Datasets:

DOI:

Issue trying to load dataset

#1
by velvetthunder92 - opened

Hi! I'm new to Hugging face.
After logging in with cli and getting access to the dataset,
I tried to load the data using the following snippet.

from datasets import load_dataset
hoho_dataset = load_dataset("usm3d/hoho-train-set")

The data download completes, and I then get the error
ValueError: The TAR archives of the dataset should be in WebDataset format, but the files in the archive don't share the same prefix or the same types.

Any help is appreciated! Thanks.

Urban Scene Modeling Competition CVPR 2024 (Image Track) org

Hi,

Because of the complex nature of the dataset (and possibly our unfamiliarity with HF dataset intrinsics), this way indeed doesn't work.
Instead, we recommend to use the way, provided in example notebook.

https://huggingface.co/usm3d/handcrafted_baseline_submission/blob/main/notebooks/example_on_training.ipynb

--
Best, Dmytro.

Okay, thanks for the quick response Dmytro.
Will try the other way.

Hi again!

I think the documentation is not quite clear on the steps and requirements. For example, the documentation on the example submission repo suggests that there should be a file hoho.py which does not exist and should actually be installed via the instructions at https://huggingface.co/usm3d/tools. One should also install the library webdataset.
After this, the code snippet in the example notebook to download dataset incorrectly has (possibly old) repo-id : 'usm3d/usm-training-data', which should be 'usm3d/hoho-train-set'.

Wanted to list them for anyone else having the same issues.

Urban Scene Modeling Competition CVPR 2024 (Image Track) org

Ouch! It seems that we accidentally used old code there.
I will fix that today, thank you for the catch!

Btw, this is to download a single shard. Should one itereate over all filenames like 'hoho_v3_000-of-032.tar.gz' from 000 to 032 to get the whole training data?

Urban Scene Modeling Competition CVPR 2024 (Image Track) org
β€’
edited May 22

The path is fixed and notebook updated.
To download the whole dataset you can use smth like (also in updated notebook):

# Download all the datasets:
for i in range(26):
    dataset = wds.WebDataset(hf_hub_download(repo_id='usm3d/hoho-train-set',
            filename=f'data/train/hoho_v3_{str(i).zfill(3)}-of-032.tar.gz',
            repo_type="dataset"))
Urban Scene Modeling Competition CVPR 2024 (Image Track) org
This comment has been hidden

Sign up or log in to comment