-
Grigory Sizov authored
Summary: Closes T136364380 Added [WavLM Model](https://github.com/microsoft/UniSpeech/tree/main/WavLM): - Added `WavLMSelfAttention` class (from [original implementation](https://github.com/microsoft/UniSpeech/blob/2e9dde8bf815a5f5fd958e3435e5641f59f96928/WavLM/modules.py)) and adjusted existing Encoder and Transformer classes to be compatible with it - Added factory functions `wavlm_model`, `wavlm_base`, `wavlm_large` to `models/wav2vec2/model.py` - Added bundles for base and large models to pipelines. **TODO**: pre-trained model weights are not yet uploaded to `download.pytorch.org`, permissions not granted yet. ## Tests - Expanded HuggingFace integration tests to cover WavLM. For there tests, added JSON configs for base and large models from HF ([base](https://huggingface.co/microsoft/wavlm-base/blob/main/config.json), [large](https://huggingface.co/microsoft/wavlm-large/blob/main/config.json)) into test assets - Expanded TorchScript and quantization tests to cover WavLM ## Comments There are a few workarounds I had to introduce: - Quantization tests for WavLM were breaking down at [`torch.cat`](https://github.com/pytorch/audio/pull/2822/files#diff-6f1486901c94320ec0610a460dc674638fab9d104a61564ff7b59353a8b8547cR466) ~~until I excluded the arguments of `torch.cat` from quantization [here](https://github.com/pytorch/audio/pull/2822/files#diff-6f1486901c94320ec0610a460dc674638fab9d104a61564ff7b59353a8b8547cR368-R369). I haven't found a better way to fix it, let me know if there is one~~ The reason for this seems to be that quantization replaces `.bias` and `.weight` attributes of a `Linear` module with methods. Since we are using weights and biases directly, the code was break. The final solution suggested by nateanl was to define attention weights and biases directly in `WavLMSelfAttention`, skipping the `Linear` layers - ~~WavLM uses position embedding in the first layer of encoder, but not in the subsequent ones. So [UniSpeech](https://github.com/microsoft/UniSpeech/blob/2e9dde8bf815a5f5fd958e3435e5641f59f96928/WavLM/modules.py#L342) and [HF](https://github.com/huggingface/transformers/blob/b047472650cba259621549ac27b18fd2066ce18e/src/transformers/models/wavlm/modeling_wavlm.py#L441-L442) implementations only create this embedding module in the layers where it's used. However, we can't do this here because it breaks TorchScript. So as a solution I add a dummy `Identity` module to `WavLMSelfAttention` when the actual embedding is not needed: [here](https://github.com/pytorch/audio/pull/2822/files#diff-6f1486901c94320ec0610a460dc674638fab9d104a61564ff7b59353a8b8547cR361-R368).~~ Thanks nateanl for resolving this! - I had to add dummy `position_bias` and `key_padding_mask` arguments to `SelfAttention.forward` to make TorchScript tests pass. Since both `SelfAttention` and `WavLMSelfAttention` are called from `EncoderLayer`, they need to have compatible signatures. Having a variable number of arguments with `**kwargs` or checking object class doesn't seem to work with TorchScript, so I instead made both types of attention accept `position_bias` and `key_padding_mask` arguments. Nit: do we still need to specify `__all__` if there are no wildcard imports in `__init__.py`, e.g. in `torchaudio/models/__init__.py`? Pull Request resolved: https://github.com/pytorch/audio/pull/2822 Reviewed By: nateanl Differential Revision: D41121855 Pulled By: sgrigory fbshipit-source-id: 9f4f787e5810010de4e74cb704063a26c66767d7
bd76d3d7
