Views
No views yet
facebook/hubert-large-ls960-ft with a classification head trained
on GAN-augmented data using the CAARMA framework.epoch18_val_acc7997.ckpt — PyTorch Lightning checkpoint (epoch 18, val acc 79.97%)inference.py — minimal loader + embedding extraction script1pip install torch torchaudio transformers pytorch-lightning huggingface_hub
2
3# Download the checkpoint
4huggingface-cli download MassaBaali/HowFar-Caarma epoch18_val_acc7997.ckpt --local-dir .
5
6# Run inference
7python inference.py --ckpt epoch18_val_acc7997.ckpt --audio sample.wav1from inference import load_model, extract_embedding
2
3model = load_model("epoch18_val_acc7997.ckpt", device="cuda")
4embedding = extract_embedding(model, "sample.wav", device="cuda")
5print(embedding.shape)strict=False is used on
load to tolerate minor state-dict key differences.transformers-native
format, so standard AutoModel.from_pretrained will not work.1@article{baali2025caarma,
2 title={CAARMA: Class augmentation with adversarial mixup regularization},
3 author={Baali, Massa and Li, Xiang and Chen, Hao and Hannan, Syed Abdul and Singh, Rita and Raj, Bhiksha},
4 journal={Findings of the Association for Computational Linguistics: EMNLP},
5 volume={2025},
6 pages={9732--9742},
7 year={2025}
8}