ankitapasad / s3adapter

s3adapter

Setting up

1. directly clone this repo
2. cd ./s3adapter/s3prl

pip install -e .

4. pip install necessary packages

pip install huggingface-hub
pip install editdistance
pip install loralib
pip install tensorboardX

5. You should be able to run s3prl with adapter version!

Starting adapter-based tuning

After modifying the three files, you can start adapter-based tuning with the following command line:

run_downstream.py --adapter houlsby -u hubert -d asr -m train -f -n hubert_asr_houlsby 

make sure the last sys.argv contains the name of adapter, and -f is set to True

The supported adapters includes

• adapter for AdapterBias (https://arxiv.org/abs/2205.00305)
• houlsby for Houlsby adapters (https://arxiv.org/abs/1902.00751)
• bitfit for Bitfit method (https://arxiv.org/abs/2106.10199)
• lora for LoRA (https://github.com/microsoft/LoRA)
• cnn for CNN adapters (One can add cnn adapters & houlsby adapters at the same time by specifying name in sys.argv[-1] )

for example:

run_downstream.py --adapter houlsby -u hubert -d asr -m train -f -n hubert_asr_houlsby_cnn

means using houlsby and cnn adapters at the same time.

Simply tune the whole upstream model using the last hidden representation:

run_downstream.py  -u hubert -d asr -m train -f -l -1 -n hubert_asr_finetune

modified files

1. s3prl/s3prl/run_downstream.py (add adapter option parser)
2. s3prl/s3prl/upstream/wav2vec2/wav2vec2_model.py (add adapter to upstream modules)
3. s3prl/s3prl/upstream/wav2vec2/expert.py (set load_state_dict strict=False to enable the model to load the pretrained checkpoint)
4. s3prl/s3prl/downstream/runner.py (set require_grad = True for the adapter parameters, and freeze the remaining parameters )