How long does it take you to train a model from scratch?

[Bob2000]

I follow the Generative Adversarial Networks paper and their codebase, I tried to train it from scratch with 5,000 images but it took so long to reduce the loss. I trained for 5000 epoch now but the loss still high. Does anyone face the same issue when training something from scratch?

 

[sonkaranjali]

Training a GAN with only about 5,000 images can be tricky, and it's common to see challenges. GANs are sensitive to things like the quality of the data, the learning rate, and the model architecture. Just because the loss is high doesn’t mean the model isn’t learning; GANs don’t follow the same rules as regular supervised models.

You might want to try using smaller batch sizes or different techniques like spectral normalization or gradient penalty. Switching to more stable models like DCGAN or WGAN-GP can also help. Also, make sure your generator and discriminator are balanced if one is learning too fast, it can throw off the training.

 

[Bob2000]

Training a GAN with only about 5,000 images can be tricky, and it's common to see challenges. GANs are sensitive to things like the quality of the data, the learning rate, and the model architecture. Just because the loss is high doesn’t mean the model isn’t learning; GANs don’t follow the same rules as regular supervised models.

You might want to try using smaller batch sizes or different techniques like spectral normalization or gradient penalty. Switching to more stable models like DCGAN or WGAN-GP can also help. Also, make sure your generator and discriminator are balanced if one is learning too fast, it can throw off the training.

Thank you for your comment, I double checked and it was in fact still learning but very slowly. I'm trying to increase the dataset size now to see if it improves