results
This model is a fine-tuned version of facebook/wav2vec2-large-xlsr-53 on the None dataset. It achieves the following results on the evaluation set:
- Loss: 0.1638
- Accuracy: 0.975
Model description
The model Urdu audio and classify in following categories
- Angry
- Happy
- Neutral
- Sad
Training and evaluation data
The dataset is available at https://www.kaggle.com/datasets/kingabzpro/urdu-emotion-dataset
Training procedure
Training code is available at https://www.kaggle.com/code/chtalhaanwar/urdu-emotions-hf
Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 5e-05
- train_batch_size: 32
- eval_batch_size: 32
- seed: 42
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- num_epochs: 50
- mixed_precision_training: Native AMP
Training results
Training Loss | Epoch | Step | Validation Loss | Accuracy |
---|---|---|---|---|
1.3838 | 1.0 | 10 | 1.3907 | 0.225 |
1.3732 | 2.0 | 20 | 1.3872 | 0.2125 |
1.3354 | 3.0 | 30 | 1.3116 | 0.6625 |
1.2689 | 4.0 | 40 | 1.1820 | 0.6375 |
1.1179 | 5.0 | 50 | 1.0075 | 0.7 |
0.9962 | 6.0 | 60 | 0.8707 | 0.7125 |
0.8842 | 7.0 | 70 | 0.7485 | 0.7625 |
0.786 | 8.0 | 80 | 0.6326 | 0.8 |
0.6757 | 9.0 | 90 | 0.5995 | 0.8 |
0.6104 | 10.0 | 100 | 0.4835 | 0.825 |
0.5821 | 11.0 | 110 | 0.3886 | 0.9 |
0.4721 | 12.0 | 120 | 0.3935 | 0.8625 |
0.3976 | 13.0 | 130 | 0.3020 | 0.925 |
0.4483 | 14.0 | 140 | 0.3171 | 0.9 |
0.2665 | 15.0 | 150 | 0.3016 | 0.9125 |
0.2119 | 16.0 | 160 | 0.2722 | 0.925 |
0.3376 | 17.0 | 170 | 0.3163 | 0.8875 |
0.1518 | 18.0 | 180 | 0.2681 | 0.9125 |
0.1559 | 19.0 | 190 | 0.3074 | 0.925 |
0.1031 | 20.0 | 200 | 0.3526 | 0.8875 |
0.1557 | 21.0 | 210 | 0.2254 | 0.9375 |
0.0846 | 22.0 | 220 | 0.2410 | 0.9375 |
0.0733 | 23.0 | 230 | 0.2369 | 0.925 |
0.0964 | 24.0 | 240 | 0.2273 | 0.9375 |
0.0574 | 25.0 | 250 | 0.2066 | 0.95 |
0.1113 | 26.0 | 260 | 0.2941 | 0.9125 |
0.1313 | 27.0 | 270 | 0.2715 | 0.925 |
0.0851 | 28.0 | 280 | 0.1725 | 0.9625 |
0.0402 | 29.0 | 290 | 0.2221 | 0.95 |
0.1075 | 30.0 | 300 | 0.2199 | 0.9625 |
0.0418 | 31.0 | 310 | 0.1699 | 0.95 |
0.1869 | 32.0 | 320 | 0.2287 | 0.9625 |
0.0637 | 33.0 | 330 | 0.3230 | 0.9125 |
0.0483 | 34.0 | 340 | 0.1602 | 0.975 |
0.0891 | 35.0 | 350 | 0.1615 | 0.975 |
0.0359 | 36.0 | 360 | 0.1571 | 0.975 |
0.1006 | 37.0 | 370 | 0.1809 | 0.9625 |
0.0417 | 38.0 | 380 | 0.1923 | 0.9625 |
0.0346 | 39.0 | 390 | 0.2035 | 0.9625 |
0.0417 | 40.0 | 400 | 0.1737 | 0.9625 |
0.0396 | 41.0 | 410 | 0.1833 | 0.9625 |
0.0202 | 42.0 | 420 | 0.1946 | 0.9625 |
0.0137 | 43.0 | 430 | 0.1785 | 0.9625 |
0.0214 | 44.0 | 440 | 0.1841 | 0.9625 |
0.0304 | 45.0 | 450 | 0.1690 | 0.9625 |
0.0199 | 46.0 | 460 | 0.1646 | 0.975 |
0.0122 | 47.0 | 470 | 0.1622 | 0.975 |
0.0324 | 48.0 | 480 | 0.1615 | 0.975 |
0.0269 | 49.0 | 490 | 0.1625 | 0.975 |
0.0245 | 50.0 | 500 | 0.1638 | 0.975 |
Framework versions
- Transformers 4.18.0
- Pytorch 1.11.0
- Datasets 2.1.0
- Tokenizers 0.12.1
- Downloads last month
- 103
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social
visibility and check back later, or deploy to Inference Endpoints (dedicated)
instead.