Stripping keras_learning

[Lakedaemon]

I am building an app for android with a keras model exported (through tenserflow).

Yet, it crashes at runtime because there are DT_BOOL node in the graph and the tensorflow library I use doesn't have KernelOp for DT_BOOL registered out of the box. So, I would like to try getting read of all DT_BOOL noes (if it is possible) from the model that I only use for inference on the device.

Is it possible to strip a keras model from all DT_BOOL nodes that depens on keras_learning_phase ?
How can it be done ?

I tried this :

K.set_learning_phase(0) # all new operations will be in test mode from now on

serialize the model and get its weights, for quick re-building

config = model.get_config()
weights = model.get_weights()

re-build a model where the learning phase is now hard-coded to 0

new_model = Sequential.from_config(config)
new_model.set_weights(weights)

And I also tried a few things on the tensorflow side (optimize_for_inference, freeze_graph, (wrongly) turning Switch node into Identity (I corrupted my graph), ...)

[Lakedaemon]

I worked around this issue by exporting the model to pbtxt (protocol buffer text mode)
and then replacing
node {
name: "dropout_1/keras_learning_phase"
op: "Placeholder"
attr {
key: "dtype"
value {
type: DT_BOOL
}
}
attr {
key: "shape"
value {
shape {
unknown_rank: true
}
}
}
}

with a Const op
node {
name: "dropout_1/keras_learning_phase"
op: "Const"
attr {
key: "dtype"
value {
type: DT_BOOL
}
}
attr {
key: "value"
value {
tensor {
dtype: DT_BOOL
tensor_shape {
}
bool_val: false
}
}
}
}

and then I was able to freeze/transform the model and use it in an android application

Sadly, the nodes that depend from this boolean values aren't stripped though. So this isn't the best solution

[Lakedaemon]

I finally found the right way to do this:

In a completely new python process (ie. not in a process that I previously used for training), I use
K.set_learning_phase(0) model = buildMyModel() model.load_weights('myWeights.hdf5') saver = tf.train.Saver() saver.save(K.get_session(), modelName)
and the model is saved without the nodes for learning/training