Same image - Different output

[groot-1313]

After training the model, I tested it from the latest checkpoint.
The model predicts different outputs for the same input image!!!

[groot-1313]

I switched off the dropout to avoid it.

[julien2512]

dropout garanties non specialization of each neuron in a NxN matrix (you do not need to apply it on convolutional networks). That mean your network would be more flexible to change.

I do not know your usecase exactly, but you should keep it while learning.
Eventhough, may be it is possible to keep it for the test.

[mattchiang-gsp]

@groot-1313 How exactly did you switch off dropout?