CN111033532A - Training method and system for generating countermeasure network, electronic device, and storage medium - Google Patents
Training method and system for generating countermeasure network, electronic device, and storage medium Download PDFInfo
- Publication number
- CN111033532A CN111033532A CN201980002600.2A CN201980002600A CN111033532A CN 111033532 A CN111033532 A CN 111033532A CN 201980002600 A CN201980002600 A CN 201980002600A CN 111033532 A CN111033532 A CN 111033532A
- Authority
- CN
- China
- Prior art keywords
- network
- generating
- networks
- picture
- generated
- 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
- 238000012549 training Methods 0.000 title claims abstract description 114
- 238000000034 method Methods 0.000 title claims abstract description 64
- 238000003860 storage Methods 0.000 title claims description 9
- 230000003042 antagnostic effect Effects 0.000 claims description 16
- 238000012545 processing Methods 0.000 claims description 9
- 238000010586 diagram Methods 0.000 description 8
- 230000008569 process Effects 0.000 description 8
- 230000006870 function Effects 0.000 description 6
- 230000000694 effects Effects 0.000 description 5
- 238000013527 convolutional neural network Methods 0.000 description 4
- 230000009471 action Effects 0.000 description 3
- 238000004891 communication Methods 0.000 description 3
- 238000009826 distribution Methods 0.000 description 3
- 238000004519 manufacturing process Methods 0.000 description 3
- 238000005457 optimization Methods 0.000 description 3
- 239000013598 vector Substances 0.000 description 3
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 230000005540 biological transmission Effects 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 238000013136 deep learning model Methods 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 230000002452 interceptive effect Effects 0.000 description 1
- 238000003672 processing method Methods 0.000 description 1
- 230000011218 segmentation Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Software Systems (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Medical Informatics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Artificial Intelligence (AREA)
- Image Analysis (AREA)
Abstract
The application discloses a training method and a training system for generating a confrontation network. The generating countermeasure network includes a plurality of generating networks and a plurality of discriminating networks, each of the generating networks connecting a plurality of discriminating networks, the method includes: generating a picture based on the generation network; judging the probability that the picture is a real picture or a generated picture by utilizing a plurality of judging networks connected with the generating network; selecting an appointed judging network corresponding to the generating network according to the judged probability of the real picture or the generated picture; and updating the parameters of the plurality of generated networks according to the specified discriminating network.
Description
Technical Field
The embodiment of the disclosure relates to the field of deep learning, in particular to a training method and system for generating a confrontation network, an electronic device and a storage medium.
Background
Generation of a countermeasure network (GAN) is a deep learning model, and is one of the most promising methods for unsupervised learning in complex distribution in recent years. The model passes through (at least) two modules in the framework: the mutual game learning of the generation network and the discrimination network produces a rather good output.
At present, generation of countermeasure networks has wide applications including, for example, image generation, image conversion, watermarking, semantic segmentation, high-resolution picture generation, and the like. However, there are still at least two problems in the current application:
1, convergence is difficult when the confrontation network is generated for training; a large amount of training data is required, for example, pictures are generated based on high-dimensional vectors and trained;
2, it is difficult to generate a high-resolution picture; the accuracy of the production countermeasure network needs to be continuously improved, and when the accuracy of the model is high enough, the generated high-resolution picture can meet the requirement.
The above description of the discovery process of the problems is only for the purpose of aiding understanding of the technical solutions of the present disclosure, and does not represent an admission that the above is prior art.
Disclosure of Invention
To solve at least one problem of the prior art, at least one embodiment of the present disclosure provides a training method for generating a countermeasure network, where the countermeasure network includes a plurality of generation networks and a plurality of discriminant networks, each of the generation networks connects a plurality of discriminant networks, and the method includes:
generating a picture based on the generation network;
judging the probability that the picture is a real picture or a generated picture by utilizing a plurality of judging networks connected with the generating network;
selecting an appointed judging network corresponding to the generating network according to the probability of the real picture or the generating picture judged by each judging network;
and updating the parameters of the plurality of generated networks according to the specified discriminating network.
In an embodiment of the training method for generating an antagonistic network, the multiple generating networks have the same structure and different initial parameters; the plurality of judgment networks have the same structure and different initial parameters.
In an embodiment of the training method for generating an antagonistic network, before the step of generating a picture based on the generated network, the method further comprises:
judging whether the generated confrontation network meets a training target or not;
wherein the training targets comprise: for each generated network of the generated countermeasure network, the probability that the picture judged by any one of the connected discrimination networks is a real picture or the probability that the picture is generated is converged.
In an embodiment of the training method for generating an antagonistic network, the specified discriminative network is determined based on the updated picture generated by the generating network.
