VXallset
commented
3 years ago Hello, I found that in the definition of
conv_2d, src/util.py, kernel_initializer will be created if no initializer is given explicitly. Moreover,conv_2dis called repeatedly indown_sampling, here. To avoid recreating rebundant nodes in tf computation graph, I think the kernel_initializer should be created before the loop indown_sampling, and pass it toconv_2d.Looking forward to your reply. Btw, I am very glad to create a PR to fix it if you are too busy.
Hi @JamesCao2048 ,
Sorry for the late reply that I was too busy recently and didn't notice this issue.
For the problem that you mentioned, I think that down_sampling part is necessary. In HRNet, there are x2, x4, x8 times down-sampling. You will need to write them separately if you don't use the while-loop. What's more, kernel_initializer in each conv_2d layer should be different and that's why I didn't create it and pass it to conv_2d.
That's my personal opinion and not sure whether I am correct.
Thanks again.
Hello, I found that in the definition of
conv_2d, src/util.py, kernel_initializer will be created if no initializer is given explicitly. Moreover,conv_2dis called repeatedly indown_sampling, here. To avoid recreating rebundant nodes in tf computation graph, I think the kernel_initializer should be created before the loop indown_sampling, and pass it toconv_2d.Looking forward to your reply. Btw, I am very glad to create a PR to fix it if you are too busy.