The returned masks were all nan

[irwenqiang]

After predicted 237 correctly. It failed becaused the returned masks were all nan
`# Compute the vanilla mask and the smoothed mask.

vanilla_mask_3d = gradient_saliency.GetMask(im, feed_dict = {neuron_selector: prediction_class})
`

saliency/saliency/base.py

Line 97 in e962539

return self.session.run(self.gradients_node, feed_dict=feed_dict)[0]

values of vanilla_mask_3d:
[[[nan nan nan]
[nan nan nan]
[nan nan nan]
...
[nan nan nan]
[nan nan nan]
[nan nan nan]]

[irwenqiang]

It seems that x is not a parameter of y, so get not any dy/dx

[irwenqiang]

tensorflow r1.2 works well

[JarvisDevon]

I am getting the same result using Tensorflow Version 1.4.0

This problem however only occurs on models which do not perform well in the domain I am using (the majority of the saliency maps are NaN, however on occasion the model will provide a very dark saliency map). I have trained models which perform well in the domain and provide clear saliency maps.

Thus, I am curious if there is an interpretation for the NaN's or a conclusion that can be made. My feeling is that the NaN's occur due to the model's inability to develop strong or decisive weights. Would this be a logical cause for the NaNs?