Automatically annotated layout data generation algorithm based on DoubleGAN

Abstract

The present invention relates to the field of data generation statistics and discloses an automatic annotated map data generation algorithm based on DoubleGAN. The algorithm comprises a prediction model, which includes a generative model training and data expansion process. The generative model training includes S1: building a feature map generation model G1; S2: building a feature map discrimination model D1; S3: building an automatic annotation model G2; S4: building a generated label map discrimination model D2; S5: defining a loss function; S6: using an RMS optimizer to calculate the exponentially weighted moving average of the squared gradient; and S7: adversarial training. The algorithm uses U-net to build the automatic annotation model and uses Wasserstein and RMS optimizers to perform adversarial training on the DoubleGAN model. The algorithm automatically annotates randomly generated data, obtaining a large number of high-quality data samples and improving the generalization of downstream prediction models. The DoubleGAN algorithm also uses two GAN models with different structures to enhance existing data.

Description

Technical Field

[0001] The present invention relates to the field of data generation statistics technology, and specifically to an automatic annotated layout data generation algorithm based on DoubleGAN. Background Art

[0002] As deep learning theory matures, artificial intelligence technology has been widely used in many fields, and AI for EDA has also emerged, such as the layout congestion prediction model based on the fully convolutional neural network, the DRC prediction model based on RouteNet, and the voltage drop prediction model based on U-net. The above deep learning models can efficiently extract various verification information based on the complex physical information of the layout circuit to achieve the purpose of early prediction, thereby assisting the development process and improving R&D efficiency. Although the above deep learning models have achieved excellent performance in EDA-assisted design, due to the confidential nature of IC design layouts, public layout data is very scarce and training data is limited, which makes such models face problems such as unstable training and low model generalization. Data enhancement methods are very necessary.

[0003] The existing data augmentation scheme uses a generative adversarial network (GAN) to train a data generator. GAN consists of a generator network and a discriminator network D. The generator network is used to generate data, and the discriminator network is used to judge the authenticity of the data. Its principle is to let the two continuously train against each other, so that the data generated by the generator network is highly close to the real data. This method has been applied in image fields such as the face dataset WIDER-FACE, the SAR ship dataset, and the emotion recognition dataset FER2013. It can generate high-quality images and improve the generalization of the model.

[0004] In the aforementioned dataset augmentation applications, GAN-based augmentation methods can only generate data without label information, so manual annotation of the generated data is usually required to obtain its label information. For discrete label information, such as the labels for cat and dog image classification, manual annotation is still feasible. However, for continuous label information, such as the labels of pathological and normal areas in medical images, manual annotation is not only time-consuming and labor-intensive, but also prone to errors. In view of this, we propose an algorithm for automatically generating annotated map data based on DoubleGAN. Summary of the Invention

[0005] The purpose of the present invention is to provide an automatic annotation layout data generation algorithm based on DoubleGAN to solve the problems raised in the above background technology.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an automatic annotated layout data generation algorithm based on DoubleGAN, including a prediction model, wherein the prediction model includes a generative model training and a data expansion process. By cooperating with the generative model training and the data expansion process, the training effect of the prediction model is improved, and the existing data can be effectively enhanced. The generative model training includes the following steps:

[0007] S1. Build the feature map generation model G1 and design a 5-layer TranConv module. First, perform deconvolution and ReLU activation and normalization. Finally, use a tanh activation layer to reduce the 1×1 random variable to a size of 256×256. This helps restore the feature information of the real data as much as possible, providing a data accuracy foundation for subsequent data transmission and recognition definition, thereby ensuring the efficiency of the training model.

[0008] S2. Build a feature map discrimination model D1, using the generated feature map or the real feature map as the input variable, obtain a 4-layer ConvReLU module, and perform convolution + LeakyReLU function activation and BN processing. Finally, output the probability value through the fully connected layer, effectively ensuring that the feature map will output the optimal feature map generation model in the continuous convergence detection process;

[0009] S3. Build the automatic labeling model G2. The input feature map passes through the ConvPool module in sequence, is activated by convolution + pooling + LeakyReLU function and batch normalization, and then alternately passes through the TranConv module and the Conv module to restore the size. Finally, the label map is output through the Sigmoid activation layer.

