Saved graph returns non-deterministic results about darkflow HOT 4 CLOSED

Indeed it was related to dropout, it was saved as 0.5 instead of 1.0

from darkflow.

Can you elaborate on this? If this is a bug I am welcoming your pull requests!

from darkflow.

I am not sure if it is a bug or planned feature, but TensorFlow graphs saved using --savepb contain a dropout layer with a keep probability of 0.5. For evaluation/testing purposes, it is not good, because

1. It makes the net "weaker"
2. It makes the results non deterministic, each run a bit different result.

It might be good if one is planning to re-train the network in another framework.

I don't know how to fix it correctly. Currently I hard-coded the keep probability to 1.0 when the layer is being built. That is of course not good for training.
I guess that you can put a code right before the actual saving that changes the probability?

class dropout_layer(Layer):
def setup(self, p):
self.h['pdrop'] = dict({
#'feed': p, # for training
'feed': 1.0, # for training
'dfault': 1.0, # for testing
'shape': ()
})

from darkflow.

Thanks for spotting this silly bug. Please pull the new code and check if it is working.

from darkflow.