In an embodiment of the training method for generating an antagonistic network, the step of selecting the designated discrimination network corresponding to the generation network according to the determined probability of the real picture or the generation picture includes:
and selecting the discrimination network with the lowest true picture probability or the highest generated picture probability as the designated discrimination network.
In an embodiment of the training method for generating the countermeasure network, the number of the generation networks and the number of the discrimination networks in the generation countermeasure network are both two or more.
In an embodiment of the training method of generating an antagonistic network, for the plurality of generating networks, when one generating network satisfies the training target, a next generating network is trained.
In an embodiment of the training method for generating the countermeasure network, for the plurality of generation networks, the parameters are sequentially updated for each generation network in an iterative manner until the plurality of generation networks all meet the training target.
In an embodiment of the training method for generating an antagonistic network, the method further comprises:
and training the generated network to be trained by using the generated confrontation network after training.
In an embodiment of the training method for generating the countermeasure network, the training the generation network to be trained by using the generated countermeasure network after training includes:
calculating the Loss value between each generated network and the generated network to be trained to obtain a plurality of Loss values;
and training the network to be trained according to the plurality of Loss values, so that the sum of the plurality of Loss values is converged to the minimum.
In an embodiment of the training method for generating the countermeasure network, each discriminant network is connected to only one generating network in the generating countermeasure network.
Another aspect of the present disclosure also provides a computing processing device, including: a processor and a memory; the processor is configured to perform the aforementioned steps of the training method for generating an antagonistic network by calling a program or instructions stored in the memory.
In another aspect, the present disclosure also proposes a non-transitory computer-readable storage medium storing a program or instructions for causing a computer to perform the aforementioned steps of the training method for generating an antagonistic network.
It is apparent that at least one of the embodiments of the present disclosure provides a training scheme for generating an anti-network, so as to achieve the effects of increasing the convergence speed of the generated anti-network and reducing the convergence difficulty.
Drawings
To more clearly illustrate the technical solutions of the embodiments of the present disclosure, the drawings used in the description of the embodiments or the prior art will be briefly described below, and it is obvious that the drawings in the following description are only some embodiments of the present disclosure, and other drawings can be obtained by those skilled in the art according to the drawings.
Fig. 1 is a flowchart of a training method for generating a countermeasure network according to an embodiment of the present disclosure.
Fig. 2A and 2B are schematic block diagrams of two generation countermeasure networks provided by the embodiment of the disclosure.
Fig. 3 is a flowchart of a training method for generating a countermeasure network according to a second embodiment of the disclosure.
Fig. 4 is a flowchart of the substeps of step S106 of the second embodiment of the present disclosure.
Fig. 5 is a block diagram of a system for generating a confrontation network training system according to an embodiment of the present disclosure.
Fig. 6 is a block diagram of a computing processing device provided by an embodiment of the present disclosure.
Detailed Description
In order that the above objects, features and advantages of the present disclosure can be more clearly understood, the present disclosure will be described in further detail with reference to the accompanying drawings and examples. It is to be understood that the embodiments described are only a few embodiments of the present disclosure, and not all embodiments. The specific embodiments described herein are merely illustrative of the disclosure and are not intended to be limiting. All other embodiments derived by one of ordinary skill in the art from the described embodiments of the disclosure are intended to be within the scope of the disclosure.
It is noted that, in this document, relational terms such as "first" and "second," and the like, may be used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions.
Aiming at the problems of slow convergence and increased convergence difficulty of the generated countermeasure network in the prior art, the embodiment of the disclosure provides a training scheme for generating the countermeasure network, and the effects of improving the convergence speed of the generated countermeasure network and reducing the convergence difficulty are achieved, so that a basis is provided for bringing better effects to the countermeasure network in application.
In the training method for generating a countermeasure network according to the embodiment of the present disclosure, the generating countermeasure network includes a plurality of generating networks and a plurality of discriminating networks, each of the generating networks is connected to the plurality of discriminating networks, fig. 1 is a flowchart of the training method for generating a countermeasure network according to an embodiment of the present disclosure, and as shown in fig. 1, the method may include the following steps:
s101, generating a picture based on the generation network;
s102, judging the probability that the picture is a real picture or a generated picture by utilizing a plurality of judging networks connected with the generating network;
s103, selecting an appointed judging network corresponding to the generating network according to the judged real picture or the probability of the generating picture;
s104, updating the parameters of the plurality of generating networks according to the specified judging network;
these steps are described in detail below.
Step S101, generating pictures based on the generation network;
fig. 2A is a schematic diagram of a generation countermeasure network 100a according to an embodiment of the disclosure. As described above, the generative countermeasure network 100a includes N generative networks and M discriminative networks. In fig. 2A, N production networks are represented by generating networks 110, 120, 130, and M discriminating networks are represented by three discriminating networks 210, 220, 230. In some embodiments, after the training of the generative warfare network 100a is completed, the generative warfare network 100a may be used to train the generative network 200, and after the training of the generative warfare network 200 based on the generative warfare network 100a is completed, the network may be used for subsequent processing. In some embodiments, the training of the generation network 200 may also use the discriminant network 500 for auxiliary updating, and the training process of the generation network 200 can be completed more efficiently by using the discriminant network 500 for auxiliary updating.