[0010] S4. Build the generated label image discrimination model D2, which takes the generated label image or the real label image as the input variable, passes it through a 4-layer ConvReLU module, uses convolution + LeakyReLU function activation and BN processing, and finally outputs the probability value through a fully connected layer;

[0011] S5. Define the loss function, let f w is the neural network function, V is the loss function, E is the Wasserstein expected value, P r is a set of real data, P f is a set of generated data, x is real data, z is generated data, and the loss function can be used to accurately calculate the accuracy of the data form and the distortion during transmission, so as to better help the model integrate and identify the data, thereby improving the working efficiency of the model.

[0012] The designed loss function is as follows:

[0013]

[0014] S6. Using the RMS optimizer, calculate the exponentially weighted moving average of the square of the gradient:

[0015] s t =βs t-1 +(1-β)g t 2 ;

[0016] Update parameters:

[0017]

[0018] Among them, g t is the gradient at time step t, s t is the moving average of the cumulative squared gradient up to time step t, β is the coefficient of the moving average, α is the learning rate, and ε is a small constant added to prevent the denominator from being zero. The Wasserstein and RMS optimizers are used to perform adversarial training on the DoubleGAN model, automatically annotate randomly generated data, and obtain a large number of high-quality data samples, thereby improving the generalization of downstream prediction models. In addition, by using the Wasserstein distance as the loss function, the possibility of gradient dissipation in GAN model training is reduced, making the training more stable and the generated data quality higher.

[0019] S7. Adversarial training is performed until the model reaches convergence conditions, and the model file is saved. Through continuous adversarial training, the accuracy of the model at output is improved, thereby providing better data support for subsequent training models.

[0020] Preferably, the coefficient of β in S6 is 1.

[0021] Preferably, the data expansion process includes the following steps:

[0022] S11, load the feature map generation model G1, and randomly generate several feature map data;

[0023] S12, load the automatic annotation model G2, input the above-mentioned several feature map data into the model, obtain the corresponding label map, and form a generated data set;

[0024] S13. Merge the real dataset and the generated dataset as the training set for the congestion prediction model.

[0025] Preferably, the real data set in S13 includes a real label map and a real feature map, wherein the real label map and the real feature map are collected from G1 and G2. At the same time, the generated data set includes a generated label map and a generated feature map, wherein the generated label map and the generated feature map are obtained by a feature map generation model. Two GAN models with different structures are used to enhance the existing data, which can randomly generate a large amount of high-quality data while completing automatic labeling.

[0026] Preferably, the expanded size of the training set of the congestion prediction model in S13 is twice the size of the original training set.

[0027] Preferably, the data format of the data set generated in S12 is circuit physical information extracted based on IC layout information, specifically including pin density and macro layout area feature map, each feature map has a corresponding label map.

[0028] Preferably, the feature map discriminant model D1 in S2 will be fed back to the feature map discriminant model D1 after obtaining the data of the model loss value, and the data of the model loss value will be synchronously transmitted to the feature map generation model G1.

[0029] Preferably, the generated label graph discrimination model D2 in S4 will be fed back to the generated label graph discrimination model D2 after obtaining the data of the model loss value, and the data of the model loss value will be synchronously transmitted to the automatic labeling model G2.

[0030] Preferably, the feature map generation model G1 obtains the feature map generation model inputted by the data expansion process after reaching the convergence condition, and the automatic annotation model G2 obtains the automatic annotation model inputted by the data expansion process after reaching the convergence condition.

[0031] Preferably, the automatic labeling model G2 in S3 adopts a U-net structure. By using the U-net structure to build an automatic labeling model, its cross-layer splicing technology enables the extraction of high-dimensional latent features, which is conducive to generating latent features of model learning data and making the labeling results closer to the true value.

[0032] Compared with the existing technology, the present invention provides an automatic annotation layout data generation algorithm based on DoubleGAN, which has the following beneficial effects:

[0033] 1. This DoubleGAN-based automatic annotation layout data generation algorithm uses U-net to build an automatic annotation model, adopts Wasserstein and RMS optimizers to conduct adversarial training on the DoubleGAN model, and automatically annotates randomly generated data to obtain a large number of high-quality data samples, thereby improving the generalization of downstream prediction models. In addition, the DoubleGAN algorithm uses two GAN models with different structures to enhance the existing data, which can randomly generate a large amount of high-quality data while completing the automatic annotation work.

