unofficial mxnet reimplement
I had refer to this repo Bag_of_Tricks_for_Image_Classification_with_Convolutional_Neural_Networks in Pytorch.
4 * TITAN Xp
resnet34_v2, lr=0.01 batch_size=64, epoch=450,lr_step: 300,350,400
baseline: 0.594
Usage: python3 cache/run_train.sh
I had try some tricks which were from the Paper Bag of Tricks for Image Classification with Convolutional Neural Networks :
+wramup 60.3
+warmup +label smooth 0.644
+warmup +label smooth +cutout 0.618
+warmup +label smooth +RandomErasing 0.622
| trick | acc |
|---|---|
| baseline | 0.594 |
| +warmup | 0.603 |
| +label smooth | 0.644 |
| +cutout | 0.618,it doesn't work |
| +RandomErasing | 0.622,it doesn't work |