Each of the N generation networks is connected to the M discrimination networks. That is, generation network 110 as shown in FIG. 2A is connected to the three illustrated discrimination networks 210, 220, and 230, and generation network 120 and generation network 130 are also connected to the three illustrated discrimination networks 210, 220, and 230, respectively.
In step S101, each of the generation networks 110, 120, 130 in the generation countermeasure network 100a generates pictures, which may be pictures generated based on random vectors, referred to as "generation pictures", respectively; accordingly, a picture actually taken may become a "real picture". The generated picture can be used as training data, is input into a generated countermeasure network, is used for training the generated countermeasure network together with a plurality of real pictures, enables a judgment network in the generated countermeasure network to judge the received picture to be the real picture and the generated picture more accurately through continuous iteration, accordingly feeds back the generated network, enables the generated network to generate pictures which are in a false or a false state through continuous countermeasure iteration, and completes a training target.
The training target can be a preset target for generating whether the confrontation network generated picture meets the requirement or not. The sum of the probability that a picture is a real picture and the probability that the picture is generated is 1, and due to the convergence property of the function, the training target for generating the countermeasure network can be that the predicted picture is the real picture or the predicted picture is close to 0.5. When the training target is judged not to be reached, the foregoing steps S101 to S104 may be continuously performed until the probability of judging as a real picture or generating a picture meets the requirement.
After step S101 is performed, step S102 may be performed as follows:
s102, judging the probability that the picture is a real picture or a generated picture by utilizing a plurality of judging networks connected with the generating network;
in an embodiment, the sum of the probability that the picture is a real picture and the probability that the picture is generated is 1.
In this step, the discrimination network may discriminate whether the picture is a real picture or a generated picture, and output a probability of the picture being a real picture or a probability of the generated picture. Each discrimination network independently performs discrimination, and has a probability value.
In some embodiments, the input content of the discrimination network is only from the generation network, and the pictures received by the discrimination network are all generated pictures, and for a certain discrimination network, when the probability that the discrimination network determines that a picture is generated is higher, the higher the discrimination network determines that the discrimination network is accurate, the parameters of the generation network can be modified by back propagation by using the discrimination network as a reference, so that the generation network is optimized, and a picture which cannot be accurately determined by the discrimination network is generated. The plurality of generation networks can independently generate pictures at the same time, and transmit the pictures to the plurality of judgment networks connected with the generation networks, and the pictures are judged by the plurality of judgment networks respectively. As shown in fig. 2A, the plurality of decision networks 210, 220, and 230 may, for example, first decide the pictures generated by the generation network 110 to obtain a plurality of probabilities, respectively, and then decide the pictures generated by the generation network 120 to obtain a plurality of probabilities, respectively; finally, the pictures generated by the generation network 130 are discriminated to obtain a plurality of probabilities.
After step S102 is performed, step S103 may be performed as follows:
s103, selecting an appointed judging network corresponding to the generating network according to the judged real picture or the probability of the generating picture;
in this step, when a judgment network judges that the probability of the picture is the generated picture is higher, or the probability value of the picture is judged to be really the picture is lower, the judgment result of the judgment network is proved to be more accurate, and the judgment network is taken as the designated judgment network corresponding to the generated network.
After step S103 is performed, step S104 may be performed as follows:
s104, updating the parameters of the plurality of generating networks according to the specified judging network;
in this step, after the designated discriminating network is determined, the network parameters of the plurality of generating networks may be updated in a back propagation manner according to the probability of the designated discriminating network, for example, after the designated discriminating network is determined, the parameters of the other N-1 generating networks may be modified according to the probability of the designated discriminating network, so that the pictures generated by the generating networks are more spurious.
For example, assume that the generating network is g (z), where z is a random noise, and the generating network converts this random noise into data type x. Using the example of a picture generation scenario, the output of the generation network is a picture. For the discriminant network, the output for any input x, d (x) is a real number in the range of 0-1, which is used to determine how likely the picture is a real picture or the generated picture is. Let Pr and Pg represent the distribution of the real image and the distribution of the generated image, respectively, and obtain the objective function of the discrimination network as follows:
similar generation networks aim to make the discrimination network unable to distinguish between real pictures and generated pictures, so the whole optimization objective function is as follows:
the optimization method of the maximum minimization objective function has various optimization modes, and the most intuitive processing method can be that interactive iteration is respectively carried out on the judgment network parameter and the generation network parameter, the generation network parameter is fixed, the judgment network parameter is optimized, and after a period of time, the judgment network parameter is fixed and the generation network parameter is optimized again until the process is converged.