[0034] 2. This DoubleGAN-based automatic annotation layout data generation algorithm uses Wasserstein distance as the loss function, which reduces the possibility of gradient dissipation in GAN model training, making training more stable and generating higher quality data.

[0035] 3. This DoubleGAN-based automatic annotation layout data generation algorithm uses the U-net structure to build an automatic annotation model. Its cross-layer splicing technology enables the extraction of high-dimensional latent features, which is conducive to generating the latent features of the model learning data and making the annotation results closer to the real value.

[0036] 4. This DoubleGAN-based automatic annotation layout data generation algorithm can effectively improve the performance of downstream task prediction models by expanding existing IC layout data using the DoubleGAN algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 Schematic diagram of the system framework of the present invention;

[0038] Figure 2 A schematic diagram of a model structure for generating a characteristic graph of the present invention;

[0039] Figure 3 Schematic diagram of the structure of the feature graph discrimination model of the present invention;

[0040] Figure 4 Automatically mark the model structure schematic diagram of the present invention;

[0041] Figure 5 Generate a schematic diagram of the label graph discrimination model structure for the present invention;

[0042] Figure 6 Schematic diagram comparing the generation effects of different loss functions of the present invention;

[0043] Figure 7 This is a curve diagram of the expansion algorithms of the present invention changing with the expansion multiple. DETAILED DESCRIPTION

[0044] like Figure 1-7 As shown, the present invention provides a technical solution: an automatic annotation layout data generation algorithm based on DoubleGAN, including a prediction model, which includes a generation model training and a data expansion process. Through the mutual cooperation of the generation model training and the data expansion process, the training effect of the prediction model is improved, and the existing data can be effectively enhanced. The generation model training includes the following steps:

[0045] S1. Build the feature map generation model G1 and design a 5-layer TranConv module. First, perform deconvolution and ReLU activation and normalization. Finally, use a tanh activation layer to reduce the 1×1 random variable to a size of 256×256. This helps restore the feature information of the real data as much as possible, providing a data accuracy foundation for subsequent data transmission and recognition definition, thereby ensuring the efficiency of the training model.

[0046] S2. Build a feature map discrimination model D1, using the generated feature map or the real feature map as the input variable, obtain a 4-layer ConvReLU module, and perform convolution + LeakyReLU function activation and BN processing. Finally, output the probability value through the fully connected layer, effectively ensuring that the feature map will output the optimal feature map generation model in the continuous convergence detection process;

[0047] S3. Build the automatic labeling model G2. The input feature map passes through the ConvPool module in sequence, is activated by convolution + pooling + LeakyReLU function and batch normalization, and then alternately passes through the TranConv module and the Conv module to restore the size. Finally, the label map is output through the Sigmoid activation layer.

[0048] S4. Build the generated label image discrimination model D2, which takes the generated label image or the real label image as the input variable, passes it through a 4-layer ConvReLU module, uses convolution + LeakyReLU function activation and BN processing, and finally outputs the probability value through a fully connected layer;

[0049] S5. Define the loss function, let f w is the neural network function, V is the loss function, E is the Wasserstein expected value, P r is a set of real data, P f is a set of generated data, x is real data, z is generated data, and the loss function can be used to accurately calculate the accuracy of the data form and the distortion during transmission, so as to better help the model integrate and identify the data, thereby improving the working efficiency of the model.

[0050] The designed loss function is as follows:

[0051]

[0052] S6. Using the RMS optimizer, calculate the exponentially weighted moving average of the square of the gradient:

[0053] s t =βs t-1 +(1-β)g t 2 ;

[0054] Update parameters:

[0055]

[0056] Among them, g t is the gradient at time step t, s t is the moving average of the cumulative squared gradient up to time step t, β is the coefficient of the moving average, α is the learning rate, and ε is a small constant added to prevent the denominator from being zero. The Wasserstein and RMS optimizers are used to perform adversarial training on the DoubleGAN model, automatically annotate randomly generated data, and obtain a large number of high-quality data samples, thereby improving the generalization of downstream prediction models. In addition, by using the Wasserstein distance as the loss function, the possibility of gradient dissipation in GAN model training is reduced, making the training more stable and the generated data quality higher.

