Near constant loss
nscotto opened this issue · comments
Does someone have managed to get results with this repository?
My loss is near constant, oscillating around 1.50.
I should mention that I have removed the instance segmentation module, but that shouldn't impact the pose estimation module.
Have you solved this? Just curious, I haven't worked with this lib before.
@AroMorin
I had more success with the other repository based on Tensorflow, so I stopped trying with this one.
I see. The other libraries usually operate with TF 1.X. The advantage of this library is that it uses Keras, so the interface doesn't change much between TF1.X and TF2.X. Also, it uses MIT License which is permissive. Others typically have restrictive licenses or none at all.
@nscotto can you tell me what other libraries worked for you? Did you measure the accuracy using COCO ?
It's the Tensorflow Personlab repository, roughly the equivalent of this one.
I don't remember the mAP value sorry, they provide pre-trained weights in case you want to try.