After step S104 is executed, when it is determined that the preset condition is not met, the operation of generating the picture in step S101 may be continuously executed by using the updated generation network, and the updated generation network may be sent to the plurality of discrimination networks for determination, so as to select a new discrimination network. After one iteration, the probability of judging the network output of the picture for generating the picture is reduced to 60% but the critical value is not reached, and then a new iteration can be continued until the probability of judging the network output is continuously reduced.
In addition, for the multiple generation networks in the generation countermeasure network, these generation networks 110-.
As shown in fig. 2A, the generation countermeasure network 100a, for a plurality of generation networks among which, when one generation network satisfies the training target, the next generation network is trained. For example, the generation network 110 may be trained first, a plurality of probabilities are obtained by using the plurality of discrimination networks 210 and 230, a designated discrimination network is selected, parameters of the generation network 110 are updated in a back propagation manner, the designated discrimination network is selected based on the updated pictures output by the generation network 110, and then the parameters of the generation network 110 are updated. Until the training of the generation network 110 is completed, the generation network 120 is trained by using the plurality of discrimination networks 210 and 230 until the training of the generation network 120 is completed, and the generation network 130 is trained until the training of the generation network 130 is completed, thereby completing the training target of the generation of the confrontation network 100 a.
Similarly, for the multiple generation networks in the generation countermeasure network 100a, the generation networks 110 and 130 may be alternately trained before the training targets are not met, and the parameters are updated for each generation network in turn until the multiple generation networks meet the training targets. For example, the generation network 110 may be trained first, a plurality of probabilities are obtained by using the plurality of discrimination networks 210 and 230, a designated discrimination network is selected, the parameters of the generation network 110 are updated in a back propagation manner, the designated discrimination network is selected based on the updated pictures output by the generation network 110, the generation network 120 is trained after the updating, a plurality of probabilities are obtained by using the plurality of discrimination networks 210 and 230, the designated discrimination network is selected, the parameters of the generation network 120 are updated in a back propagation manner, the designated discrimination network is selected based on the updated pictures output by the generation network 130, and the process is repeated until the generation countermeasure network 100b completes the training target.
In other embodiments, the discriminative network to which each generating network is connected is different from the discriminative networks to which other generating networks are connected. Fig. 2B is a schematic diagram of another generation countermeasure network 100B of the present invention. As shown in the generation countermeasure network 100B of fig. 2B, the number of discrimination networks may be N times the generation network, each of which is connected to a unique generation network. As shown in fig. 2B, the determination networks 210 and 230 are connected to the generation network 110, the determination networks 310 and 330 are connected to the generation network 120, and the determination networks 410 and 430 are connected to the generation network 130. In the training for generating the countermeasure network 100b, each of the generating networks 110 and 130 can be independently trained until the probability value output by each of the discriminating networks 210 and 430 in the generating countermeasure network 100b converges to 0.5.
It is to be noted that, in addition to the generation countermeasure network 100B, the generation network 200 and the discrimination network 500 to which the generation countermeasure network 100B is connected are also shown in fig. 2B. In some embodiments, after the training of the generative warhead network 100b is completed, the generative warhead network 200 may be trained using the generative warhead network 100b, and after the training of the generative warhead network 200 based on the generative warhead network 100b is completed, the training may be used for subsequent processing. In some embodiments, the training of the generation network 200 may also use the discriminant network 500 for auxiliary updating, and the training process of the generation network 200 can be completed more efficiently by using the discriminant network 500 for auxiliary updating.
In view of the above, the embodiment of the present disclosure provides a training method for generating an anti-network, which is directed at the problems of slow convergence and high convergence difficulty of the anti-network generated in the prior art, and provides a training scheme for generating an anti-network, in which a plurality of generation networks and a plurality of training networks are connected to each other, and a designated discrimination network selected by one generation network is used to perform back propagation on parameters of other generation networks, so as to achieve the effects of increasing the convergence speed of the generated anti-network and reducing the convergence difficulty.
Fig. 3 is a flow chart of generating a confrontation model according to another embodiment of the present invention, and as shown in fig. 3, in an alternative embodiment, before step S101 of the present invention, step S100 may be performed as follows:
and S100, judging whether the generated confrontation network meets the training target.
In this step, a determination may be made as to whether the antagonistic network meets the training objectives. As described above, the generated countermeasure network model is verified to have a convergent property, and when the condition that the discrimination value of each discrimination network converges to 0.5 is not satisfied, the training of the generated countermeasure network can be continued until the training target of the generated countermeasure network is satisfied through multiple iterations, that is, for each picture output by the generation network, the probability that the connected discrimination network judges that the picture is a real picture or the generated picture converges to 0.5, and when the condition that the discrimination value of each discrimination network converges to 0.5 is determined, the training of the generated countermeasure network is qualified, and the training is stopped.
