Model Parameters and size

[R7788380]

Hi, thanks for your sharing code.

According to your code, the numbers of parameters in separation layer in DPTNet is 2,800,769 (2.8M), and my hyperparamters show below:

N=64 # the number of filters in encoder
C=2 # the number of speechs
L=128 the number of Feedforward dimension in Transformer
H=4 # the number of heads in Transfomer
K=250 # Segmentation chunk size
B=6 # the number of Dual-Path Blocks

I am so confuse about the model size in the original paper, the whole model size in paper is 2.69MB, but the separation layer here is 10.68MB (suppose 32 bits per param, 2,800,769*32/8/1024/1024=10.68MB), the size of separation layer far exceeds that of whole model in paper.

I found your issue, and I would like to know if you have the same concerns, maybe you had already solved this problem.