Open ghost opened 6 years ago
Yea. Batching has a bunch of other consequences. Its been on my mind for a long time now to support batched operations. That is the first thing coming in the next release.
I'm wondering if keras-vis's "saliency" and "filter" looks something like this:
https://jupyter.brynmawr.edu/services/public/dblank/Examples/conx/Sample%20Output%20in%20Conx.ipynb
Not competing with keras-vis, but making an easy to use system for keras non-experts. Happy to collaborate, too!
I build the following model
The first thing i observed is that even if i pass input sizes of [1, 28, 28] (MNIST), the r.shape is (28, 28, 3), but I guess this is intended behaviour since we need more than a gray image to visualize saliency. But if I choose to visualize all filters of the second layer, why i am getting back only one heatmap? I was expecting r.shape to be of size [12, 28, 28, 3], but i only get [28, 28, 3].