Evaluation problem

[drx-code]

Thank you for your great work! I tried to train the RQ-VAE model on the FFHQ dataset with the default config configs/ffhq/stage1/ffhq256-rqvae-8x8x4.yaml. The batch size is 8 and I used one NVIDIA RTX-3090 GPU. However, I found the evaluation result is extremely worse (running it with the eval function from trainer_vqvae.py)

While in the model.train() mode the result is much better

So how to solve that question? Thank you a lot.

[xiaoShen110141]

Have you found a solution now? I also encountered the same problem.

[xiaoShen110141]

@drx-code I think overfitting is normal when training only on FFHQ. The author seems to have mentioned in the project that when training on datasets such as FFHQ, it is necessary to use models trained on ImageNet for fine-tuning.