a random gradient descent method, an optimization based on momentum or a root-mean-square back propagation method; when the discriminator updates the parameters, the adopted solving method is a random gradient descent, momentum-based optimization or a root-mean-square back propagation method.
In a further preferred aspect, the generator countermeasure network organization structure based on arbiter sharing is characterized in that: each generator and the arbiter form a GAN structure, an ACGAN structure or a WGAN structure.
The training method for generating the countermeasure network based on the arbiter sharing is characterized in that: the asynchronous training method is adopted and comprises the following steps:
step 1: initializing a generator number i to be 1, wherein the generator number i is a positive integer between [1 and n ], and represents the currently selected generator number, and n is the number of generators;
step 2: initializing the training times j of the single step discriminator to be 1; the training times j of the single step discriminator is an integer between [1, k ], wherein k represents that the discriminator needs to update k times of parameters every time the generator updates the parameters in the alternate training process of the generator and the discriminator;
and step 3: slave generator GiIs an input variable ziRandomly sampling m samples z in the distribution ofi (1)，…，zi (m)(ii) a Distributing p from a target datasetdataIn the random sampling of m samples x(1)，…，x(m)(ii) a According to a cost function formula of a discriminator
Resolving update discriminator parameter theta(D)WhereinIs a generator GiParameter of D (x)(l)) Indication inputIs x(l)The output of the discriminator in time is,is expressed as inputThe output of the discriminator in time is,is a generator GiAn output of (d);
and 4, step 4: judging whether the training frequency j of the single step discriminator is less than k, if j is less than k, j is added by 1, returning to the step 3 to continue to execute the single step of the discriminator, and if j is equal to k, entering the step 5;
and 5: slave generator GiIs an input variable ziRandomly sampling m samples z in the distribution ofi (1)，…，zi (m)(ii) a According to cost function formula of generator
Calculation update generator GiParameter(s)
Step 6: judging whether the generator number i is smaller than n, if i is smaller than n, adding 1 to i, returning to the step 2, and performing alternate training of the discriminator and the generator on the next generator; if i is equal to n, go to step 7;
and 7: and (4) judging whether a training termination condition is reached, if so, finishing the asynchronous training, otherwise, returning to the step 1, and iteratively executing the processes.
In a further preferred aspect, the training method for generating an anti-network based on arbiter sharing is characterized in that: when the generator parameters are solved and updated in the step 5, the adopted solving method is a random gradient descent, momentum-based optimization or root-mean-square back propagation method; and 3, when the parameters of the arbiter are solved and updated in the step 3, the adopted solving method is a random gradient descent, momentum-based optimization or root-mean-square back propagation method.
The training method for generating the countermeasure network based on the arbiter sharing is characterized in that: the synchronous training method comprises the following steps:
step 1: initializing the training times j of the single step discriminator to be 1; the training frequency j of the single step discriminator is an integer between [1, k ], wherein k represents that the discriminator needs to update k times of parameters when the generator updates the parameters every time in the synchronous training process of the generator and the discriminator;
step 2: respectively slave generator G1，…，Gi，…，GnIs an input variable z1，…，zi，…，znEach randomly sampling m samples z in the distribution of1 (1)，…，z1 (m)，…，zi (1)，…，zi (m)，…，zn (1)，…，zn (m)(ii) a Distributing p from a target datasetdataIn the random sampling of m samples x(1)，…，x(m)(ii) a According to a cost function formula of a discriminator
Resolving update discriminator parameter theta(D)WhereinIs a generator GiParameter of D (x)(l)) Representing the input as x(l)The output of the discriminator in time is,is expressed as inputThe output of the discriminator in time is,is a generator GiAn output of (d);
and step 3: judging whether the training frequency j of the single step discriminator is less than k, if j is less than k, j is added with 1, returning to the step 2 to continue to execute the single step of the discriminator, and if j is equal to k, entering the step 4;
and 4, step 4: respectively slave generator G1，…，Gi，…，GnIs an input variable z1，…，zi，…，znEach randomly sampling m samples z in the distribution of1 (1)，…，z1 (m)，…，zi (1)，…，zi (m)，…，zn (1)，…，zn (m)(ii) a Respectively resolving and updating parameters of each generator according to the cost functions of n generators in the synchronous training method, and for the ith generator GiUsing a cost function
Calculation update generator GiParameter(s)
And 5: and (4) judging whether a training termination condition is reached, if so, finishing the synchronous training, otherwise, returning to the step 1, and iteratively executing the processes.