Thus, in an alternative embodiment, the method may further comprise the steps of:
and S105, when the judged probability that the picture is a real picture is close to the probability that the picture is a generated picture aiming at each judgment network of the generated confrontation network, finishing the training of the generated confrontation network.
In this step, when it is determined that the probability of each of the real pictures or generated pictures output by the discrimination network is close to 0.5 and converges to 0.5 (for example, the probability of 0.5 is not fluctuated up and down beyond a specified range), it is considered that the generation of the countermeasure network has sufficiently converged, and the training of the generation of the countermeasure network is terminated.
In an alternative embodiment, the plurality of generating networks have the same structure and different initial parameters; the plurality of judgment networks have the same structure and different initial parameters.
The same structure here may include: the generated networks have the same structure, and the judgment networks have the same structure.
For example, each discriminant network is a convolutional neural network model (CNN) having the same layer structure, and each generation network is a structure similar to but opposite to the convolutional neural network model (CNN), and these generation networks also have the same layer structure with respect to each other.
In an optional embodiment, the steps further comprise:
and S106, training the generated network to be trained by using the generated confrontation network after training.
As shown in fig. 2A, after the training of the generation countermeasure network is completed, the generation network to be trained 200 is trained based on the trained generation network. For example, the Loss values between each generated network and the generated network to be trained 200 may be calculated to obtain N Loss values. The Loss value refers to a distance between the generated network generation result and the generated result of the generation network 200.
The generating network 200 is trained based on the sum of the N Loss values such that the sum of the N Loss values converges to a minimum.
Therefore, in an alternative embodiment, as shown in fig. 4, the step S106 of training the generated network to be trained by using the generated counterpoise network completed by training may include the following sub-steps:
s1061, calculating Loss values between each generated network and the generated network to be trained to obtain a plurality of Loss values;
s1062, training the network to be trained according to the multiple Loss values, so that the sum of the multiple Loss values is converged to the minimum.
In the sub-step, the Loss value refers to, for example, a distance between the generation result of the generation network 210 and 230 and the generation result of the generation network 200 in fig. 2A. When training the generation network 200, the synthetic distance between the picture of the generation network 200 and the N generation networks 210 and 230 generated pictures that have been trained is minimized, and the trained generation network 200 can be obtained.
In some embodiments, as shown in fig. 2A, a discriminant network 500 is connected to the generation network 200 for facilitating updates to the generation network 200.
In an optional embodiment, the number of the generating networks and the number of the discriminating networks are both more than two. In some embodiments, the number of discriminant networks is n times the number of generating networks, where the number of generating networks is greater than 2, and n is a natural number.
Fig. 5 is a block diagram of a system 300 for generating a confrontation network training provided by an embodiment of the present disclosure.
As shown in fig. 5, the generate confrontation network training system 300 may include, but is not limited to, the following elements: a generation unit 301 for generating a picture based on a generation network; a judging unit 302, configured to judge, by using multiple discrimination networks connected to the generation network, a probability that the picture is a real picture or a generated picture; a selecting unit 303, configured to select an appointed decision network corresponding to the generation network according to the determined probability of the real picture or the generated picture; an updating unit 304, configured to update parameters of the multiple generation networks according to the specified discrimination network.
The generation unit 301 is configured to generate a picture based on a generation network. In the schematic diagram of the generation countermeasure network 100a according to an embodiment of the present disclosure shown in fig. 2A, the generation countermeasure network 100a includes N generation networks and M discrimination networks. In fig. 2A, N production networks are represented by generating networks 110, 120, 130, and M discriminating networks are represented by three discriminating networks 210, 220, 230.
Each of the N generation networks is connected to the M discrimination networks. That is, generation network 110 as shown in FIG. 2A is connected to the three illustrated discrimination networks 210, 220, and 230, and generation network 120 and generation network 130 are also connected to the three illustrated discrimination networks 210, 220, and 230, respectively.
Each of the generating countermeasure networks 110, 120, 130 in the generating countermeasure network 100a generates pictures, which may be pictures generated based on random vectors, as training data, referred to as "generating pictures" to indicate that they are randomly generated pictures rather than real pictures. In contrast, a picture actually taken may be referred to as a "real picture". The generated pictures are input as training data to generate the confrontation network 100a, and are used for training the generated confrontation network together with a plurality of real pictures, so that the discrimination network in the generated confrontation network can judge the real pictures and the generated pictures more accurately through continuous iteration, the generated network is fed back, and the generated network generates pictures which are false and spurious.
As mentioned above, the "training target" may be a target that is preset to determine whether the generated confrontation network generation picture meets the requirement. In an embodiment, due to the convergent nature of the function, the training goal to generate the countermeasure network may be, for example, that the probability predicted by the discrimination network meets a specified requirement, such as near 0.5. And when the judgment fails to reach the training target, the training can be continuously and circularly carried out until the judgment that the probability meets the requirement.
