Keras converter. Lambda layer with multiple inputs.

[shamangary]

Hi. I encounter an issue when lambda layer receives multiple inputs.

# This is a custom activation function.
def swish(x,s1):
    return K.sigmoid(x)*x

def merge(x):
    return x[0]+x[1]

# Create a silly model that has our custom activation function as a new layer.
def create_model():
    a = 3
    inp = Input(shape=(256, 256, 3))
    x = Conv2D(6, (3, 3), padding="same")(inp)
    #x = Activation(swish)(x)                   # doesn't work! :-(
    x = Lambda(swish, arguments={'s1':a})(x)
    x = GlobalAveragePooling2D()(x)
    x0 = Dense(10, activation="softmax")(x)
    x1 = Dense(10, activation="softmax")(x0)
    x = Lambda(merge)([x0,x1])
    return Model(inp, x)

The conversion works fine when lambda layer only has one input. However, x = Lambda(merge)([x0,x1]) causes the following problem.

Traceback (most recent call last):
  File "test.py", line 98, in <module>
    custom_conversion_functions={ "Lambda": convert_lambda })
  File "/home/shamangary/anaconda3/envs/new/lib/python3.5/site-packages/coremltools/converters/keras/_keras_converter.py", line 745, in convert
    custom_conversion_functions=custom_conversion_functions)
  File "/home/shamangary/anaconda3/envs/new/lib/python3.5/site-packages/coremltools/converters/keras/_keras_converter.py", line 543, in convertToSpec
    custom_objects=custom_objects)
  File "/home/shamangary/anaconda3/envs/new/lib/python3.5/site-packages/coremltools/converters/keras/_keras2_converter.py", line 191, in _convert
    graph.build()
  File "/home/shamangary/anaconda3/envs/new/lib/python3.5/site-packages/coremltools/converters/keras/_topology2.py", line 643, in build
    successors = self.edge_map[layer]
KeyError: 'lambda_2'

I was trying much more complex merging function (with multiple different dim features) other than just addition.
Do you know how to resolve this? Thank you!

[aseemw]

Can you try changing this function locally to the following and see if the error persists?

def _get_first_shared_layer(self):
        for idx, layer in enumerate(self.layer_list):
            keras_layer = self.keras_layer_map[layer]
            if not _is_merge_layer(self.keras_layer_map[layer]) and \
                len(self.get_predecessors(layer)) > 1 and \
                len(keras_layer._inbound_nodes) > 1:
                return idx
        return -1

[shamangary]

Hello. The conversion works :). Despite I still have to write a custom swift file for my layer for the ios app, I guess the conversion is working now. Many thanks.

[visionscaper]

@aseemw Has the above fix been merged in the code?

[aseemw]

Yes.