python-dbn aims to make experimenting with different deep learning architectures. Right now, it's capability is limited to specifying deep feed-forward networks.

An example:

from dbn        import DBN
from dbn.layers import *

net = DBN([
		OneHotSoftmax(8),
		Sigmoid(3)
	  ],3)

net.fit(train_data,train_labels)
net.predict(test_data)

###There's a still a lot left to do

• Persistence of learnt weights Of highest importance
• Multiple visible layer sets per RBM Multiple types of input for each RBM. Hopefully this will make experimenting different ways for data input to the network possible.
• Auto-encoders Building on to the DBN class to create an intuitive way to build auto-encoders just by specifying the layer dimensions.
• Documentation I'd like the library to be as thoroughly documented as scikit-learn. I've learnt a ton from that library, largely because the documentation of the library is so complete, going right down to the theory level of things, just enough so you can quickly understand and start implementing. My hope is that this library achieves similar goals.

###Getting the damn theory right I'm still realy new at this, so I'd appreciate any help I can get with the theory side of things. Things I'm still unsure of that are already implemented:

• When do I stop training for RBMs? I've looked at Hinton's tutorial on training RBMs, but it doesn't seem to be helping for some of the tasks I've tested it with.
• With a softmax layer one-hot input, should my reconstruction be also be turning just one neuron on based in the probability densities? Or should I just leave the values as probabilities? I've heard (though I can't remember from where) that this results in slower learning due to the more distributed nature of the updates.
• Learning rate decay with RBMs: Yay or nay?