Visualizing Filters about dcgan.torch HOT 3 CLOSED

image.scale will help scale up the visualization.
image.save (as you have seen) will help visualize outside interactive notebook.
Otherwise, the display package will also work, as seen in the dcgan.torch code

from dcgan.torch.

I think I am misunderstanding how these weights can be visualized. Is the image I posted above the entire set of kernels for that layer? I'm not sure I understand exactly how to project the other layers into pixel space. For instance if I take the first conv layer (weight:size = 100x1024x4x4), does it need to be narrowed to a three channel mapping as:

filters=net:get(1).weight
filters=filters:narrow(2, 1, 3)
image.toDisplayTensor{input=filters}
image.save("filters.png", filters)

And then, how would I visualize the features that are maximally activated by the training data? I guess I am looking at the deconvnet paper and wonder if there are any examples of this in torch?

from dcgan.torch.

the filters from other layers have to be visualized as single-channel images.
so, reshape weight to become size: 100 * 1024 x 4 x 4

from dcgan.torch.