CAMs show abnormal peak at lower left corner for all input samples

[dipayandas97]

I constructed a chexnet model using DenseNet121 and loaded our pre-trained weights. On generating CAMs of the final conv layer, an unwanted region of high activations are found on all output CAMs. This is regardless of input samples.

First two are outputs from abnormal class and last two from normal class

[armiro]

Same here. @dipayandas97 could you find any probable reason for that or how to correct it?

[mwestt]

I would also be interested to hear some rationale behind what's causing this, as I've just encountered it.

[naomity]

same issue even for my own implementation/training, any possible reasons?

[Aliktk]

@armiro @naomity @dipayandas97 can you please tell me about all the dependencies including Cuda and NumPy etc version I will be very thankful. I am trying from almost one month. Please help me in this regard