There's at least one existing paper about this idea (https://arxiv.org/abs/1511.05641 https://arxiv.org/abs/1511.05641). Also, it is possible to initialize a co

There's at least one existing paper about this idea (https://arxiv.org/abs/1511.05641 https://arxiv.org/abs/1511.05641). Also, it is possible to initialize a convolutional layer so that it passes through its input, but initializing the weights properly requires a little more work than calling tf.keras.initializers