The determining unit 302 is configured to determine, by using a plurality of discrimination networks connected to the generating network, a probability that the picture is a real picture or a generated picture. The discrimination network may discriminate whether the feature value of the picture is fake or real. Each discrimination network can output the discrimination judgment probability value of which the feature value of the picture is fake. Since the feature values of the pictures are fake, for a certain discrimination network, when the probability that the discrimination network judges that the picture is fake is higher, the judgment of the discrimination network is more accurate.
The selecting unit 303 is configured to select an appointed decision network corresponding to the generation network according to the determined probability of the real picture or the generated picture.
When a judgment network judges that the probability of the picture as the generated picture is higher or the probability value of the picture as the true picture is lower, the judgment result of the judgment network is proved to be more accurate, and the judgment network is used as the designated judgment network corresponding to the generated network.
The updating unit 304 is configured to, after the designated judgment network is determined, perform back-propagation updating on the network parameters of the plurality of generation networks according to the probability of the designated judgment network, for example, after the designated judgment network is determined, modify the parameters for N-1 other generation networks by using the probability of the designated judgment network, so that the pictures generated by the generation networks are more spurious.
In some embodiments, the division of each unit in the generation countermeasure network training system 300 is only one logical function division, and there may be another division manner when the actual implementation is performed, for example, the generation unit 301, the judgment unit 302, the selection unit 303, and the update unit 304 may be implemented as one unit; the generating unit 301, the determining unit 302, the selecting unit 303, or the updating unit 304 may be divided into a plurality of sub-units. It will be understood that the various units or sub-units may be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether such functionality is implemented as hardware or software depends upon the particular application and design constraints imposed on the implementation. Skilled artisans may implement the described functionality in varying ways for each particular application.
Fig. 6 is a schematic structural diagram of a computing processing device according to an embodiment of the present disclosure. The computing processing device includes: at least one processor 401, at least one memory 402, and at least one communication interface 403. The various components in the in-vehicle device are coupled together by a bus system 404. A communication interface 403 for information transmission with an external device. Understandably, the bus system 404 is operative to enable connective communication between these components. The bus system 404 includes a power bus, a control bus, and a status signal bus in addition to a data bus. For clarity of illustration, the various buses are labeled as bus system 404 in fig. 6.
It will be appreciated that the memory 402 in this embodiment can be either volatile memory or nonvolatile memory, or can include both volatile and nonvolatile memory.
In some embodiments, memory 402 stores the following elements, executable units or data structures, or a subset thereof, or an expanded set thereof: an operating system and an application program.
The operating system includes various system programs, such as a framework layer, a core library layer, a driver layer, and the like, and is used for implementing various basic services and processing hardware-based tasks. The application programs, including various application programs such as a Media Player (Media Player), a Browser (Browser), etc., are used to implement various application services. The program for implementing the training method for generating the countermeasure network provided by the embodiment of the present disclosure may be included in the application program.
In the embodiment of the present disclosure, the processor 401 is configured to execute the steps of the embodiments of the training method for generating an anti-confrontation network provided by the embodiments of the present disclosure by calling a program or an instruction stored in the memory 402, specifically, a program or an instruction stored in an application program.
The training method for generating the countermeasure network provided by the embodiment of the disclosure can be applied to the processor 401, or implemented by the processor 401. The processor 401 may be an integrated circuit chip having signal processing capabilities. In implementation, the steps of the above method may be performed by integrated logic circuits of hardware or instructions in the form of software in the processor 401. The Processor 401 may be a general-purpose Processor, a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), an off-the-shelf Programmable Gate Array (FPGA) or other Programmable logic device, discrete Gate or transistor logic device, or discrete hardware components. A general purpose processor may be a microprocessor or the processor may be any conventional processor or the like.
The steps of the training method for generating the countermeasure network provided by the embodiment of the disclosure can be directly implemented by a hardware decoding processor, or implemented by a combination of hardware and software units in the decoding processor. The software elements may be located in ram, flash, rom, prom, or eprom, registers, among other storage media that are well known in the art. The storage medium is located in a memory 402, and the processor 401 reads information in the memory 402 and performs the steps of the method in combination with its hardware.
It is noted that, for simplicity of description, the foregoing method embodiments are described as a series of acts or combination of acts, but those skilled in the art will appreciate that the disclosed embodiments are not limited by the order of acts described, as some steps may occur in other orders or concurrently with other steps in accordance with the disclosed embodiments. In addition, those skilled in the art can appreciate that the embodiments described in the specification all belong to alternative embodiments.
