log of training about posefix_release HOT 13 CLOSED

This is very hard to see. Please summarize your problem.

from posefix_release.

During the training process, the train loss did't converge and the effectivity of the algorithm became worser and worser

from posefix_release.

some time the loss become very large:

and some time being negative:

from posefix_release.

Hmm. I haven't met that problem. Which dataset are you using?

from posefix_release.

COCO

from posefix_release.

Your downloaded code is the most recent version?

from posefix_release.

Did you change

    ## dataset
    dataset = 'PoseTrack' # 'COCO', 'PoseTrack', 'MPII'
    testset = 'val' # train, test, val (there is no validation set for MPII)

to your desired settings? (COCO, train)

from posefix_release.

I downloaded the code on the 8th of August, I suppose it is the latest code

from posefix_release.

dataset = 'COCO' # 'COCO', 'PoseTrack', 'MPII'
testset = 'val' # train, test, val (there is no validation set for MPII)

I did't change the testset and suppose it is a params only for testing process

from posefix_release.

Right. It seems the number of iteration for one epoch is smaller than mine. Do you use other version of COCO? Anyway, I haven't met this problem, so I cannot give you clear answer. Sorry. I seems there is some issue with heatmap loss. Could you debug yourself and let me know any progress? Thank you!

One thing I want to check is your trained model gives bad performance because of this issue?

from posefix_release.

ok, as i have modified the code as discussion in issue #8 to fixed the bug of InvalidArgumentError
Could you please provide me with the change of loss during the training process?

from posefix_release.

Sorry I don't keep training log :(

from posefix_release.

Never mind, thanks for your patience!

from posefix_release.