Deep nets with stochastic depth about recipes

[f0k]

As you can see, if the next layer and the current layer have different output shapes, you need to linearly project the output of the current layer so that it matches the dimensions of the output of the following layer.

The identity transform in Eq. 2 is the same as in Eq. 1, which is just the shortcut connection of the ResNet. For shortcuts changing the spatial size (dashed arrows in http://arxiv.org/abs/1512.03385, Fig 3), there are two options, explained in the first paragraph of page 4 of http://arxiv.org/abs/1512.03385. An implementation of that paper is given in https://github.com/Lasagne/Recipes/blob/master/papers/deep_residual_learning/Deep_Residual_Learning_CIFAR-10.py, including these two options.

On page 8 of the stochastic depth paper, they mention that for blocks changing the number of filters and spatial dimension, they "replace the identity connections in these blocks by an average pooling layer followed by zero paddings to match the dimensions." This is neither of the two options in the ResNet paper, but it's easy enough to modify the existing Lasagne Recipe to do so.

If in doubt about what they did in the stochastic depth paper, refer to the source code at https://github.com/yueatsprograms/Stochastic_Depth.

/edit: If you manage to reproduce the results of the stochastic depth paper (CIFAR-10 will be the easiest target), we'd appreciate a PR to this repository.
For additional fun, note that there's also a second ResNet paper (https://arxiv.org/abs/1603.05027) which was done concurrently to the stochastic depth paper. It's possible that combining these two would yield even better results. A Lasagne implementation is here: https://github.com/FlorianMuellerklein/Identity-Mapping-ResNet-Lasagne.

[christopher-beckham]

Ah, thank you! How did I not notice the implementation detail section in the paper... I'll take a crack at this and see what I can come up with!