AST Fine-Tuned on ESC-50 (Fold 1)
This checkpoint is an AST audio classifier fine-tuned on ESC-50 from the base model MIT/ast-finetuned-audioset-10-10-0.4593 (took the feature extractor and replaced the classification head to match ESC-50 dataset).
Training Setup
- Dataset: ESC-50
- Split protocol: Fold 1 as test, Fold 2 as validation, remaining folds as train
- Base model: MIT/ast-finetuned-audioset-10-10-0.4593
- Epochs: 30
- Batch size: 32
- Learning rate: 5e-5
- Weight decay: 1e-4
- Warmup ratio: 0.1
- Seed: 42
Results
- Best validation accuracy: 0.9675 (epoch 7)
- Test accuracy (fold 1): 0.9350
- Test loss (fold 1): 0.2708
Usage
from transformers import AutoFeatureExtractor, ASTForAudioClassification
repo_id = "Adam-ousse/ast-esc50-finetuned-fold1"
feature_extractor = AutoFeatureExtractor.from_pretrained(repo_id)
model = ASTForAudioClassification.from_pretrained(repo_id)
model.eval()
Notes
- This checkpoint corresponds to one fold setup (test fold 1).
- For publication-grade reporting, train and report all 5 folds and provide mean accuracy.