In a further preferred aspect, the training method for generating an anti-network based on arbiter sharing is characterized in that: when the generator parameters are solved and updated in the step 4, the adopted solving method is a random gradient descent, momentum-based optimization or root-mean-square back propagation method; and (3) when the parameters of the discriminator are solved and updated in the step (2), the adopted solving method is a random gradient descent method, an optimization method based on momentum or a root-mean-square back propagation method.
Advantageous effects
In the traditional GAN network, the confrontation games between a generator and a discriminator are mutually promoted, and in each confrontation iteration process, generated data and real data generated by the generator are mixed and then are firstly sent to the discriminator; the discriminator distinguishes the real data and generates the data and then outputs the own judgment result, and according to the own judgment result and the actual result, the discriminator absorbs the characteristics of the data generated by the learning generator, adjusts the own network parameters and improves the distinguishing capability; and the generator absorbs the judgment principle of the learning judger according to the deception effect of the data generated by the generator, adjusts the network parameters of the generator and improves the reality of the data generated by the generator. This is the principle of GAN network learning promotion. However, the data generated by a single generator cannot completely cover the precise distribution of the real data, so that the learning promotion process of the whole GAN is limited in an incomplete data distribution, and the reality and diversity of the generated data are limited.
The invention fully understands the principle of GAN network learning promotion and supplements and strengthens the distribution integrity of the generated data by introducing generators with the same generation target (aiming at the same real data distribution) from other angles. The generated data characteristics of a plurality of generators are absorbed by the discriminator, so that the discrimination capability of the discriminator is more comprehensively improved; therefore, the generator learns more discrimination principles, obtains knowledge of other generators indirectly from feedback of the discriminator, and better promotes the reality of generated data. Therefore, compared with the traditional GAN, the method can better promote the reality and diversity of the generated data.
Additional aspects and advantages of the invention will be set forth in part in the description which follows and, in part, will be obvious from the description, or may be learned by practice of the invention.
Drawings
The above and/or additional aspects and advantages of the present invention will become apparent and readily appreciated from the following description of the embodiments, taken in conjunction with the accompanying drawings of which:
FIG. 1 is a schematic diagram of the organization of a discriminator-shared GAN model;
FIG. 2 is a schematic flow chart of an asynchronous training method based on a discriminator-shared GAN model;
fig. 3 is a flow chart of a synchronous training method based on a arbiter-shared GAN model.
Detailed Description
The invention provides a GAN organizational structure and a training method based on arbiter sharing, and ACGAN, WGAN and other improved methods aiming at the traditional GAN mentioned in the background art can be simply and directly transplanted into the invention. For clarity, the present embodiment only describes the present invention by taking a conventional GAN as an example, and the implementation process is as follows:
in the first step, the organization structure of the GAN model based on the arbiter sharing is designed, as shown in fig. 1. The structure comprises a discriminator and a plurality of (two or more) generators, any generator and the discriminator can form a traditional GAN structure, and the discriminator and the generators can be formed by any neural network. The input of the generator is a random latent variable z, the output is generated data G (z), and the output has the same data structure as the target data set x; when updating the parameters, the generator receives the feedback result output by the discriminator, and instructs the generator to update the parameters by a random gradient descent method (other solving methods can replace the random gradient descent method, and the embodiment only takes the random gradient descent as a representative for explanation, which is the same below). The inputs to the arbiter comprise the target data set x and the generator output, the output of the arbiter being a real number (typically 0, 1)]Real numbers In between), representing the likelihood of the data being true or false, i.e., the likelihood of the data In coming from the target data set x; when updating parameters, the discriminator will determine the data label (indicating whether the data source is x or G)iGenerated data) and self-output feedback results, and guiding the discriminator to update parameters by a random gradient descent method.
And secondly, designing a training method of the GAN model based on the arbiter sharing. According to the organizational structure characteristics of the GAN model shared by the discriminators, the invention designs two training methods of asynchronous training and synchronous training and corresponding cost functions, and any method can finish the training of the GAN model shared by the discriminators, wherein the asynchronous training method integrates the knowledge of all generators step by step to guide the training and learning of the discriminators, and the synchronous training method can integrate the knowledge of all generators globally to guide the training and learning of the discriminators, but the calculation density (unit time or the calculation complexity of single-step training) is higher.
