Request for tutorial on how to convert a channel first Keras model to channel last model? #467
Description
Hello,
I have a pre-trained Keras model (in h5 format) where all the layers operate on channel first data format. I want to convert this model to operate on the channel last data format (the default data format).
For some clarity, the current model summary looks like this:
Model: "model"
__________________________________________________________________________________________________
Layer (type) Output Shape Param # Connected to
==================================================================================================
input0 (InputLayer) [(None, 3, 240, 320) 0
__________________________________________________________________________________________________
259_pad (ZeroPadding2D) (None, 3, 242, 322) 0 input0[0][0]
__________________________________________________________________________________________________
259 (Conv2D) (None, 16, 120, 160) 432 259_pad[0][0]
__________________________________________________________________________________________________
260 (BatchNormalization) (None, 16, 120, 160) 64 259[0][0]
__________________________________________________________________________________________________
261 (Activation) (None, 16, 120, 160) 0 260[0][0]
__________________________________________________________________________________________________
262_pad (ZeroPadding2D) (None, 16, 122, 162) 0 261[0][0]
__________________________________________________________________________________________________
262 (DepthwiseConv2D) (None, 16, 120, 160) 144 262_pad[0][0]
As you can see, it's in the channel first format. I want to convert each layer in the model to operate on channel last format. So the ideal model summary will be as follows:
Model: "model"
__________________________________________________________________________________________________
Layer (type) Output Shape Param # Connected to
==================================================================================================
input0 (InputLayer) [(None, 240, 320, 3) 0
__________________________________________________________________________________________________
259_pad (ZeroPadding2D) (None, 242, 322, 3) 0 input0[0][0]
__________________________________________________________________________________________________
259 (Conv2D) (None, 120, 160, 16) 432 259_pad[0][0]
__________________________________________________________________________________________________
260 (BatchNormalization) (None, 120, 160, 16) 64 259[0][0]
__________________________________________________________________________________________________
261 (Activation) (None, 120, 160, 16) 0 260[0][0]
__________________________________________________________________________________________________
262_pad (ZeroPadding2D) (None, 122, 162, 16) 0 261[0][0]
__________________________________________________________________________________________________
262 (DepthwiseConv2D) (None, 120, 160, 16) 144 262_pad[0][0]
I raised an issue at TensorFlow and from a bit of searching I understood that this situation is not uncommon.
From this comment, I understood that we have to do a "network surgery" to accomplish this.
The above solution works when the training files were available. In the case where I had only the output of tf.estimator.export_savedmodel, I had to do a network surgery. I made a clone of the graph, in channel_last format, loaded it in and assigned all variables, values from the trained model. This new graph runs fine on CPU.
It would be great if you could provide a sample code on how to do this.
Any help/pointers are really appreciated.
Thank you.