Size mismatch during the forward-propagation?

[zjysteven]

Hi! First, thank you for this well-designed benchmark!

However, I have a question about the tensor size before and after the CNN. The input of CNN is (Cin x T x 1024) (I'm using the channel-first notion for convenience) and there are three max-pooling operations with the kernel size of 8, 8, 4 along the last dimension. I expect this will lead the output of CNN to be (Cout x T x 4), which is inconsistent with (Cout x T x 2) as illustrated in your image of the network architecture. Please correct me if I miss something.

Besides, I'm wondering why you choose to use only a one-layer linear projection without any non-linear activation for the DOA prediction? Is it related to the performance according to your experiments? Thanks!

[sharathadavanne]

Hi @zjysteven thanks for pointing the error. You are right, the CNN output dimension should be (Cout x T x 4). I will update the image soon.

I remember trying ReLU and tanh activations in the penultimate layer of the DOA output, but I dont think I got good results. Similarly, more number of fully-connected layers didn't help either. Both these studies were done for Cartesian DOA output as discussed in the original SELDnet paper, and I used the same model here with only one change, i.e., Spherical DOA coordinates output. So I am not sure if the number of fully-connected layers and different activations would help for Spherical coordinates.

[zjysteven]

@sharathadavanne Thanks for your quick response! That makes sense since while I am reproducing the baseline, I used more number of fully-connected layers together with ReLU activation, but I couldn't achieve similar DOA results to yours. Thought I might suffer from over-fitting so I asked the above question. Thanks again for your confirmation!