• Lasagne Tutorial
  • MLP Example
  • CNN

Start: 2015-09-16

Lasagne Tutorial

As of 16 September 2015, Nolearn only supports Lasagne, though it's trying to support Keras due to GitHub issue discussion. I need the cross_validation and random_hyper components of Sklearn to work with deep neural networks.

Keras hides Theano from users; whereas Lasagne exposes Theano to users to some extent. As far as I know, when training RNNs, I don't have to use theano.scan() manually in Lasagne, which is good enough.

MLP Example

http://lasagne.readthedocs.org/en/latest/user/tutorial.html The sample code is in the /examples folder

input_var is the Theano variable that the network's input layer will be linked to. In the example, input_var is linked to a variable given as an argument to the build_mlp() function.

After loading the input data, it applies 20% dropout (this is the first time I know this trick):

l_in_drop = lasagne.layers.DropoutLayer(l_in, p=0.2)

Lasagne does not support .yaml, because it believes that is's more flexible to encode architecture in Python itself.

CNN

dropout tends not to work well for convolutional layers.