usage inquiry: retrieving weights for hinton

[DPR-Sanchez]

I was reviewing the documentation and I haven't managed to figure out how to retrieve the matrix of weights from a trained neural network to pass to the hinton graph. I was reviewing the below documentation and only a demo numpy matrix is utilized as an example.

http://neupy.com/docs/secondary/visualizations.html

Any insight into this matter would be much appreciated. Thank you again for creating this wonderful neural network library!

[itdxer]

That's a good point, I think I'll need to add this as a feature or at least mention it in the documentation. For now, you can do it like

from neupy.utils import tensorflow_session

sess = tensorflow_session()
layer = optimizer.network.layers[1]  # or network.layers[1]
weights = sess.run(layer.weight)

[DPR-Sanchez]

I tried my hand at a rough mock-up of the feature as a function since I think it would only be a small for loop as seen below:

from neupy.utils import tensorflow_session
import numpy as np

def weights():
	sess = tensorflow_session()
	layers = optimizer.network.layers  

	weights = []

	for layer in layers:	
		weights.append(sess.run(layer.weight))

	return np.toarray(weights)

hope it helps, and thank you again for your feedback.

[itdxer]

you also need to remember that not every layer has weights (thelayer.weight operation might fail) and some layers also can have bias (layer.bias) which you can also visualise with a Hinton diagram.

[DPR-Sanchez]

I know the below is an ugly mockup, but I gave another stab at it

from neupy.utils import tensorflow_session
import numpy as np

def weights():
	sess = tensorflow_session()
	layers = optimizer.network.layers  

	weights = []

	for layer in layers:
		try:
			weights.append(sess.run(layer.bias))
		except Exception as e:
			pass 
		try:
			weights.append(sess.run(layer.weight))	
		except Exception as e:
			pass 

	return np.toarray(weights)