Multi-class multi-label classification with BERT CLS token representation.
SemEval-2018 Task1 E-c-English. See the official website.
- python == 3.7.7
- torch == 1.5.1
- transformers == 3.0.2
- numpy == 1.19.4
- tqdm == 4.53.0
- sklearn == 0.21.3
- deepspeed == 0.4.1
python run.py- GPU: Titan Xp * 1 (12GB)
- CPU: Intel(R) Xeon(R) CPU E5-2650 v4 @ 2.20GHz
- Batch size: 64
| Name | Epoch | Max GPU Memory (MB) | Train Time/epoch (s) | Test Time/epoch (s) | Macro F1 (%) |
|---|---|---|---|---|---|
| Baseline | 3 | 8835 | 26 | 3 | 46.632 |
| DeepSpeed (stage 0) | 3 | 5647 | 26 | 3 | 46.632 |
| DeepSpeed (stage 1) | 3 | 5757 | 29 | 3 | 46.155 |
| DeepSpeed (stage 2) | 3 | 5757 | 30 | 3 | 46.155 |
| DeepSpeed (stage 3, disabled) | 3 | 7143 | 51 | 8 | 46.491 |
This project is under MIT licence. For licence of dataset, refer this for more information.