output heatmap visualisation (5.4) changes everytime I load my network

fchollet / deep-learning-with-python-notebooks

Jupyter notebooks for the code samples of the book "Deep Learning with Python"

MIT License

18.68k stars 8.65k forks source link

output heatmap visualisation (5.4) changes everytime I load my network #58

Open JaspervDalen opened 6 years ago

JaspervDalen commented 6 years ago

I tested notebook 5.4 and it works as expected. However when I change the network from VGG16 to my own network (mobilenet with 2 dense and dropout layers behind it), the heatmaps change every time I load the network (load model from keras using .h5 file). Any idea why this happens? The output of the network (softmax probabilities) are equal for every network load by the way

JaspervDalen commented 6 years ago

I tried it again with just the mobilenet (applications.mobilenet) trained on imagenet to just replicate the african elephant. The prediction is correct however the heatmap does not work...