Add Support for Deconvolution Layer for Keras model

[dsmilkov]

From @clover113 on April 2, 2018 8:9

Currently, when converting keras model with deconvolution2d layer, the converter went fine. However, when importing the model in Tensorflow.js, it outputs "Error: Unknown layer: Conv2DTranspose".
Also, conv2dtranspose is supported in tensorflow.js.

Copied from original issue: tensorflow/tfjs-converter#67

[dsmilkov]

From @carlthome on April 2, 2018 9:56

Similarly when going through the SavedModel converter with tf.layers.conv2d_transpose you'll end up with

Unsupported Ops in the model
StridedSlice

[dsmilkov]

From @clover113 on April 2, 2018 9:59

Yeah, it seems that tensorflow.js's support for layers is too limited. But it is not good not to have Conv2dTransposed layer.

[dsmilkov]

From @nsthorat on April 2, 2018 13:37

@caisq can we add this layer next? The op exists in core in forward mode (we'll eventually get to the backprop for it but it's non-trivial).

[dsmilkov]

From @caisq on April 2, 2018 13:58

@clover113 Thanks for filing the bug! We can add this layer, but it'll support inference first, without supporting training because the gradients for conv2DTranspose is still missing.

Now the deconvolution itself seems to work. But I still cannot import the model. It says: "Error: Could not find weights in manifest with names: deconvolution/kernel_1". I only have one deconvolution layer and I import the model using tf.loadModel() only once. Other layers are just fine, so I think the implementation of importing keras model with deconvolution layer is still broken.

[caisq]

@clover113 Can you let us know what Keras model you are trying to import, so we can try to reproduce the error on our end?

[oeway]

+1 support training for conv2DTranspose

Currently, it seems Conv2DTranspose do not have a gradient function for training, I got an error like this:

Uncaught (in promise) Error: Cannot compute gradient: gradient function not found for conv2dDerInput.