gradient of the LP problem about deepemd HOT 3 CLOSED

hi, thank you for your interests in our work.
it is a good observation from your demo code with a toy example, but for our task the generated flow must be used together with the similarity matrix to generate the final score, which is further used for cross-entropy loss. if you directly sum the flow, the grad back-propagated does seem very small in the example.
I just do a test where I add a detach() to the similarity maxtrix in

from deepemd.

you may also consider it from another view that, if you sum the flow as the loss, you are actually encouraging a small overall flow. however, the total flow is a fixed value, so the gradients generated from it is meaningless thus closed to zero.

from deepemd.

Thank you for your reply! I now get the point.

from deepemd.