---
language: en
license: apache-2.0
tags:
- text-classfication
- int8
- PostTrainingStatic
datasets: 
- mrpc
metrics:
- f1
---

# INT8 BERT base uncased finetuned MRPC

### Post-training static quantization

This is an INT8  PyTorch model quantized with [Intel® Neural Compressor](https://github.com/intel/neural-compressor). 

The original fp32 model comes from the fine-tuned model [Intel/bert-base-uncased-mrpc](https://huggingface.co/Intel/bert-base-uncased-mrpc).

The calibration dataloader is the train dataloader. The default calibration sampling size 300 isn't divisible exactly by batch size 8, so the real sampling size is 304.

The linear module **bert.encoder.layer.9.output.dense, bert.encoder.layer.10.output.dense** falls back to fp32 to meet the 1% relative accuracy loss.

### Test result

- Batch size = 8
- [Amazon Web Services](https://aws.amazon.com/) c6i.xlarge (Intel ICE Lake: 4 vCPUs, 8g Memory) instance.

|   |INT8|FP32|
|---|:---:|:---:|
| **Throughput (samples/sec)**  |23.986|11.202|
| **Accuracy (eval-f1)** |0.8997|0.9042|
| **Model size (MB)**  |120|418|

### Load with Intel® Neural Compressor (build from source):

```python
from neural_compressor.utils.load_huggingface import OptimizedModel
int8_model = OptimizedModel.from_pretrained(
    'Intel/bert-base-uncased-mrpc-int8-static',
)
```

Notes:  
 - The INT8 model has better performance than the FP32 model when the CPU is fully occupied. Otherwise, there will be the illusion that INT8 is inferior to FP32.