xinhe's picture
Update README.md
7125487
|
raw
history blame
1.9 kB
metadata
language:
  - en
license: apache-2.0
tags:
  - multiple-choice
  - int8
  - PostTrainingStatic
datasets:
  - swag
metrics:
  - accuracy
model-index:
  - name: bert-base-uncased-finetuned-swag-int8-static
    results:
      - task:
          name: Multiple-choice
          type: multiple-choice
        dataset:
          name: Swag
          type: swag
        metrics:
          - name: Accuracy
            type: accuracy
            value: 0.7838148474693298

INT8 bert-base-uncased-finetuned-swag

Post-training static quantization

This is an INT8 PyTorch model quantized with Intel® Neural Compressor.

The original fp32 model comes from the fine-tuned model thyagosme/bert-base-uncased-finetuned-swag.

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 bert.encoder.layer.2.output.dense, bert.encoder.layer.5.intermediate.dense, bert.encoder.layer.9.output.dense, bert.encoder.layer.10.output.dense fall back to fp32 to meet the 1% relative accuracy loss.

Test result

INT8 FP32
Throughput (samples/sec) 16.55 9.333
Accuracy (eval-accuracy) 0.7838 0.7915
Model size (MB) 133 418

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

from neural_compressor.utils.load_huggingface import OptimizedModel
int8_model = OptimizedModel.from_pretrained(
    'Intel/bert-base-uncased-finetuned-swag-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.