Deep Residual Net
Example code for Deep Residual Learning for Image Recognition
- Run this script by
python resnet-small.pyfor 100 epochs get a train accuracy around 89.47% and validation accuracy around 85.95%
- Then change the learning rate to 0.01, running this training from 100th epoch for 50 iterations, and get a train accuracy around 98.72% and test accuracy around 89.77%
Differences to the Paper
- 1*1 convolution operators are used for increasing dimensions.
- This is a small residual net consists of 52 layers(can change to 20, 32, 44 layers by changing
ResidualSymbolto 3, 5, 7)
- Using mxnet default data augmentation options include center crop (instead of random crop) and random mirror, no paddings on raw image data and the input image size is 28*28(instead of 32*32).