[0057] S7. Adversarial training is performed until the model reaches convergence conditions, and the model file is saved. Through continuous adversarial training, the accuracy of the model at output is improved, thereby providing better data support for subsequent training models.

[0058] In one embodiment of the present invention, the coefficient of β in S6 is 1.

[0059] In addition, the data augmentation process includes the following steps:

[0060] S11, load the feature map generation model G1, and randomly generate several feature map data;

[0061] S12, load the automatic annotation model G2, input the above-mentioned several feature map data into the model, obtain the corresponding label map, and form a generated data set;

[0062] S13. Merge the real dataset and the generated dataset as the training set for the congestion prediction model.

[0063] In an embodiment of the present invention, the real data set in S13 includes a real label map and a real feature map, wherein the real label map and the real feature map are collected from G1 and G2. At the same time, the generated data set includes a generated label map and a generated feature map, wherein the generated label map and the generated feature map are obtained from the feature map generation model, and two GAN models with different structures are used to enhance the existing data, which can randomly generate a large amount of high-quality data while completing the automatic labeling work. Among them, the expanded size of the training set of the congestion prediction model in S13 is twice the original training set. In addition, the data form of the data set generated in S12 is the circuit physical information extracted based on the IC layout information, specifically, including pin density, macro layout area feature map, and each feature map has a corresponding label map. Specifically, the feature map discrimination model D1 in S2 will feed back the model loss value data to the feature map discrimination model D1 after obtaining the model loss value data, and the model loss value data will be synchronously transmitted to the feature map generation model G1.

[0064] In an embodiment of the present invention, the generated label map discriminant model D2 in S4 will be fed back to the generated label map discriminant model D2 after obtaining the data of the model loss value, and the data of the model loss value will be synchronously transmitted to the automatic annotation model G2. Furthermore, the feature map generation model G1 obtains the feature map generation model inputted by the data expansion process after reaching the convergence condition, and the automatic annotation model G2 obtains the automatic annotation model inputted by the data expansion process after reaching the convergence condition. At the same time, the automatic annotation model G2 in S3 adopts the U-net structure, and by using the U-net structure to build the automatic annotation model, its cross-layer splicing technology enables the potential features in high dimensions to be extracted, which is beneficial to the potential features of the generated model learning data, so that the annotation results are closer to the true values.

[0065] In an embodiment of the present invention, the algorithm test experiment of the present invention is completed on a personal PC with GPU: NVIDIA RTX 3080Ti12GB; CPU: Intel CoreTMi7-10700@2.90GHz x16; memory: 32GB; operating system: Ubuntu20.04; programming language: Python3.10.4; Pytorch1.13.1. The experiment uses the CircuitNet dataset to evaluate the improved performance of the present invention. First, the cross entropy loss function, least squares loss function, and Wassersetin distance loss function are used to train the same number of times to generate the FID value of the feature map and the corresponding generated label map, and after expanding the training set to twice the original, the experimental data of the prediction model performance index are tested. Then, the traditional expansion methods such as rotation and flipping, the expansion method based on DCGAN, and the DoubleGAN expansion method proposed in this article are explored. The impact of different expansion multiples on the performance of the congestion prediction model is used, and the performance index of the congestion prediction model before expansion is used as the baseline. Please refer to the table below for details.

[0066]

[0067] Combined with the above table and the instructions Figure 6 The data generated using the cross entropy loss function is quite different from the real data, and DoubleGAN basically cannot learn the distribution of the real data; the data generated using the least squares loss function initially has the prototype of the real data, but there is still a certain gap with the real data; the data generated using the Wasserstein distance loss function is very similar to the real data in terms of morphology and structure, indicating that DoubleGAN has learned the distribution of the real data and has the best generation effect.

[0068] Please continue to refer to the instructions attached Figure 7 It can be seen that this data model is best when expanded by 2 times, which improves the accuracy of the prediction model, and the improvement is greater than the traditional expansion and DCGAN-based expansion methods, which plays a positive reference role in IC layout data expansion methods.

[0069] In summary, by using U-net to build an automatic labeling model, using Wasserstein and RMS optimizers to perform adversarial training on the DoubleGAN model, and automatically labeling randomly generated data, a large number of high-quality data samples are obtained, thereby improving the generalization of downstream prediction models.

