vit-base-vocalsound-logmel

This model is a fine-tuned version of google/vit-base-patch16-224 on VocalSound dataset. It achieves the following results on the evaluation set:

• accuracy: 88.8
• precision (micro): 91.3
• recall (micro): 87.1
• f1 score (micro): 89.1
• f1 score (macro): 89.1

Training and evaluation data

Training: VocalSound training split (#samples = 15570)

Evaluation: VocalSound test split(#samples = 3594)

Training hyperparameters

The following hyperparameters were used during training:

• optimizer: AdamW
  • weight_decay: 0
  • learning_rate: 5e-5
• batch_size: 32
• training_precision: float32

Preprocessing

Differently from vit-base-vocalsound, the log-melspectrogram is used(log was applied as an addition) and the preprocessor normalization step uses VocalSound statistics(i.e. mean and std) instead of the default IMAGENET ones.

Framework versions

• Transformers 4.27.4
• TensorFlow 2.12.0
• Tokenizers 0.13.3