dblN / stochastic_depth_keras

Keras implementation for "Deep Networks with Stochastic Depth" http://arxiv.org/abs/1603.09382

[Deprecated] No longer maintained (to catch up with the latest keras/theano/tensorflow but you'll get the idea of how to implement stochastic depth)

Deep Networks with Stochastic Depth

Keras implementation for "Deep Networks with Stochastic Depth" http://arxiv.org/abs/1603.09382

Original code(in lua/torch) is at https://github.com/yueatsprograms/Stochastic_Depth.

Usage

1. Install Theano following its instruction.
2. Install Keras (I use new API at keras-1 branch)

$ cd keras
$ git checkout keras-1
$ python setup.py install

3. Just run python train.py

Known Issues

• Error related to maximum recursion depth
  • When the network is deep, there happens error saying it reaches to maximum recursion depth.
  • You can resolve this issue by using sys.setrecursionlimit(max_recursion_depth). You should increase max_recursion_depth until you get no error (Increasing this value might cause segmentation fault if you don't have enough memory).

Results

CIFAR10

• Number of layers == 50
• (other configs are same as train.py)