The asynchronous training method selects n generators in turn and alternates with the arbiter for training and learning, and the process is as shown in fig. 2. The training process adopts a random gradient descent method to guide parameter updating aiming at the corresponding cost function, and a generator G of an asynchronous training methodiThe cost function of the discriminator D corresponds to equation (5) and equation (6), respectively, where the index i represents the current selection of the ith generator. In actual operation, the mean value calculation is used to approximate the expected calculation, and the equations (5) and (6) are converted into the equations (7) and (8) after the substitution.
The synchronous training method synchronously integrates all generators and alternately performs training learning with the arbiter, and the process is as shown in fig. 3. The training process adopts a random gradient descent method to guide parameter updating aiming at the corresponding cost function, and a generator G of the training method is synchronizediThe cost function is the same as the asynchronous training method in formula (5), and the cost function of the discriminator D is in formula (9). In actual operation, the mean value calculation is used to approximate the expected calculation, and the formula (5) and the formula (9) are converted into the formula (7) and the formula (10) after the substitution.
The following detailed description of embodiments of the invention is intended to be illustrative, and not to be construed as limiting the invention.
Fig. 1 is a schematic organization structure diagram of a GAN model based on arbiter sharing. The structure includes a discriminantAnd n (n is more than or equal to 2) generators. Wherein the generator Gi(1. ltoreq. i. ltoreq.n) as the input of a random latent variable ziOutput as generated data Gi(zi) Outputting a data structure having the same as the target data set x; when updating the parameters, the generator receives the output result D (In G) of the discriminatori(zi) Guiding to adjust the internal network parameters of the generator by a random gradient descent method)The input In of the arbiter uses different input data at different stages, including the target data set x and the n generator outputs G1(z1)，……，Gn(zn) The output of the discriminator is a real number (usually [0,1 ]]Real numbers In between), representing the likelihood of the data being true or false, i.e., the likelihood of the data In coming from the target data set x; when updating the parameters, the discriminator guides and adjusts the internal network parameters theta of the discriminator by a random gradient descent method according to the data labels and the feedback result D (in) output by the discriminator(D)。
Fig. 2 is a schematic flow chart of an asynchronous training method. Given that the training termination condition is that the specified iteration number or the generated data reaches the specified index, the asynchronous training method is specifically described as follows:
(1) the initialization generator number i is 1. The generator number i is a positive integer between [1, n ] and represents the currently selected generator number, and n is the number of generators.
(2) And initializing the number j of the single step discriminator training times to be 1. The training time j of the single step discriminator is an integer between [1, k ], where k represents that the discriminator needs to be updated k times every time the generator updates the parameters in the alternating training process of the generator and the discriminator, and k can be 1 generally.
(3) Slave generator GiIs an input variable ziRandomly sampling m samples z in the distribution ofi (1)，…，zi (m). m represents the number of sample samples whose mean value calculation approximates the desired calculation, and this sampling process provides the calculation data for the updating of the discriminator parameter in the step (5).
(4) Distributing p from a target datasetdataIn the random sampling of m samples x(1)，…，x(m). m represents the number of sample samples whose mean value calculation approximates the desired calculation, and this sampling process provides the calculation data for the updating of the discriminator parameter in the step (5).
(5) Discriminator cost function for asynchronous training method(equation (8)), the discriminator parameter θ is updated using a random gradient descent method(D)。
(6) And judging whether the training times j of the single step discriminator is less than k. If j is less than k, j is added by 1, and the process jumps to the step (3) to continue executing the single-step training of the discriminator; and if j is equal to k, sequentially executing the step (7).
(7) Slave generator GiIs an input variable ziRandomly sampling m samples z in the distribution ofi (1)，…，zi (m). m represents the number of sample samples whose mean value calculation approximates the desired calculation, which provides the calculation data for the generator parameter update in step (8).
(8) Generator cost function for asynchronous training method(equation (7)), the generator parameters are updated using a random gradient descent method
(9) It is determined whether the generator number i is less than n. If i is less than n, i is added by 1, the process jumps to the step (2), and alternate training of the discriminator and the generator is executed aiming at the next generator; and if i is equal to n, sequentially executing the step (10).
(10) And judging whether the training termination condition is reached. If so, the asynchronous training is finished. If not, the process jumps to the step (1) and the above processes are executed iteratively.
