Wav2Vec2 pipeline?

[jvel07]

Congrats on the challenge paper!
The paper reports results with w2v2, but couldn't find a fine-tuning pipeline that is reproducible as per the paper.
It would be good to have this in order to be able to reproduce the paper (officially).

[mauricege]

Hi!

You can find the code for fine-tuning wav2vec on the HC-C branch src/ml/wav2vec.py. For HP-C, we only performed feature extraction with wav2vec.

You can reference dvc.yaml for all the steps that are run in the baseline:

ComParE2023/dvc.yaml

Lines 111 to 125 in 89769c5

	wav2vec:
	cmd: python src/ml/wav2vec.py
	deps:
	- ./data/wav
	- ./data/lab/train.csv
	- ./data/lab/devel.csv
	- ./data/lab/test.csv
	- ./src/ml/wav2vec.py
	params:
	- wav2vec
	- target
	outs:
	- results/wav2vec
	- ./tb_logs/wav2vec:
	persist: True

Alternatively, dvc dag will print a graph of the pipeline.

Hope that helps!