resnet.md

Resnet Architecture

To use the resnet architecture, one can set the arch option to either resnet18/34/50/101 or pytorch-resnet18/34/50/101. The former is realized in model/resnet_.py and the later is implemented in model/resnet.py. The difference between the choice is that the later one is from the pytorch offical and the former is more flexible. Users can customize the Resnet architecture variants with different combination of options for resnet18/34/50/101. The combination is possible by setting segments in keyword.

For example, by add origin,cbsa,fix_pooling,singleconv,fix in the keyword, we obtain the pytorch official architecture (the same with the pytorch-resnet18/34/50/101). Fine-grain control is supported for the stem and body.

Download the QTool.png to view in a whole.

Stem

1. imagenet or CIFAR10 / CIFAR100 to change the first convolution hyperparameter (as well as the maxpool layer).

For imagenet, the conv is kernel_size7x7, stride=2, pad=3.

For CIFAR10/CIFAR100, the conv is kernel_size and no maxpool.

Besides, set alpha_width = 0.25 for CIFAR architecture to shrink the channel width.

2. fix_pooling to choose whether the bn layer should be moved before maxpool layer.

3. TRe is an alternative implementation of the Stem.

Body

• Default buildup

• Lossless downsample network

• Prone: Point-wise and Reshape Only Network

• Specific Activation and Normalization