Non local model does not yield better results

[hnyz979]

Dear nnUyi:
Thank you for your code, which provides us a clear and concise of the Non Local Model.
However, I have trained the network with and without Non Local Model for comparison. For the network with Non Local Model, I use your code; for the network without None Local Model, I just disabled the "nonlocal_block1 = NonLocalBlock(cnv1_pool, 32, scope='nonlocal_block1')" and "nonlocal_block2 = NonLocalBlock(cnv2_pool, 64, scope='nonlocal_block2')". The results show that both training and testing accuracy are higher WITHOUT Non Local Model.
Have you experienced this strange circumstance?
Thanks again!

[hnyz979]

PS I then find that you modified your codes by updating reshape ops in non-local block. However, this does not provide better performance...