Dummy models for CI

[javiermtorres]

For CI purposes, we would need dummy models that return a predictable (e.g. constant) value, with as little processing as possible. These will honor the general interface (token classification, sequence classification, etc) but just return a dummy value. This allows testing integration and non-ML related changes efficiently.

Closes #98

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.

Files with missing lines	Coverage Δ
encoderfile/src/build_cli/tokenizer.rs	92.73% <100.00%> (+1.28%)	⬆️
...i/transforms/validation/sequence_classification.rs	89.61% <100.00%> (+0.42%)	⬆️
..._cli/transforms/validation/token_classification.rs	89.18% <100.00%> (ø)

... and 5 files with indirect coverage changes

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[besaleli]

nit: i'd put configs in the respective subfolders of ./models and name them all encoderfile.yml

[javiermtorres]

ah, good idea 👍
rather than in whatever_encoderfile.yaml within ./models, perhaps?

[javiermtorres]

Hmmm, currently models is in gitgnore. Maybe we could have some specific dir for encoderfiles?

[besaleli]

where is code where models weights themselves are generated?

[javiermtorres]

It's in

encoderfile/scripts/create_dummy_model.py

Line 78 in 62e0b0a

class DummySequenceClassifier(PreTrainedModel):

, which basically generates and/or tests the dummy models (usable from cmd line)
I preferred to download the weights as standard, maybe we can optionally generate them from scratch; the procedure is there, in any case.

[javiermtorres]

BTW, fun fact: since the output is dynamically generated, there are actually no weights. So the torch model exporter refuses to write anything, and the onnx exporter fails because it sees no weights. We need to include a dummy val in the state so it gets exported and everything works ok.

[javiermtorres]

Here:

encoderfile/scripts/create_dummy_model.py

Line 95 in 62e0b0a

self.register_buffer(

We could have hardcoded weights instead (or maybe hardcoded outputs rather) if you'd prefer.