Deep Residual Learning for Image Recognition

This is a Torch implementation of "Deep Residual Learning for Image Recognition",Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun the winners of the 2015 ILSVRC and COCO challenges.

How to use

• Install Torch, CUDNN, and required torch libraries (at least fs, display, image, cudnn, nngraph, ...)
• Download Imagenet to a folder.
• Run th train-imagenet.lua --dataTrainRoot /mnt/imagenet/train --dataValRoot /mnt/imagenet/val Images are assumed to have the path /mnt/imagenet/train/n1234/photo.JPEG, where n1234 is the synset ID.
• If you want to monitor progress, run th -ldisplay.start in a separate shell and visit http://localhost:8000. A log of the training loss is reported, as well as the network's instantaneous weights, gradients, and activations.

Accuracy

TODO

Notes and shortcomings

• This is only a classification task for now. No semantic segmentation yet.

• I haven't tested this network yet. It appears to begin converging, but not a whole lot.

• The size of the network is so big that I can only fit a small version in memory. (Focusing on correctness rather than efficiency for now...)