Numerical gradient doesn't match computed gradient

[briancheung]

This is an image of the difference between the numerical and the computed gradient. There seems to be a bug. Strangely, the differences are specific to the bits that are turned on in the image.

Either the scale is all wrong for my epsilon value in the gradient check or the test and/or ExTrain is broken.

[JesseLivezey]

Interesting, Maybe there is still a bug. How small was you step?

[briancheung]

epsilon was 1e-4. The gradients agree at areas of the image that have no 'on' pixel values.

Maybe we should be calculating the expected flow and then taking the derivative of that rather than calculating the expected derivative of the flow.

[JesseLivezey]

Maybe we should be calculating the expected flow and then taking the
derivative

That is what we should be doing, although I'm not sure they aren't the same
thing... I'm heading into campus soon.

On Mon, May 5, 2014 at 9:38 AM, Brian Cheung [email protected]:

epsilon was 1e-4. The gradients agree at areas of the image that have no
'on' pixel values.

Maybe we should be calculating the expected flow and then taking the
derivative of that rather than calculating the expected derivative of the
flow.

—
Reply to this email directly or view it on GitHubhttps://github.com//issues/13#issuecomment-42208079
.