If the input depth is three, is the depth of the Keras Conv2d filter three?
I noticed that the Conv2d layer was used in the code where the input image has three dimensions. But since we only input a filter in two dimensions, how does matrix multiplication work?
Does the 2D filter convolve (or use broadcast) each input channel separately (and then add the results)?
Or does the depth of the filter automatically match the depth of the input (3 if it is a color image)? If this is the case, the 3x3x3 filter should have 27 weights that can be trained, not 9 in the previous case.
Tensorflow has a more explicit filter size for conv2d (you must enter height, width, channel, output channel).
https://www.tensorflow.org/api_docs/python/tf/nn/conv2d
Solution
The convolution filter of the Keras Conv2d layer automatically has n input channels, where n is the depth of the layer/number of channels before it. The previous layer is fed into the Conv2d layer as input data.
These assumptions make Keras easier to use for common use cases, such as chaining Conv2d together in deep convolutional networks.