File size: 1,846 Bytes
53737cf
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
# HuBERT fine-tuned on DUSHA dataset for speech emotion recognition in russian language

The pre-trained model is this one - [facebook/hubert-large-ls960-ft](https://huggingface.co/facebook/hubert-large-ls960-ft)

The DUSHA dataset used can be found [here](https://github.com/salute-developers/golos/tree/master/dusha#dataset-structure)

# Fine-tuning

Fine-tuned in Google Colab using Pro account with A100 GPU

Freezed all layers exept projector, classifier and all 24 HubertEncoderLayerStableLayerNorm layers

Used half of the train dataset

# Training parameters

2 epochs \
train batch size = 8 \
eval batch size = 8 \
gradient accumulation steps = 4 \
learning rate = 5e-5 without warm up and decay

# Metrics

Achieved \
accuracy = 0.86 \
balanced = 0.76 \
macro f1 score = 0.81 \
on test set, improving accucary and f1 score compared to dataset baseline

# Usage

```python
from transformers import HubertForSequenceClassification, Wav2Vec2FeatureExtractor
import torchaudio
import torch

feature_extractor = Wav2Vec2FeatureExtractor.from_pretrained("facebook/hubert-large-ls960-ft")
model = HubertForSequenceClassification.from_pretrained("xbgoose/hubert-dusha-finetuned")
num2emotion = {0: 'neutral', 1: 'angry', 2: 'positive', 3: 'sad', 4: 'other'}

filepath = "path/to/audio.wav"

waveform, sample_rate = torchaudio.load(filepath, normalize=True)
transform = torchaudio.transforms.Resample(sample_rate, 16000)
waveform = transform(waveform)

inputs = feature_extractor(
        waveform, 
        sampling_rate=feature_extractor.sampling_rate, 
        return_tensors="pt",
        padding=True,
        max_length=16000 * 10,
        truncation=True
    )

logits = model(inputs['input_values'][0]).logits
predictions = torch.argmax(logits, dim=-1)
predicted_emotion = num2emotion[predictions.numpy()[0]]
print(predicted_emotion)
```