| Title |
Improvement of Training Stability and Controlability for GAN via Reusing Pre-trained Autoencoder and Multi-label Classifier |
| Authors |
김지수(Jisoo Kim) ; 지준(Jun Ji) ; 오희석(Heeseok Oh) |
| DOI |
https://doi.org/10.5573/ieie.2024.61.2.88 |
| Keywords |
Generative adversarial network; Training stability; Controllable generation; Transfer learning |
| Abstract |
Generative Adversarial Networks (GANs) are generative models that face the challenge of sampling high-resolution images due to the generator's hostility towards the discriminator in a pixel-wise manner, as well as the instability of the dataset distribution estimation arising from adversarial training. To generate images that satisfy certain conditions, previous studies have embedded conditions into latent vectors, which are then fed into the generator. However, due to GANs’ training instability, there is a limit to generating images having plausible quality with controlled conditions. One solution is to utilize a pre-trained feature mapper through transfer learning to increase the efficiency of learning. In this paper, we apply transfer learning to improve the fidelity of the generated image and promote training stability, ultimately generating high-resolution images that meet the combined condition and latent representation. We pre-trained the encoder-decoder for embedding an image into a low-dimensional latent space and reconstructing it into a spatial domain, learning the latent space by reducing the errors between the predictions and input labels through the classifier. The weights of the generator were initialized by the pre-trained decoder that reconstructs the image context, while an encoder that embeds an image into latent space was set as a frozen classifier that receives the generator's output. As a result, we verified that retraining the PSC-GAN (Pre-weighted Stabilized Conditioning Generative Adversarial Network) with improved training stability and controllability under the same conditions led to about 10% quantitative improvement compared to previous studies. |