deep-vision icon indicating copy to clipboard operation
deep-vision copied to clipboard

Published 20 hours ago • verified publisher icon ethanyanjiali

Hourglass demo: fix MPII record generation

Open peabody124 opened this issue 4 years ago • 2 comments

Also set a default of 5 for epochs as it crashed otherwise.

peabody124 avatar Oct 26 '20 06:10 peabody124

I'm not sure how the hourglass demo ran previously, but these were needed for me to have valid tfrecords with all the data needed. This is using TF 2.3.1.

I have not run a model for a long time to confirm it converges to a good value, though. What are the units on the loss and what I should expect?

peabody124 avatar Oct 26 '20 06:10 peabody124

Thanks for the PR. I didn't have time to fully test this part, and i do remember I added these extra values like scale in my other MPII implementation so I think this is indeed a necessary change.

i don't really remember the scale of a good loss tbh, but you could rely on a PCKh metrics to understand how good your model is. The loss can't reflect the real accuracy of pose estimation.

ethanyanjiali avatar Oct 31 '20 05:10 ethanyanjiali