search

facebookresearch / mmf

A modular framework for vision & language multimodal research from Facebook AI Research (FAIR)
https://mmf.sh/
Other
5.44k stars 925 forks source link

TextVQA Checksum - doesn't match and varies between models #1308

Closed ab3308 closed 1 year ago

ab3308 commented 1 year ago

When training with TextVQA I have experienced different checksums when training with different models. Also, when training on M4C I observed checksum error.

M4C - TextVQA:

Command line:

CUDA_VISIBLE_DEVICES=0 mmf_run config=projects/m4c/configs/textvqa/defaults.yaml model=m4c dataset=textvqa run_type=train checkpoint.resume_pretrained=True checkpoint.resume_zoo=m4c.textvqa.alone

Logs:

AssertionError: Checksum from dl.fbaipublicfiles.com/mmf/data/datasets/textvqa/defaults/features.tar.gz does not match the expected checksum. Please try again.

LoRRA - TextVQA

Command line:

CUDA_VISIBLE_DEVICES=0 mmf_run config=projects/lorra/configs/textvqa/defaults.yaml model=lorra dataset=textvqa run_type=train

Logs:

Checksum changed for <file_path>. Redownloading.
Downloading https://dl.fbaipublicfiles.com/mmf/data/datasets/textvqa/defaults/features.tar.gz to <file_path>

I am unsure as to whether the downloaded files vary, and whether or not simply bypassing the checksum would be a viable option here? Both are of the same size and the download process completed.

Any help would be much appreciated.

ab3308 commented 1 year ago

After running train with LoRRA, then reattempting M4C train, dataset seemed to be accepted. Resolved.