Created by: Sand3r-
Provides a fix for accuracy drop obtained with QAT models by enabling uint8 variable type on conv-relu outputs. Also enables executing VGG16 and VGG19 by:
- Adding additional way of fetching scales (by associating tensor with its tensor scale by the *.scale suffix)
- Fixing bug with empty list of disabled op ids Restores fake-quant mul to achieve best accuracy.
Accuracy comparison
Resnet50
top1 acc | top 5 acc | |
---|---|---|
Original QAT | 0.7655 | 0.9304 |
Transformed QAT | 0.7653 | 0.9298 |
Diff | 0.0002 | 0.0006 |
Mobilenet V1
top1 acc | top 5 acc | |
---|---|---|
Original QAT | 0.7077 | 0.8954 |
Transformed QAT | 0.7076 | 0.8943 |
Diff | 0.0001 | 0.0011 |