Additional ONNX Operators for Keras Convolution1D

[ThomasManz]

Hello,

I´d like to use tf2onnx to convert Keras model, because the old keras2onnx project seems to be stopped.
I´ve realized that a simple Convolution1D network is translated in different ways, the code is simple:

model = Sequential()
model.add(Convolution1D(filters=4, kernel_size=5, strides=3, activation='tanh', input_shape=(10, 3), name='CL1',
                            bias_initializer='random_uniform', padding='same'))
model.add(keras.layers.Flatten())
tf2onnx.convert.from_keras(model, ...)

While keras2onnx does this rather simple, tf2onnx creates some additional operators (see picture: left=tf2onnx, right=keras2onnx): it creates a further dimension with Unsqueeze/Squeeze and also does not use the parameter 'bias' in Conv but adds a further Add.

Can you explain why it´s done this more complex way?
Are there any plans to implement a simpler transformation of Conv1D?

Thanks for your great work
Thomas