Embodiments of the present disclosure also provide a non-transitory computer-readable storage medium, where the non-transitory computer-readable storage medium stores a program or instructions, and the program or instructions cause a computer to perform steps of various embodiments of a training method for generating an anti-confrontation network, which are not described herein again to avoid repeated descriptions.
It should be noted that, in this document, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising … …" does not exclude the presence of other identical elements in a process, method, article, or apparatus that comprises the element.
Those skilled in the art will appreciate that while some embodiments described herein include some features included in other embodiments, rather than others, combinations of features of different embodiments are meant to be within the scope of the disclosure and form different embodiments.
Those skilled in the art will appreciate that the description of each embodiment has a respective emphasis, and reference may be made to the related description of other embodiments for those parts of an embodiment that are not described in detail.
Although the embodiments of the present disclosure have been described in conjunction with the accompanying drawings, those skilled in the art may make various modifications and variations without departing from the spirit and scope of the present disclosure, and such modifications and variations fall within the scope defined by the appended claims.
Industrial applicability
At least one of the embodiments of the present disclosure provides a training scheme for generating an countermeasure network, which achieves the effects of increasing the convergence speed of the generation of the countermeasure network and reducing the difficulty of convergence, and has industrial applicability.
Claims (13)
1. A training method for generating a countermeasure network, the generating countermeasure network including a plurality of generating networks and a plurality of discriminating networks, each of the generating networks connecting a plurality of discriminating networks, the method comprising:
generating a picture based on the generation network;
judging the probability that the picture is a real picture or a generated picture by utilizing a plurality of judging networks connected with the generating network;
selecting an appointed judging network corresponding to the generating network according to the probability of the real picture or the generating picture judged by each judging network;
and updating the parameters of the plurality of generated networks according to the specified discriminating network.
2. Training method for generation of antagonistic networks according to claim 1, characterized in that said plurality of generation networks have the same structure and different initial parameters; the plurality of judgment networks have the same structure and different initial parameters.
3. The training method for generating an antagonistic network according to claim 1, characterized in that before the step of generating a picture based on said generated network, said method further comprises:
judging whether the generated confrontation network meets a training target or not;
wherein the training targets comprise: for each generated network of the generated countermeasure network, the probability that the picture judged by any one of the connected discrimination networks is a real picture or the probability that the picture is generated is converged.
4. The training method for generating an antagonistic network as claimed in claim 1,
the specified discriminatory network is determined based on the updated picture generated by the generating network.
5. The training method for generating an antagonistic network according to claim 1, wherein the step of selecting the designated discrimination network corresponding to the generation network according to the judged probability of the real picture or the generation picture comprises:
and selecting the discrimination network with the lowest true picture probability or the highest generated picture probability as the designated discrimination network.
6. The training method for generating countermeasure networks according to claim 1, wherein the number of the generation networks and the number of the discrimination networks in the generation countermeasure networks are two or more.
7. The training method for generating an antagonistic network according to claim 1, wherein, for the plurality of generating networks, when one generating network satisfies the training target, a next generating network is trained.
8. The training method for generating an antagonistic network according to claim 1, wherein the updating parameters for each of the plurality of generating networks are iterated in turn until the plurality of generating networks are all satisfied with the training target.
9. The training method for generating an antagonistic network according to one of the claims 1 to 8, characterized in that it further comprises:
and training the generated network to be trained by using the generated confrontation network after training.
10. The training method for generating a countermeasure network according to claim 9, wherein the step of training the generation network to be trained using the generated countermeasure network completed by training includes:
calculating the Loss value between each generated network and the generated network to be trained to obtain a plurality of Loss values;
and training the network to be trained according to the plurality of Loss values, so that the sum of the plurality of Loss values is converged to the minimum.
11. The training method for generating countermeasure networks according to claim 1, wherein each discrimination network is connected to only one generation network in the generation countermeasure networks.
12. A computing processing device, comprising: a processor and a memory;
the processor is adapted to perform the steps of the method of any one of claims 1 to 11 by calling a program or instructions stored in the memory.
13. A non-transitory computer-readable storage medium storing a program or instructions for causing a computer to perform the steps of the method according to any one of claims 1 to 11.
