token type ids can be set by optional argument up to python wrapper

[hachall]

Description

This PR adds an optional token_type_ids argument to Encoder.translate_batch, to facilitate BERT models with multiple input features (sentence-sentence entailment, context and question answering etc.). The token_type_ids vector is optional at the pybind level, and overloaded into the future_batch_async functions, with all previous endpoints preserved for backwards compatibility.

Closes #1383.

...
output = ct2_model.forward_batch(input_ids, token_type_ids=token_type_ids)
pooler_output = output.pooler_output
pooler_output = np.array(pooler_output)
pooler_output = torch.as_tensor(pooler_output)
ct_logits = classifier(pooler_output)

Type of change

Please delete options that are not relevant.

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Docs update
• Refactor

Testing

The code passes all C++ and python tests. For validation, the modified encoder was tested with both a bert-base-uncased and fine tuned bert-tiny for inference on the train split of the MRPC dataset (sentence pairs), and all output logits verified equal (<10e-12) to the output of a hugging face loaded model at full precision, when passing token_type_ids to both. When quantized, the logits are no longer equal, but the token_type_ids demonstrably improve the classifier accuracy.

Checklist:

• My code follows the style guidelines of this project
• I have made corresponding changes to the documentation
• I have added tests that prove my fix is effective or that my feature works

[hachall]

#1383

[hachall]

Thank you for the notes @guillaumekln! I have implemented the changes, let me know if there's anything else I can do