about the ImageNet

[DDgogogo]

Can you provide the link of the ImageNet datasets that you have used? I use my ImageNet datasets when run the eval_toolbox/resize.p , it does not work.

[hao-pt]

Hi folk, you just need to ensure the folder containing only images, no subfolder as in original ImageNet data. Or, you can modify the code to load your images accordingly. Here, we only use the training set of 1.28M images.

[DDgogogo]

大家好，你只需要确保文件夹只包含图片，没有像原始 ImageNet 数据那样的子文件夹即可。或者，你可以修改代码来相应地加载图片。这里我们只使用了包含 1.28M 张图片的训练集。

Thanks you for your replay! I train the CelebA datasets that you provide by myself with 250 epoch, but when I run the eval.sh with my checkpoints(0000250.pt), my fid is 13. It's quite different from your 4.(I use the same setting as your checkpoint.) Do you know why?

# CelebA 256
CUDA_VISIBLE_DEVICES=0 torchrun --nnodes=1 --rdzv_endpoint 0.0.0.0:$MASTER_PORT --nproc_per_node=1 dimsum/sample_ddp.py ODE \
    --model DiM-L/2 \
    --per-proc-batch-size 64 \
    --image-size 256 \
    --ckpt checkpoints/0000250.pt \       **#My checkpoints were trained for 250 epochs, and the training settings are the same as yours.**
    --num-fid-samples 50_000 \
    --path-type GVP \
    --num-classes 1 \
    --sampling-method dopri5 \
    --num-sampling-steps 250 \
    --diffusion-form none \
    --sample-dir samples-50k \
    --block-type combined \
    --bimamba-type none \
    --eval-refdir real_samples/celeba_256 \
    --eval-metric fid50k_full,pr50k3_full \
    --rms-norm \
    --fused-add-norm \
    --learnable-pe \
    --cond-mamba \
    --use-attn-every-k-layers 4 \ 

[hao-pt]

Something's wrong here. Does your real_samples/celeba_256 contain valid GT images in .jpg format? Could you send the the training command with detailed configs? Just for sanity check.