[0070] The above generally describes the present invention in detail. However, it is obvious to those skilled in the art that modifications or improvements may be made based on the present invention. Therefore, modifications or improvements that do not depart from the spirit of the present invention are within the scope of protection of the present invention.

Claims

1. An automatic annotated layout data generation algorithm based on DoubleGAN, characterized by: The prediction model includes a generative model training and data expansion process, and the generative model training includes the following steps: S1. Build the feature map generation model G1 and design a 5-layer TranConv module. First, perform deconvolution + ReLU activation and normalization. Finally, use a tanh activation layer to reduce the 1×1 random variable to a size of 256×256 to restore the feature information of the real data as much as possible. S2. Build a feature map discrimination model D1, using the generated feature map or the real feature map as the input variable, obtain a 4-layer ConvReLU module, perform convolution + LeakyReLU function activation and BN processing, and finally output the probability value through the fully connected layer; S3. Build the automatic labeling model G2. The input feature map passes through the ConvPool module in sequence, is activated by convolution + pooling + LeakyReLU function and batch normalization, and then alternately passes through the TranConv module and the Conv module to restore the size. Finally, the label map is output through the Sigmoid activation layer. S4. Build the generated label image discrimination model D2, which takes the generated label image or the real label image as the input variable, passes it through a 4-layer ConvReLU module, uses convolution + LeakyReLU function activation and BN processing, and finally outputs the probability value through a fully connected layer; S5. Define the loss function, let f w is the neural network function, V is the loss function, E is the Wasserstein expected value, P r is a set of real data, P f is the set of generated data, x is the real data, z is the generated data, The designed loss function is as follows: S6. Using the RMS optimizer, calculate the exponentially weighted moving average of the square of the gradient: s t =βs t-1 +(1-β)g t 2 ; Update parameters: Among them, g t is the gradient at time step t, s t is the moving average of the cumulative squared gradient up to time step t, while β is the coefficient of the moving average, α is the learning rate, and ε is a small constant added to prevent the denominator from being zero; S7, adversarial training, until the model reaches convergence conditions, save the model file; The feature map discriminant model D1 in S2 will be fed back to the feature map discriminant model D1 after obtaining the data of the model loss value, and the data of the model loss value will be synchronously transmitted to the feature map generation model G1. The generated label map discriminant model D2 in S4 will be fed back to the generated label map discriminant model D2 after obtaining the data of the model loss value, and the data of the model loss value will be synchronously transmitted to the automatic labeling model G2.

2. The DoubleGAN-based automatic annotation layout data generation algorithm according to claim 1, characterized in that: The coefficient of β in S6 is 1.

3. The DoubleGAN-based automatic annotation layout data generation algorithm according to claim 1, characterized in that: The data expansion process includes the following steps: S11, load the feature map generation model G1, and randomly generate several feature map data; S12, load the automatic annotation model G2, input the above-mentioned several feature map data into the model, obtain the corresponding label map, and form a generated data set; S13. Merge the real dataset and the generated dataset as the training set for the congestion prediction model.

4. The DoubleGAN-based automatic annotation layout data generation algorithm according to claim 3, characterized in that: The real data set in S13 includes a real label map and a real feature map, wherein the real label map and the real feature map are collected from G1 and G2. At the same time, the generated data set includes a generated label map and a generated feature map, wherein the generated label map and the generated feature map are obtained by a feature map generation model.

5. The DoubleGAN-based automatic annotation layout data generation algorithm according to claim 3, characterized in that: The expanded size of the training set of the congestion prediction model in S13 is twice the size of the original training set.

6. The DoubleGAN-based automatic annotation layout data generation algorithm according to claim 3, characterized in that: The data set generated in S12 is in the form of circuit physical information extracted based on IC layout information, specifically including pin density and macro layout area feature maps, each feature map having a corresponding label map.

7. The DoubleGAN-based automatic annotation layout data generation algorithm according to claim 1, characterized in that: The feature map generation model G1 obtains the feature map generation model inputted by the data expansion process after reaching the convergence condition, and the automatic annotation model G2 obtains the automatic annotation model inputted by the data expansion process after reaching the convergence condition.

8. The DoubleGAN-based automatic annotation layout data generation algorithm according to claim 1, characterized in that: The automatic labeling model G2 in S3 adopts a U-net structure.

Images