In full-convolutional-network model such as InceptionV3 case, it might not be understood input/output data shape of each layers before starting train/predict.
That is model definition is construction like (None, None, None, channels) .
So it might not be able to correctly work visualize_cam .
For details see #144 .
Improvement
It get tensor's shape from actual input/output data, not from model definition.
Problem
In full-convolutional-network model such as InceptionV3 case, it might not be understood input/output data shape of each layers before starting train/predict. That is model definition is construction like
(None, None, None, channels)
. So it might not be able to correctly workvisualize_cam
.For details see #144 .
Improvement
It get tensor's shape from actual input/output data, not from model definition.