I have a (trained) tf.keras model which I would like to convert to a quantized model and retrain with tensorflow's fake quant strategy (using python as frontend).
Could I somehow apply tf.contrib.quantize.create_training_graph directly to the keras model (graph) then retrain? Seems like there's some problem with the fact that the session is already created when taking the graph from K.get_session().graph.
For example, the following approach:
import tensorflow.contrib.lite as tflite
keras_graph = tf.keras.backend.get_session().graph
from tensorflow.contrib.quantize import create_training_graph
create_training_graph(input_graph=keras_graph,
quant_delay=int(0*(len(X_train) / batch_size)))
...
model.compile(...)
model.fit_generator(...)
results with the message: "Operation '{name:'act_softmax/sub' id:2294 op device:{} def:{{{node act_softmax/sub}} = Sub[T=DT_FLOAT](conv_preds/act_quant/FakeQuantWithMinMaxVars:0, act_softmax/Max)}}' was changed by updating input tensor after it was run by a session. This mutation will have no effect, and will trigger an error in the future. Either don't modify nodes after running them or create a new session."
And true enough, the error: tensorflow.python.framework.errors_impl.FailedPreconditionError: Attempting to use uninitialized value conv_preds/act_quant/conv_preds/act_quant/max/biased
(i.e. create_training_graph needs the graph before the session was created? is it possible to get the graph from a keras model before the session was instantiated?)
Alternatively, if this doesn't work, could I convert the (h5) model to a checkpoint, then somehow load the model from this checkpoint to a tensorflow graph and continue working with pure tensorflow?
Would appreciate any help or pointers. Thank you!
