|
--- |
|
license: cc-by-4.0 |
|
tags: |
|
- int8 |
|
- Intel® Neural Compressor |
|
- PostTrainingStatic |
|
datasets: |
|
- squad2 |
|
metrics: |
|
- f1 |
|
--- |
|
|
|
# INT8 RoBERT base finetuned on Squad2 |
|
|
|
### Post-training static quantization |
|
|
|
This is an INT8 PyTorch model quantized with [huggingface/optimum-intel](https://github.com/huggingface/optimum-intel) through the usage of [Intel® Neural Compressor](https://github.com/intel/neural-compressor). |
|
|
|
The original fp32 model comes from the fine-tuned model [deepset/roberta-base-squad2](https://huggingface.co/deepset/roberta-base-squad2). |
|
|
|
The calibration dataloader is the train dataloader. The default calibration sampling size 100 isn't divisible exactly by batch size 8, so the real sampling size is 104. |
|
|
|
The linear modules **roberta.encoder.layer.7.output.dense**, **roberta.encoder.layer.8.output.dense**, **roberta.encoder.layer.9.output.dense**, fall back to fp32 for less than 1% relative accuracy loss. |
|
|
|
### Evaluation result |
|
|
|
| |INT8|FP32| |
|
|---|:---:|:---:| |
|
| **Accuracy (eval-f1)** |82.3122|82.9231| |
|
| **Model size (MB)** |141|474| |
|
|
|
### Load with optimum: |
|
|
|
```python |
|
from optimum.intel.neural_compressor.quantization import IncQuantizedModelForQuestionAnswering |
|
int8_model = IncQuantizedModelForQuestionAnswering.from_pretrained( |
|
'Intel/roberta-base-squad2-int8-static', |
|
) |
|
``` |
|
|
