Number of parameters of transposed convolution

Course Q&A Deep Learning Specialization Convolutional Neural Networks
1

I am trying to understand how to calculate the number of parameters for the transposed convolution. For example, let’s take the U-net in the second assignment of Week 3 of Course 4.

According to the dump of the model, the first transposed convolution takes as input a volume with dimensions (6, 8, 512) and outputs a volume with dimensions (12, 16, 256). In our case, the filter is 2x2. So, I would say that each filter is really 2x2x512 and there are 256 of such filters, so that we get about half a million parameters. The dump of the model says that there are more than a million parameters for this layer. Why is this the case?

2

Oh, too bad. I am getting so confused by the picture reflecting incorrectly the actual network being built. The filter being actually used for transposed convolution is 3x3, not 2x2.
Other differences I noticed:

  • the first volume is 32 channels, not 64.
  • there are dropout layers.