Adam-ousse/ast-cremad-finetuned

AST fine-tuned on CREMA-D

This model is an ASTForAudioClassification checkpoint fine-tuned from:

• MIT/ast-finetuned-audioset-10-10-0.4593

Dataset

• Dataset: MahiA/CREMA-D
• Classes: anger, disgust, fear, happy, neutral, sad
• Split protocol:
  • official train.csv was split into train and validation
  • validation was stratified 10% of official train
  • official test.csv was used as test

Training notes

• Loss: class-weighted cross-entropy to account for class imbalance

Validation result

• Best epoch: 17
• Best validation accuracy: 0.662751677852349

Test result

• Split: test
• Samples: 1489
• Loss: 1.767368
• Accuracy: 0.70047

Per-class accuracy

• anger: 0.7280 (174/239)
• disgust: 0.3520 (44/125)
• fear: 0.4529 (77/170)
• happy: 0.4474 (34/76)
• neutral: 0.8608 (699/812)
• sad: 0.2239 (15/67)

Usage

import torch
from transformers import AutoFeatureExtractor, ASTForAudioClassification

model_id = "Adam-ousse/ast-cremad-finetuned"
feature_extractor = AutoFeatureExtractor.from_pretrained(model_id)
model = ASTForAudioClassification.from_pretrained(model_id)
model.eval()

# waveform: 1D float32 numpy array at 16 kHz
# inputs = feature_extractor([waveform], sampling_rate=16000, return_tensors="pt", padding=True)
# with torch.no_grad():
#     logits = model(**inputs).logits
#     pred = int(torch.argmax(logits, dim=1)[0])

Labels

The model config includes:

• id2label: {0: anger, 1: disgust, 2: fear, 3: happy, 4: neutral, 5: sad}
• label2id: {anger: 0, disgust: 1, fear: 2, happy: 3, neutral: 4, sad: 5}