by yueatsprograms

yueatsprograms / Stochastic_Depth

Deep Networks with Stochastic Depth

444 Stars 79 Forks Last release: Not found 21 Commits 0 Releases

Available items

No Items, yet!

The developer of this repository has not created any items for sale yet. Need a bug fixed? Help with integration? A different license? Create a request here:

Deep Networks with Stochastic Depth

This repository hosts the Torch 7 code for the paper Deep Networks with Stochastic Depth available at For now, the code reproduces the results in Figure 3 for CIFAR-10 and CIFAR-100, and Figure 4 left for SVHN. The code for the 1202-layer network is easily modified from the repo

using our provided module for stochastic depth.

Table of Contents


Please see the latest implementation of stochastic depth and other cool models (DenseNet etc.) in PyTorch, by Felix Wu and Danlu Chen. Their code is much more memory efficient, more user friendly and better maintained. The 1202-layer architecture on CIFAR-10 can be trained on one TITAN X (amazingly!) under our standard settings.


  • Torch 7 and CUDA with the basic packages (nn, optim, image, cutorch, cunn).
  • cudnn and torch bindings.
  • nninit;
    luarocks install nninit
    should do the trick.
  • CIFAR-10 and CIFAR-100 datasets in Torch format; this script should very conveniently handle it for you.
  • SVHN dataset in Torch format, available here. Please note that running on SVHN requires roughly 28GB of RAM for dataset loading.

Getting Started on CIFAR-10

git clone
cd Stochastic_Depth
git clone
cd cifar.torch
th Cifar10BinToTensor.lua
cd ..
mkdir results
th main.lua -dataRoot cifar.torch/ -resultFolder results/ -deathRate 0.5

Usage Details

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 decay survival probabilities ending in 0.5. The
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 5.25%.

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 (
), data augmentation (
), batch size (
) 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 every epoch to
, which has a table of tuples containing your test and validation errors until that epoch.

The architecture and number of epochs for SVHN used in our paper are slightly different from the code's default, please use the following command if you would like to replicate our result of 1.75% on SVHN:

th main.lua -dataRoot path_to_data -resultFolder path_to_save -dataset svhn -N 25 -maxEpochs 50 -deathRate 0.5

Known Problems

  • It is normal to get a +/- 0.2% difference from our reported results on CIFAR-10, and analogously for the other datasets. Networks are initialized differently, and most importantly, the validation set is chosen at random (determined by your seed).
  • If you train on SVHN and the model doesn't converge for the first 1600 or so iterations, that's ok, just wait for a little longer.
  • Xavier reported that the model is able to converge for him on CIFAR-10 only after he uses the following initalization for Batch Normalization
    model:add(cudnn.SpatialBatchNormalization(_dim_):init('weight', nninit.normal, 1.0, 0.002):init('bias', nninit.constant, 0))
    . We could not replicate the non-convergence and thus won't put this initialization into our code, but recognize that machines (or the versions of Torch installed) might be different.


My email is ys646 at I'm happy to answer any of your questions, and I'd very much appreciate your suggestions. My academic website is at

We use cookies. If you continue to browse the site, you agree to the use of cookies. For more information on our use of cookies please see our Privacy Policy.