Big Refactor

[tcapelle]

Hey, this code has made me play and have so much fun =)
I refactored the code to do multiple things:

• It uses mixed precision training (2x speed up).
• Refactor the modules to make them more readable and not repeat the code
• For the moment only refactores the conditional model with argparse and defaults. Added a bunch of params to the script argparse.
• Logging to Weights and Biases so you get intermediate results and can follow training. For instance, this is my recent Cifar10 run
• Added scripts to download the dataset, Cifar10, Cifar100 and Cifar10_64px from Kaggle.
• Refactored the training script to make it modular, also added validation metrics if available.
• Added checkpoint loading.

Please take a look, and feel free to contact me.

• Added FFCV dataloading option, this makes training dramatically faster.

[dome272]

Hey this is an awesome addition and thank you so much for the work. The only problem I have is that people might be confused who are coming from the video and suddenly see so much new code. As a result I probably will push your code to a new branch. Im very keen to try all of this out!

[tcapelle]

I know, but this is very cool as it produces a nice wandb.workspace, you can see the logging of the run here:
https://wandb.ai/capecape/train_sd?workspace=user-capecape

[tcapelle]

I am writing an article based on your code and video: https://wandb.ai/capecape/train_sd/reports/Training-a-Conditional-Diffusion-model-from-scratch--VmlldzoyODMxNjE3

[dome272]

Very very cool. Thank you!! I added a reference to your repo and blog in the readme

[noushineftekhari]

I am writing an article based on your code and video: https://wandb.ai/capecape/train_sd/reports/Training-a-Conditional-Diffusion-model-from-scratch--VmlldzoyODMxNjE3

Hello, Can I ask why we have three.PT files after training? Which one is for test and did you used diffusion for feature representation?

[tcapelle]

https://wandb.ai/capecape/train_sd/reports/How-To-Train-a-Conditional-Diffusion-Model-From-Scratch--VmlldzoyNzIzNTQ1

[noushineftekhari]

Hi Thank you for the link. I have another question. I run the model and I have the following weight: ***@***.*** No I am trying to test the model I used the below script: n = 10 device = "cpu" model = UNet_conditional(num_classes=4).to(device) # ckpt = torch.load(r"models\DDPM_conditional\ckpt.pt") file_path = r"C:\Users\noueft\Downloads\256x256_classifier.pt" if os.path.exists(file_path): ckpt = torch.load(file_path) else: print("The file does not exist at the specified path.") model.load_state_dict(ckpt) diffusion = Diffusion(img_size=64, device=device) y = torch.Tensor([6] * n).long().to(device) x = diffusion.sample(model, n, y, cfg_scale=3) plot_images(x) But the following error : n = len(labels) TypeError: object of type 'int' has no len() I have the classes in separate folder: ***@***.*** Can you help me with it? From: Thomas Capelle ***@***.***> Sent: Monday, October 23, 2023 8:14 AM To: dome272/Diffusion-Models-pytorch ***@***.***> Cc: Eftekhari, Noushin ***@***.***>; Comment ***@***.***> Subject: Re: [dome272/Diffusion-Models-pytorch] Big Refactor (PR #1) CAUTION: External email. Ensure this message is from a trusted source before clicking links/attachments. https://wandb.ai/capecape/train_sd/reports/How-To-Train-a-Conditional-Diffusion-Model-From-Scratch--VmlldzoyNzIzNTQ1 - Reply to this email directly, view it on GitHub<#1 (comment)>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/AOJFARL7TAF2WBTXKBRCBWLYAYKMTAVCNFSM6AAAAAAQ2LFQHWVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTONZUGU3DMMRWGI>. You are receiving this because you commented.Message ID: ***@***.***>