Question about Trainable Weights in Section 3.4 of GIB Paper

Thank you for your informative paper on Generative Image Bidding (GIB). I found it very helpful in understanding the concepts. While reviewing the code implementation, I noticed a difference between Section 3.4 of the paper and the actual code. In Section 3.4,  However, in the code, the weights seem to be initialized with a value and then trained. I was wondering if there's a specific reason for this difference. Why in the implementation, you just make it as initial weight filling and let it trainable? Should not be frozen weights? Best regards.