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| PCT/CN2019/121028 WO2021102697A1 (en) | 2019-11-26 | 2019-11-26 | Method and system for training generative adversarial network, and electronic device and storage medium |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN111033532A true CN111033532A (en) | 2020-04-17 |
| CN111033532B CN111033532B (en) | 2024-04-02 |
Family
ID=70198326
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201980002600.2A Active CN111033532B (en) | 2019-11-26 | 2019-11-26 | Training method and system for generating countermeasure network, electronic device and storage medium |
Country Status (2)
| Country | Link |
|---|---|
| CN (1) | CN111033532B (en) |
| WO (1) | WO2021102697A1 (en) |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107563995A (en) * | 2017-08-14 | 2018-01-09 | 华南理工大学 | A kind of confrontation network method of more arbiter error-duration models |
| CN109785258A (en) * | 2019-01-10 | 2019-05-21 | 华南理工大学 | A kind of facial image restorative procedure generating confrontation network based on more arbiters |
| CN109801292A (en) * | 2018-12-11 | 2019-05-24 | 西南交通大学 | A kind of bituminous highway crack image partition method based on generation confrontation network |
| CN110059465A (en) * | 2019-04-24 | 2019-07-26 | 腾讯科技（深圳）有限公司 | Auth method, confrontation generate training method, device and the equipment of network |
| CN110222705A (en) * | 2019-04-23 | 2019-09-10 | 华为技术有限公司 | A kind of training method and relevant apparatus of network model |
Family Cites Families (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20170249339A1 (en) * | 2016-02-25 | 2017-08-31 | Shutterstock, Inc. | Selected image subset based search |
| CN107220600B (en) * | 2017-05-17 | 2019-09-10 | 清华大学深圳研究生院 | A kind of Picture Generation Method and generation confrontation network based on deep learning |
| CN107392973B (en) * | 2017-06-06 | 2020-01-10 | 中国科学院自动化研究所 | Pixel-level handwritten Chinese character automatic generation method, storage device and processing device |
| CN107945204B (en) * | 2017-10-27 | 2021-06-25 | 西安电子科技大学 | Pixel-level image matting method based on generation countermeasure network |
-
2019
- 2019-11-26 CN CN201980002600.2A patent/CN111033532B/en active Active
- 2019-11-26 WO PCT/CN2019/121028 patent/WO2021102697A1/en active Application Filing
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107563995A (en) * | 2017-08-14 | 2018-01-09 | 华南理工大学 | A kind of confrontation network method of more arbiter error-duration models |
| CN109801292A (en) * | 2018-12-11 | 2019-05-24 | 西南交通大学 | A kind of bituminous highway crack image partition method based on generation confrontation network |
| CN109785258A (en) * | 2019-01-10 | 2019-05-21 | 华南理工大学 | A kind of facial image restorative procedure generating confrontation network based on more arbiters |
| CN110222705A (en) * | 2019-04-23 | 2019-09-10 | 华为技术有限公司 | A kind of training method and relevant apparatus of network model |
| CN110059465A (en) * | 2019-04-24 | 2019-07-26 | 腾讯科技（深圳）有限公司 | Auth method, confrontation generate training method, device and the equipment of network |
Also Published As
| Publication number | Publication date |
|---|---|
| CN111033532B (en) | 2024-04-02 |
| WO2021102697A1 (en) | 2021-06-03 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| KR20180092810A (en) | Automatic thresholds for neural network pruning and retraining | |
| US20220147877A1 (en) | System and method for automatic building of learning machines using learning machines | |
| WO2021254114A1 (en) | Method and apparatus for constructing multitask learning model, electronic device and storage medium | |
| US20210224692A1 (en) | Hyperparameter tuning method, device, and program | |
| CN106610867B (en) | On-chip network task scheduling method and device | |
| CN113467487B (en) | Path planning model training method, path planning device and electronic equipment | |
| US20230274150A1 (en) | Performing Inference And Training Using Sparse Neural Network | |
| TW202011280A (en) | Method of operating a searching framework system | |
| CN113011529B (en) | Training method, training device, training equipment and training equipment for text classification model and readable storage medium | |
| CN114492279A (en) | Parameter optimization method and system for analog integrated circuit | |
| CN112966754B (en) | Sample screening method, sample screening device and terminal equipment | |
| CN111565065B (en) | Unmanned aerial vehicle base station deployment method and device and electronic equipment | |
| CN110728359B (en) | Method, device, equipment and storage medium for searching model structure | |
| WO2023143121A1 (en) | Data processing method and related device thereof | |
| CN116779061A (en) | Interactive drug molecule design method, device, electronic equipment and medium | |
| CN115412401B (en) | Method and device for training virtual network embedding model and virtual network embedding | |
| CN111033532A (en) | Training method and system for generating countermeasure network, electronic device, and storage medium | |
| CN110874553A (en) | Recognition model training method and device | |
| CN113239034A (en) | Big data resource integration method and system based on artificial intelligence and cloud platform | |
| KR20220134627A (en) | Hardware-optimized neural architecture discovery | |
| CN114692888A (en) | System parameter processing method, device, equipment and storage medium | |
| JP2020198135A (en) | Hyper parameter tuning method, device and program | |
| JPWO2020054402A1 (en) | Neural network processing device, computer program, neural network manufacturing method, neural network data manufacturing method, neural network utilization device, and neural network miniaturization method | |
| CN113657592B (en) | Software-defined satellite self-adaptive pruning model compression method | |
| CN113780518B (en) | Network architecture optimization method, terminal equipment and computer readable storage medium |
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 |