Update README.md
Browse files
README.md
CHANGED
|
@@ -34,6 +34,8 @@ The original fp32 model comes from the fine-tuned model [thyagosme/bert-base-unc
|
|
| 34 |
|
| 35 |
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.
|
| 36 |
|
|
|
|
|
|
|
| 37 |
### Test result
|
| 38 |
|
| 39 |
- Batch size = 8
|
|
|
|
| 34 |
|
| 35 |
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.
|
| 36 |
|
| 37 |
+
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.
|
| 38 |
+
|
| 39 |
### Test result
|
| 40 |
|
| 41 |
- Batch size = 8
|