Train CIFAR10 with PyTorch

I'm playing with PyTorch on the CIFAR10 dataset.

Pros & cons

Pros:

• Built-in data loading and augmentation, very nice!
• Training is fast, maybe even a little bit faster.
• Very memory efficient!

Cons:

• No progress bar, sad :(
• No built-in log.

Accuracy

Model	Acc.
VGG16	92.64%
ResNet18	93.02%
ResNet50	93.62%
ResNet101	93.75%
MobileNetV2	94.43%
ResNeXt29(32x4d)	94.73%
ResNeXt29(2x64d)	94.82%
DenseNet121	95.04%
PreActResNet18	95.11%
DPN92	95.16%

Learning rate adjustment

I manually change the lr during training:

• 0.1 for epoch [0,150)
• 0.01 for epoch [150,250)
• 0.001 for epoch [250,350)

Resume the training with python main.py --resume --lr=0.01

Training the model

python main.py

Specify a batch size by using -b <batch_size>. By default it is 128.

By default the output is not saved to a file.

Making time graphs with varying batch sizes

python compare_batches.py <name>

When training the model, have the output written to log/<name>-<batch_size>, where <name> will be used in the title of the produced graph.

The batch sizes used in the graph are specified in the batch_no variable.