Extracting Latents

[arjunrkaushik]

Issue: Missing Step for Extracting Latents / Dataset Path Errors

After following the training instructions in the README (including the prepare dataset step), running the training script produces the following error:

AssertionError: Data path /data/kaushik3/CoresetSelection/datasets/ImageNet1k/imagenet_folder_1x256x256_sd_vae_fixed/train/
is not an S3 url and does not exist.

• Location: Line 55 in src/data/__init__.py
• Current ImageNet directory paths:
  • Training: /data/kaushik3/CoresetSelection/datasets/ImageNet1k/train
  • Validation: /data/kaushik3/CoresetSelection/datasets/ImageNet1k/val

---

Attempted Fix

Edited the data paths in ./infra/experiments/experiments-alphaflow.yaml (Lines 75-76) to:

src: ${env.folder_datasets_path}/train/
 src_val: ${env.folder_datasets_path}/val/

New Error

After correcting the path, a new error occurs:

assert len(self._data_fnames) > 0, f'No data files found in the specified path: {self._src}.'
AssertionError: No data files found in the specified path: /data/kaushik3/CoresetSelection/datasets/ImageNet1k/train/.

• Location: Line 107 in src/data/datasets.py
• Interpretation: The dataset loader did not find any image files in the specified directory

Please help with a fix.

[huijieZH]

Hi,

Could you provide a bit more information?

1. Have you successfully installed ImageNet? Do you have the dataset downloaded in the following directories?

/data/kaushik3/CoresetSelection/datasets/ImageNet1k/train
/data/kaushik3/CoresetSelection/datasets/ImageNet1k/val

2. The error is raised in /data/kaushik3/CoresetSelection/datasets/ImageNet1k/imagenet_folder_1x256x256_sd_vae_fixed/train/
   because you’re using pre-extracted latents. The code is supposed to extract SD-VAE features into this path the first time you run it. It looks like this step failed for some reason.

Could you try running the following command?

torchrun --max-restarts=0 --standalone --nproc-per-node=8 infra/launch.py alphaflow-256x256-b256-ditb2-sdvae-base-folder direct_launch=true wandb.enabled=false

This will disable the pre-extracted latents.