Huggingface multi-modal transformers

Description

This PR introduces a new HuggingfaceEmbedding component, which wraps Huggingface models (such as LayoutLM or LILT). Compared with using the raw huggingface model, this wrapper offers a simple mechanism for splitting long documents into sliding windows before sending them to the model (since the maximum number of tokens sent to the transformer is capped at 512, and sending entire sequences all at once can be memory-intensive).

Example

Here is an example of how to define a pipeline with the HuggingfaceEmbedding component:

from edspdf import Pipeline

pipeline = Pipeline()
pipeline.add_pipe("pdfminer-extractor", name="extractor")
pipeline.add_pipe(
    "huggingface-embedding",
    name="embedding",
    config={
        "model": "microsoft/layoutlmv3-base",
        "use_image": False,
        "window": 128,
        "stride": 64,
        "line_pooling": "mean",
    },
)
model.add_pipe(
    "trainable-classifier",
    name="classifier",
    config={
        "embedding": model.get_pipe("embedding"),
        "labels": [],
    },
)

This model can then be trained following the training recipe.

[x] If this PR is a bug fix, the bug is documented in the test suite.
[ ] Changes were documented in the changelog (pending section).
[x] If necessary, changes were made to the documentation.

aphp / edspdf

Huggingface multi-modal transformers #15

Description

Example

Codecov Report