This repository integrates Flower and SpeechBrain to achieve training ASR models in FL setting.
- Flower and SpeechBrain can be installed based on the instruction on their GitHub repo (Flower GitHub and SpeechBrain GitHub).
- PyTorch version >= 1.5.0
- Python version >= 3.7
The server should be launched first.
python server.py --data_path /path/to/dataset --tr_path /path/to/train.tsv --test_path /path/to/test.tsv --tr_add_path /path/to/train_add.tsv --weight_strategy num
After obtaining server's IP address, the clients can be launched. Here is an example to start one client.
python client.py --cid 0 --server_address xx.xx.xx.xx --data_path /path/to/dataset --tr_path /path/to/train.tsv --dev_path /path/to/dev.tsv --pre_train_model_path /path/to/model.ckpt --tokenizer_path /path/to/tokenizer
Note that the tokenizer_path
can be the train.csv
file from model pre-training.