Open chrisranderson opened 7 years ago
Hey @chrisranderson,
I worked a bit on visualizing convolution filters and this is what I came up with:
I used a diverging colormap to show the difference between positive and negative, and added a bit of spacing to clarify the convolution layer structure. What do you think?
Padding, proper axes with labels, and a colorscale are all nice looking and superior to what I've got, but I might leave them for the TensorBoard team to implement after Beholder is merged, if they want. Adding padding could make rendering the images substantially slower, too. I'm doing some black magic reshaping that would have to be rethought.
That's the second time I've got the suggestion for a diverging colormap. It seems like a good choice for e.g. networks that use batch normalization. Now to decide on a specific one... I might try a few and try and pick the one that seems like it goes the best with TensorBoard orange. Or put it up to a survey by my labmates.
It could be useful to use hue and value instead of just value. What would hue represent? I could constrain hue to a certain range to keep the visualization from getting too noisy.