Data parsing Bug for dataloader. about openfold HOT 3 CLOSED

Could you share the command you're running?

from openfold.

python scripts/data_dir_to_fasta.py mmcif_data processed_fasta/input.fasta
python3 scripts/precompute_alignments_mmseqs.py processed_fasta/input.fasta data/mmseqs_dbs uniref30_2103_db alignment_dir_single/ ./mmseqs/bin/mmseqs --hhsearch_binary_path=lib/conda/envs/openfold_venv/bin/hhsearch --env_db colabfold_envdb_202108_db --pdb70 data/pdb70/pdb70

Here I get the 5E0Y_1_A_A folder under alignment_dir_single folder, which contains the a3m hhr files.

python3 scripts/generate_mmcif_cache.py mmcif_data/ mmcif_cache.json --no_workers 16

get the mmcif_cache.json( I change the release_data):
{ "5e0y_1_A": { "release_date": "2017-03-29", "no_chains": 1 } }

Then I run
python3 train_openfold.py mmcif_data/ alignment_dir_single/ data/pdb_mmcif/mmcif_files/ output 2021-10-10 --template_release_dates_cache_path mmcif_cache.json --precision 16 --gpus 0 --seed 42

To test dataloader. And get the bug.

from openfold.

Fixed in 79f9f03. It ended up being a bug in the distillation set pipeline.

Two notes about your usage:

1. Note that --gpus 0 means "use 0 GPU's", not "use GPU 0". To limit the model to a specific GPU, use the CUDA_VISIBLE_DEVICES environment variable. If you really did intend to run the model with 0 GPUs, I can't promise that you'll get usable performance out of OpenFold.
2. .pdb files in the mmcif_dir are interpreted as examples from the distillation set and are treated slightly differently from ground truth examples. You'll want to get rid of the one in the directory you sent if this is a real protein. This behavior wasn't at all clear in the previous README, so I added a paragraph about that too.

from openfold.