A slight problem in visualizing layers in classification.inpy

[Hossein Hasanpour]

I'm trying to get the visualization for all layers both for weights and their filters 

For this instead of repeating:

# the parameters are a list of [weights, biases]

filters = net.params['conv1'][0].data

vis_square(filters.transpose(0, 2, 3, 1))

and changing layer name, I tried using a loop like this : 

 for layer_name, param in net.params.iteritems():
    # the parameters are a list of [weights, biases]
    filters = net.params[layer_name][0].data
    vis_square(filters.transpose(0, 2, 3, 1))

now it works fine for the first layer, but gives this error for and stops continuing with the rest: 

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-16-cf7d5999a45c> in <module>()
      2     # the parameters are a list of [weights, biases]
      3     filters = net.params[layer_name][0].data
----> 4     vis_square(filters.transpose(0, 2, 3, 1))

ValueError: axes don't match array

Now what is wrong here? 

by the way what is 

filters.transpose(0, 2, 3, 1)

is doing? I'd be grateful if anyone could give me a hand on this (clearly I dont have python background)

[Hossein Hasanpour]

Any help is greatly appreciated :(

[Jan]

I don't really know how vis_square works, but I can tell about the filters variable. It is a 4D array of shape [n_out, n_in, filter_height, filter_width], which contains all filters of the convolutional layer. The transpose function just transposes the axes of this array, e.g. after that call filters.transpose(0, 2, 3, 1) the filters dimensions are [n_out, filter_height, filter_width, n_in].

Jan

[Hossein Hasanpour]

Thanks alot dear Jan, 

By any chance do you have any script for visualizing layers in python? 

and as a side question, why do we even need to visualize wieghts ? When I visualize wieghts I get color feature maps. whats the use for that? 

I get that visualizing the feature maps, shows the effect of filters, like as if a gabor filter was applied on the input image. but over all I dont get the idea how would these visualization would benefit us . 

again Thank you a great deal for all the helps you have so far given us

[Jan]

Well, I usually just use imshow() and maybe some subplot commands (matplotlib).

Yeah, visualizing the filters usually just shows _that_ the filters have a gabor-like nature, and combining filters from several consecutive layers cleverly sometimes reveals some more refined structure which is apparently extracted/recognized by this combination of filters. There are some approaches trying to understand what features are extracted by which layer/filters and how high-level these features are. E.g. you can look at http://arxiv.org/abs/1311.2901. But in the end, what counts is how well the network performs. And measuring that is not always as easy as building a test set and be done with the final accuracy on that...

Jan

[Sai Krishna]

Hi Jan,

What if the channels of the weight filter are not equal to 3. imshow() doesn't work in that case. Do we any other solution in that case?

thank you!