Fig. 3 is a schematic flow chart of a synchronization training method. Given that the training termination condition is that the specified iteration times or the generated data reach the specified index, the synchronous training method is specifically described as follows:
(1) and initializing the number j of the single step discriminator training times to be 1. The training time j of the single step discriminator is an integer between [1, k ], where k represents that the discriminator needs to be updated k times, and k can be 1, every time the generator updates the parameters in the synchronous training process of the generator and the discriminator.
(2) Respectively slave generator G1，…，Gi，…，GnIs an input variable z1，…，zi，…，znEach randomly sampling m samples z in the distribution of1 (1)，…，z1 (m)，…，zi (1)，…，zi (m)，…，zn (1)，…，zn (m). m represents the number of samples in each generator that are calculated as a mean value to approximate the desired calculation, and this sampling process provides the calculated data for the updating of the discriminator parameters in step (4).
(3) Distributing p from a target datasetdataIn the random sampling of m samples x(1)，…，x(m). m represents the number of sample samples that approximate the desired computation with a mean computation, which provides the computation data for the arbiter parameter update in step (4).
(4) Discriminator cost function for synchronous training method(equation (10)), the discriminator parameter θ is updated using a random gradient descent method(D)。
(5) And judging whether the training times j of the single step discriminator is less than k. If j is less than k, j is added by 1, and the process jumps to the step (2) to continue executing the single-step training of the discriminator; and if j is equal to k, sequentially executing the step (6).
(6) Respectively slave generator G1，…，Gi，…，GnIs an input variable z1，…，zi，…，znEach randomly sampling m samples z in the distribution of1 (1)，…，z1 (m)，…，zi (1)，…，zi (m)，…，zn (1)，…，zn (m). m represents the number of sample samples in each generator that are calculated as a mean value to approximate the desired calculation, which provides the calculation data for the generator parameter update in step (7).
(7) Cost function respectively aiming at n generators in synchronous training method (all in the form of equation (7)), the generator parameters are each updated using a random gradient descent method
(8) And judging whether the training termination condition is reached. If so, the synchronous training is finished. If not, the process jumps to the step (1) and the above processes are executed iteratively.
Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention, and that variations, modifications, substitutions and alterations can be made in the above embodiments by those of ordinary skill in the art without departing from the principle and spirit of the present invention.
Claims (7)
1. A discriminator sharing based generative confrontation network organizational structure, characterized by: the system comprises a discriminator and a plurality of generators, wherein each generator and the discriminator form a generation countermeasure network structure;
for each generator, its input is a random latent variable z, its output is generation data g (z), and g (z) has the same data structure as the target data set x; when updating the parameters, the generator receives a feedback result output by the discriminator and solves the feedback result to obtain the updated parameters of the generator;
the input In to the arbiter comprises the target data set x and the output of each generator, the output of the arbiter being a real number value representing the likelihood that the data In is from the target data set x; when updating parameters, the discriminator solves and obtains the updating parameters of the discriminator according to the data labels and the feedback results output by the discriminator; the data tag indicates that the data source is the target data set x or the output data of the generator.
2. The structure of claim 1, wherein the structure comprises: when the generator updates the parameters, the adopted solving method is a random gradient descent method, an optimization based on momentum or a root-mean-square back propagation method; when the discriminator updates the parameters, the adopted solving method is a random gradient descent, momentum-based optimization or a root-mean-square back propagation method.
3. The structure of claim 1, wherein the structure comprises: each generator and the arbiter form a GAN structure, an ACGAN structure or a WGAN structure.
