Fully initialize used Libraries in the Dev Container

[JulianFrattini]

Starting up the API container for the first time always triggers a download that is most likely connected to some of the connected libraries (possibly parts of the BERT model). Because this download only happens once, it does not slow down the continuous usage of that container significantly. But when frequently rebuilding the container (e.g., when performing GitHub actions for testing), this becomes an issue. Hence, identify how to move this download from the API container into the dev container.

[andreas-bauer]

This is a good idea.

I will test if the following command as one RUN command in the Dockerfile would solve the problem.

python -c 'from transformers import BertModel; BertModel.from_pretrained("bert-base-cased")'

[andreas-bauer]

Minor update on the matter.

Pretrained transformers will be cached after the first download.
If not further specified, data will be cached in ~/.cache/huggingface/transformers, or in the case of the Dockerfile.dev image in /root/.cache/huggingface/transformers.
The env HF_HOME allows users to change the caching dir but does not resolve any of the problems I encountered.

Attempt 1: Caching
A simple idea would be to run the following code during Docker image creation to cache all transformers in the image.

ENV HF_HOME="model/hf_cache_home"
RUN python -c 'from transformers import BertModel; BertModel.from_pretrained("bert-base-cased")' \
    && python -c 'from transformers import RobertaTokenizerFast; RobertaTokenizerFast.from_pretrained("roberta-base")'

While the transformers are present in the cache dir, running pytest within the container cause a new download of the files.
I tried to force a local usage of files using the local_files_only parameter, but then the test that depends on the transformers will fail.

BertTokenizer.from_pretrained(PRE_TRAINED_MODEL_NAME, local_files_only=True)

Specifying the cache dir with the argument cache_dir="..." also does not work.
I checked the permissions for the directory/files and changed them to 777 for a test. It also is not a permission problem.

Attempt 2: specify location

Another attempt of mine was to specify the location of a stored transformer.
The from_pretrained() function accepts a transformer name or a file location as an argument.

model = BertTokenizer.from_pretrained("bert-base-cased")
model.save_pretrained("model/hf/bert-base-cased")

# downloads the transformer from huggingface
BertTokenizer.from_pretrained("bert-base-cased")

# uses an already downloaded version
BertTokenizer.from_pretrained("model/hf/bert-base-cased")

model_locator.BERT could then be used to either return a path to the local files or to return the name to download it.
But this also does not work as expected. It will throw an exception that the transformer (or parts) are missing.