compare resnet 32,110 (pre-activation) with and without mixup
I have trained the following four models and record the best accuracies on validation set, which are shown below (with batchsize 128):
| Layers | without mixup | with mixup |
|---|---|---|
| 32 | 93.08% | 93.33% |
| 110 | 94.48% | 94.75% |
The loss and error on train/validation set are shown in the figures (Logarithmic coordinates):
- resnet 32 (pre-activation)
- resnet 110 (pre-activation)
- code in this depository is a implement of resnet 110 (pre-activation) with mixup
- there are some logs which lead to the figures above in floder
bak, as well as some weights checkpoints
-
python 3
-
pytorch 0.4.0