Pre-Softmax Analysis

[albermax]

Hi Luisa,

I try to create visualizations for the pre-softmax activations. My understanding from the documentation and issue 4 is that I only need to change in line blobnames = ["prob"] to blobnames =["loss3/classifier"] (for Googlenet). Unfortunately this leads an "empty" visualization:

Pre-Softmax:

compared to Post-Softmax:

Do I do something wrong?
One thing I noticed is:

/home/bbdc/alber/tmp/zintgraf/DeepVis-PredDiff2/prediction_difference_analysis.py:222: RuntimeWarning: invalid value encountered in log2
oddsTarVal = np.log2(tarVal_laplace/(1-tarVal_laplace))
/home/bbdc/alber/tmp/zintgraf/DeepVis-PredDiff2/prediction_difference_analysis.py:223: RuntimeWarning: invalid value encountered in log2
oddsAvgP = np.log2(avgP_laplace/(1-avgP_laplace))

A further questions regarding my understanding of the code/algorithm. The algorithm shows which parts of an image increase or lower the activation of a neuron, e.g., an output neuron for class c and in that case the the visualization indicates what speaks for and against a class. Where in the code is the neuron selected?
At line 125 for the sensitivity analysis the neuron with the largest activation in the pre-softmax activation is selected. Is there something similar for the prediction difference analysis? To me target_func in line 98 considers whole layers?

Thank you!
Cheers,
Max