README.md

Bag of tricks

An easy-to-reuse code for bag-of-tricks in PyTorch.

Reuse hints

Reuse the codes in utils/trick and follow main.py to modify your code.

For example,

$ git clone https://github.com/siahuat0727/bag-of-tricks
$ cp -r bag-of-tricks/utils path/to/your/repo

then slightly modify your code.

Just a kindly reminder here. For detail, please have a look at the simple and clean demo code main.py instead of the hints below.

1. Cosine learning rate

Use torch.optim.lr_scheduler.CosineAnnealingLR and .step() every training batch.

2. Warmup learning rate

It's recommended to use torch.optim.lr_scheduler for the remaining lr schedule (after warmup), and pass your original scheduler as the parameter after_scheduler in GradualWarmupScheduler. Remember to initialize your optimizer with the real initial lr (not the lr after warmup).

3. Label smoothing

CrossEntropyLossMaybeSmooth is inherited from CrossEntropyLoss. There is usually no need to calculate the loss when evaluating. If you need to do this, you can call .train() to select the smooth version before training and call .eval() to select the non-smooth version before evaluating.

4. Mixup

Aside from the code, the only thing to mention is that the accuracy shown in the training step is usually lower than the actual training accuracy, since the input is a mixture of two images.

Results - CIFAR10

ResNet18

cosine	warmup	label smoothing	mixup	accuracy (90 epochs)	accuracy (400 epochs)
				93.10	94.05
✓				92.94	94.07
	✓			93.18	93.99
		✓		93.38	94.30
			✓	93.77	95.40
✓	✓	✓	✓	94.63	96.01

To reproduce exactly the same result:

$ python main.py --gpu
$ python main.py --gpu --cosine
$ python main.py --gpu --warmup 5
$ python main.py --gpu --smooth 0.1
$ python main.py --gpu --mixup 0.3
$ python main.py --gpu --warmup 5 --smooth 0.1 --mixup 0.3 --cosine

Please star this project if you find it useful. ^.^