Closed chinayb closed 4 months ago
Thanks a lot! I'm glad that you like our work! We provided the end-to-end model training with our code. The only step left for you is to adjust the code to fit your custom dataset. For our experiments we prepared a dataset with the Pytorch ImageFolder data format, but you can prepare your own dataset with the format that you prefer the most. The dataloader is prepared in the Trainer class that you can find in both dm.py and dm_mask.py. We forgot to update the code, classifier_free_guidance was the old name for dm_masks, thanks for noticing it! I hope this will helpful.
Thank you for your help. However, the problem arises again after loading my data?
┌───────────────────── Traceback (most recent call last) ─────────────────────┐
│ G:\download\diffinfinite-master\train.py:72 in
Process finished with exit code 1
In the latent space, the diffusion model expects 4 channels. This is because the Latent Diffusion Model is composed by a VQ-VAE which maps a Bx3xCxW --> Bx3x512x512 images into Bx4x(C//8)x(W//8) --> Bx4x64x64. We prepared the data with the VQ-VAE.encoder in the Trainer class. You are free to prepare your data in the latent wherever you want.
Congratulations on your incredibly impressive work. I'm interested in training the model on my own dataset and was wondering if you could provide any specific instructions or recommendations on how to best approach this? Is there a problem with the code "from classifier_free_guidance import Unet, GaussianDiffusion, Trainer“ in train_masks.py?