Request for Code of Tiny-ImageNet

[zkcys001]

I hope this message finds you well. First and foremost, I would like to express my sincere appreciation for the incredible work you have done on your project shared on GitHub.

I am particularly interested in your Tiny-ImageNet project and would greatly appreciate it if you could share some of the code related to it.

Best regards, kecheng

[FutureXiang]

Hi Kecheng,

Thanks for your interest in this work, and I'm sorry for the delayed response.

The current state of this repo does not contain official configurations for Tiny-ImageNet training, but it supports the Tiny-ImageNet dataset and can build flexible U-Nets. Therefore, you can try to train Tiny-ImageNet models by creating a custom config file, like the following example:

# dataset params
dataset: 'tiny'
classes: 200

# other params ...

network:
  image_shape: [3, 64, 64]
  n_channels: 192
  ch_mults: [1, 2, 3, 4]
  is_attn: [False, True, True, True]
  attn_channels_per_head: 64
  dropout: 0.1
  n_blocks: 3
  use_res_for_updown: True

# other params ...

This example builds an ADM network (~300M parameters), which is similar to the one mentioned in the EDM paper (Table 7 & 8). This network should yield optimal generative & discriminative performance, but I'm not able to run it due to the large CUDA memory cost. If you have top-end GPUs you can try it :)

If you just want to reproduce the DDPM++ network used in our paper (~60M parameters), please use the config like this:

# dataset params
dataset: 'tiny'
classes: 200

# other params ...

network:
  image_shape: [3, 64, 64]
  n_channels: 128
  ch_mults: [1, 2, 2, 2]
  is_attn: [False, False, True, False]
  dropout: 0.1
  n_blocks: 4
  use_res_for_updown: True

# other params ...