4. A training method for generating an antagonistic network based on the arbiter sharing as described in claim 1, wherein: the asynchronous training method is adopted and comprises the following steps:
step 1: initializing a generator number i to be 1, wherein the generator number i is a positive integer between [1 and n ], and represents the currently selected generator number, and n is the number of generators;
step 2: initializing the training times j of the single step discriminator to be 1; the training times j of the single step discriminator is an integer between [1, k ], wherein k represents that the discriminator needs to update k times of parameters every time the generator updates the parameters in the alternate training process of the generator and the discriminator;
and step 3: slave generator GiIs an input variable ziRandomly sampling m samples z in the distribution ofi (1)，…，zi (m)(ii) a Distributing p from a target datasetdataIn the random sampling of m samples x(1)，…，x(m)(ii) a According to a cost function formula of a discriminator
Resolving update discriminator parameter theta(D)WhereinIs a generator GiParameter of D (x)(l)) Representing the input as x(l)The output of the discriminator in time is,is expressed as inputThe output of the discriminator in time is,is a generator GiAn output of (d);
and 4, step 4: judging whether the training frequency j of the single step discriminator is less than k, if j is less than k, j is added by 1, returning to the step 3 to continue to execute the single step of the discriminator, and if j is equal to k, entering the step 5;
and 5: slave generator GiIs an input variable ziRandomly sampling m samples z in the distribution ofi (1)，…，zi (m)(ii) a According to cost function formula of generator
Calculation update generator GiParameter(s)
Step 6: judging whether the generator number i is smaller than n, if i is smaller than n, adding 1 to i, returning to the step 2, and performing alternate training of the discriminator and the generator on the next generator; if i is equal to n, go to step 7;
and 7: and (4) judging whether a training termination condition is reached, if so, finishing the asynchronous training, otherwise, returning to the step 1, and iteratively executing the processes.
5. The training method for generating an antagonistic network based on the arbiter sharing according to claim 4, wherein: when the generator parameters are solved and updated in the step 5, the adopted solving method is a random gradient descent, momentum-based optimization or root-mean-square back propagation method; and 3, when the parameters of the arbiter are solved and updated in the step 3, the adopted solving method is a random gradient descent, momentum-based optimization or root-mean-square back propagation method.
6. A training method for generating an antagonistic network based on the arbiter sharing as described in claim 1, wherein: the synchronous training method comprises the following steps:
step 1: initializing the training times j of the single step discriminator to be 1; the training frequency j of the single step discriminator is an integer between [1, k ], wherein k represents that the discriminator needs to update k times of parameters when the generator updates the parameters every time in the synchronous training process of the generator and the discriminator;
step 2: respectively slave generator G1，…，Gi，…，GnIs an input variable z1，…，zi，…，znEach randomly sampling m samples z in the distribution of1 (1)，…，z1 (m)，…，zi (1)，…，zi (m)，…，zn (1)，…，zn (m)(ii) a Distributing p from a target datasetdataIn the random sampling of m samples x(1)，…，x(m)(ii) a According to a cost function formula of a discriminator
Resolving update discriminator parameter theta(D)WhereinIs a generator GiParameter of D (x)(l)) Representing the input as x(l)The output of the discriminator in time is,is expressed as inputThe output of the discriminator in time is,is a generator GiAn output of (d);
and step 3: judging whether the training frequency j of the single step discriminator is less than k, if j is less than k, j is added with 1, returning to the step 2 to continue to execute the single step of the discriminator, and if j is equal to k, entering the step 4;
and 4, step 4: respectively slave generator G1，…，Gi，…，GnIs an input variable z1，…，zi，…，znEach randomly sampling m samples z in the distribution of1 (1)，…，z1 (m)，…，zi (1)，…，zi (m)，…，zn (1)，…，zn (m)(ii) a Respectively resolving and updating parameters of each generator according to the cost functions of n generators in the synchronous training method, and for the ith generator GiUsing a cost function
Calculation update generator GiParameter(s)
And 5: and (4) judging whether a training termination condition is reached, if so, finishing the synchronous training, otherwise, returning to the step 1, and iteratively executing the processes.
7. The training method for generating an antagonistic network based on the arbiter sharing according to claim 6, wherein: when the generator parameters are solved and updated in the step 4, the adopted solving method is a random gradient descent, momentum-based optimization or root-mean-square back propagation method; and (3) when the parameters of the discriminator are solved and updated in the step (2), the adopted solving method is a random gradient descent method, an optimization method based on momentum or a root-mean-square back propagation method.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910668780.XA CN110598843B (en) | 2019-07-23 | 2019-07-23 | Training method for generating countermeasure network organization structure based on discriminant sharing |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910668780.XA CN110598843B (en) | 2019-07-23 | 2019-07-23 | Training method for generating countermeasure network organization structure based on discriminant sharing |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110598843A true CN110598843A (en) | 2019-12-20 |
| CN110598843B CN110598843B (en) | 2023-12-22 |
Family
ID=68853176
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910668780.XA Active CN110598843B (en) | 2019-07-23 | 2019-07-23 | Training method for generating countermeasure network organization structure based on discriminant sharing |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110598843B (en) |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111860782A (en) * | 2020-07-15 | 2020-10-30 | 西安交通大学 | Triple multi-scale CycleGAN, fundus fluorography generation method, computer device, and storage medium |
| CN112232129A (en) * | 2020-09-17 | 2021-01-15 | 厦门熙重电子科技有限公司 | Electromagnetic information leakage signal simulation system and method based on generation countermeasure network |
| CN112380767A (en) * | 2020-11-11 | 2021-02-19 | 山东大学 | Equipment fault diagnosis method and system based on improved generation countermeasure network |
| CN112508175A (en) * | 2020-12-10 | 2021-03-16 | 深圳先进技术研究院 | Multi-task learning type generation type confrontation network generation method and system for low-dose PET reconstruction |
| CN112565261A (en) * | 2020-12-04 | 2021-03-26 | 浙江大学 | Multi-generator AugGAN-based dynamic malicious API sequence generation method |
| CN113362403A (en) * | 2021-07-20 | 2021-09-07 | 支付宝(杭州)信息技术有限公司 | Training method and device of image processing model |
| CN114545255A (en) * | 2022-01-18 | 2022-05-27 | 广东工业大学 | Lithium battery SOC estimation method based on competitive generation type antagonistic neural network |
| CN116168789A (en) * | 2023-04-26 | 2023-05-26 | 之江实验室 | Multi-center medical data generation system and method |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20180101770A1 (en) * | 2016-10-12 | 2018-04-12 | Ricoh Company, Ltd. | Method and system of generative model learning, and program product |
| CN107944358A (en) * | 2017-11-14 | 2018-04-20 | 华南理工大学 | A kind of human face generating method based on depth convolution confrontation network model |
| CN108171266A (en) * | 2017-12-25 | 2018-06-15 | 中国矿业大学 | A kind of learning method of multiple target depth convolution production confrontation network model |
| CN108665058A (en) * | 2018-04-11 | 2018-10-16 | 徐州工程学院 | A kind of generation confrontation network method based on segmentation loss |
| CN109190620A (en) * | 2018-09-03 | 2019-01-11 | 苏州科达科技股份有限公司 | License plate sample generating method, system, equipment and storage medium |
| US20190087730A1 (en) * | 2017-09-19 | 2019-03-21 | Preferred Networks, Inc. | Non-transitory computer-readable storage medium storing improved generative adversarial network implementation program, improved generative adversarial network implementation apparatus, and learned model generation method |
| US20190171908A1 (en) * | 2017-12-01 | 2019-06-06 | The University Of Chicago | Image Transformation with a Hybrid Autoencoder and Generative Adversarial Network Machine Learning Architecture |
-
2019
- 2019-07-23 CN CN201910668780.XA patent/CN110598843B/en active Active
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20180101770A1 (en) * | 2016-10-12 | 2018-04-12 | Ricoh Company, Ltd. | Method and system of generative model learning, and program product |
| US20190087730A1 (en) * | 2017-09-19 | 2019-03-21 | Preferred Networks, Inc. | Non-transitory computer-readable storage medium storing improved generative adversarial network implementation program, improved generative adversarial network implementation apparatus, and learned model generation method |
| CN107944358A (en) * | 2017-11-14 | 2018-04-20 | 华南理工大学 | A kind of human face generating method based on depth convolution confrontation network model |
| US20190171908A1 (en) * | 2017-12-01 | 2019-06-06 | The University Of Chicago | Image Transformation with a Hybrid Autoencoder and Generative Adversarial Network Machine Learning Architecture |
| CN108171266A (en) * | 2017-12-25 | 2018-06-15 | 中国矿业大学 | A kind of learning method of multiple target depth convolution production confrontation network model |
