Unconditional Cifar10 generation FID higher than reported in the paper. · Issue #12 · feizc/DiS · GitHub
Skip to content

Unconditional Cifar10 generation FID higher than reported in the paper. #12

Description

@kongwanbianjinyu

Hello @feizc ,

Thanks for the great work! I train the DiS-S/2 model for 200K steps using train.py and sample using unconditional DDPM for 250 time steps to generate 10K samples. Then I got the FID-10K=16.446, which is higher than the result in figure 2.

I follow the hyperparameters unchanged in train.py. For sampling, I follow the DiT Code and modify the sample_ddp.py and sample_ddp.sh which can be found in the sample_ddp.zip file. For FID computation, I pass both the reference cifar10-train folder and generated sample folder to test.py.

I wonder how to reproduce the results of unconditional CIFAR10 generation. Can you please help me with that? Really appreciate it.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions