This repository hosts the Torch 7 code for the paper Deep Networks with Stochastic Depth available at http://arxiv.org/abs/1603.09382v1. For now, the code reproduces the results in Figure 3 on CIFAR-10 and CIFAR-100. The code for SVHN and 1202-layer CIFAR-10 (which requires some memory optimization) will be available very soon.
- Torch 7 and CUDA with the basic packages (nn, optim, cutorch, cunn)
- cudnn (https://developer.nvidia.com/cudnn) and torch bindings (https://github.com/soumith/cudnn.torch)
- nninit torch package (https://github.com/Kaixhin/nninit)
- CIFAR-10 and CIFAR-100 dataset in Torch format, this script https://github.com/soumith/cifar.torch should very conveniently handle it for you
th main.lua -dataRoot path_to_data -resultFolder path_to_save -deathRate 0.5
This command runs the 110-layer ResNet on CIFAR-10 with stochastic depth, using linear decaying survival probabilities ending in 0.5. The -device flag allows you to specify which GPU to run on. On our machine with a TITAN X, each epoch takes about 60 seconds, and the program ends with a test error (selected by best validation error) of 5.23%.
The default deathRate is set to 0. This is equivalent to a constant depth network, so to run our baseline, enter:
th main.lua -dataRoot path_to_data -resultFolder path_to_save
On our machine with a TITAN X, each epoch takes about 75 seconds, and this baseline program ends with a test error (selected by best validation error) of 6.41% (see Figure 3 in the paper).
You can run on CIFAR-100 by adding the flag -dataset cifar100. Our program provides other options, for example, your network depth (-N), data augmentation (-augmentation), batch size (-batchSize) etc. You can change the optimization hyperparameters in the sgdState variable, and learning rate schedule in the the main function. The program saves a file to resultFolder/errors_N_dataset_deathMode_deathRate, which is a table of nEpochs many tuples, each containing your test and validation error at the end of that epoch.
My email is ys646 at cornell.edu. I'm happy to answer any of your question, and I'd very much appreciate your suggestions.