| CN108665058A (en) * | 2018-04-11 | 2018-10-16 | 徐州工程学院 | A kind of generation confrontation network method based on segmentation loss |
| CN109190620A (en) * | 2018-09-03 | 2019-01-11 | 苏州科达科技股份有限公司 | License plate sample generating method, system, equipment and storage medium |
Non-Patent Citations (3)
| Title |
|---|
| "Rob-GAN: Generator, Discriminator, and Adversarial Attacker", ARXIV:1807.10454 * |
| 张龙, 赵杰煜, 叶绪伦, 董伟: "协作式生成对抗网络", 自动化学报, vol. 44, no. 05 * |
| 杜立平, 宋燕红, 李晓东, 金鑫: "GAN在AI领域的应用研究", 北京电子科技学院学报, vol. 26, no. 03 * |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111860782A (en) * | 2020-07-15 | 2020-10-30 | 西安交通大学 | Triple multi-scale CycleGAN, fundus fluorography generation method, computer device, and storage medium |
| CN112232129A (en) * | 2020-09-17 | 2021-01-15 | 厦门熙重电子科技有限公司 | Electromagnetic information leakage signal simulation system and method based on generation countermeasure network |
| CN112380767A (en) * | 2020-11-11 | 2021-02-19 | 山东大学 | Equipment fault diagnosis method and system based on improved generation countermeasure network |
| CN112380767B (en) * | 2020-11-11 | 2023-04-07 | 山东大学 | Equipment fault diagnosis method and system based on improved generation countermeasure network |
| CN112565261A (en) * | 2020-12-04 | 2021-03-26 | 浙江大学 | Multi-generator AugGAN-based dynamic malicious API sequence generation method |
| CN112508175A (en) * | 2020-12-10 | 2021-03-16 | 深圳先进技术研究院 | Multi-task learning type generation type confrontation network generation method and system for low-dose PET reconstruction |
| CN113362403A (en) * | 2021-07-20 | 2021-09-07 | 支付宝(杭州)信息技术有限公司 | Training method and device of image processing model |
| CN114545255A (en) * | 2022-01-18 | 2022-05-27 | 广东工业大学 | Lithium battery SOC estimation method based on competitive generation type antagonistic neural network |
| CN116168789A (en) * | 2023-04-26 | 2023-05-26 | 之江实验室 | Multi-center medical data generation system and method |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110598843B (en) | 2023-12-22 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110598843A (en) | Generation countermeasure network organization structure based on discriminator sharing and training method thereof | |
| Seo et al. | Reinforcement learning with action-free pre-training from videos | |
| CN110580501B (en) | Zero sample image classification method based on variational self-coding countermeasure network | |
| CN107437077A (en) | A kind of method that rotation face based on generation confrontation network represents study | |
| Bera et al. | Efficient trajectory extraction and parameter learning for data-driven crowd simulation. | |
| CN110427799B (en) | Human hand depth image data enhancement method based on generation of countermeasure network | |
| CN107330902B (en) | Chaotic genetic BP neural network image segmentation method based on Arnold transformation | |
| JP7144699B2 (en) | SIGNAL MODIFIER, METHOD AND PROGRAM | |
| Ulhaq et al. | Efficient diffusion models for vision: A survey | |
| Zhang et al. | Unifying generative models with GFlowNets and beyond | |
| Lin et al. | Evolutionary architectural search for generative adversarial networks | |
| Luo et al. | Diff-instruct: A universal approach for transferring knowledge from pre-trained diffusion models | |
| CN109086664A (en) | A kind of polymorphic gesture identification method of sound state fusion | |
| Valdez et al. | A framework for interactive structural design exploration | |
| Baykal et al. | Exploring deshufflegans in self-supervised generative adversarial networks | |
| CN110415261B (en) | Expression animation conversion method and system for regional training | |
| CN115914505A (en) | Video generation method and system based on voice-driven digital human model | |
| Chen et al. | PaDGAN: A generative adversarial network for performance augmented diverse designs | |
| Chen | Research on college physical education model based on virtual crowd simulation and digital media | |
| Heimel et al. | The MadNIS Reloaded | |
| Mathov et al. | Enhancing real-world adversarial patches through 3D modeling of complex target scenes | |
| ZHANG et al. | Realistic face image generation based on generative adversarial network | |
| Yin et al. | Respect: Reinforcement learning based edge scheduling on pipelined coral edge tpus | |
| CN112634456B (en) | Real-time high-realism drawing method of complex three-dimensional model based on deep learning | |
| CN115511758A (en) | Image generation method based on image interpolation generation coding technology |
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 | ||
| TR01 | Transfer of patent right | ||
| TR01 | Transfer of patent right |
Effective date of registration: 20240221 Address after: 471000 intersection of Yingzhou road and Guanlin Road, Luolong District, Luoyang City, Henan Province Patentee after: UNIT 63891 OF PLA Country or region after: China Address before: 471000 Box 061, Luoyang City, Henan Province Patentee before: 63880 UNIT OF THE PLA Country or region before: China |