Wrong gradient computation through loops

[Clemens Schwarke]

Hi,

when differentiating through a loop, the gradient seems to be wrong if the loop overwrites a variable. In my understanding, this is caused by the fact that the replay of the backward pass skips the loop. Therefore, putting the loop in a seperate function solves the issue, as the function is called again in the replay pass.

Looking forward to your opinions on this! (Unfortunately, I don't have time for a minimal example at the moment, but I am happy to clarify my issue if this post is confusing)

Cheers,
Clemens

[Nicolas Capens]

Sorry for the late reply!

when differentiating through a loop, the gradient seems to be wrong if the loop overwrites a variable.

In general overwriting a variable in a loop makes it non-differentiable. There should be a warning produced about this.

If the iteration count is constant and small (by default up to 16), we attempt to unroll the loop, and this makes it differentiable.