About model performance after training

[lucaskyle]

i modified some lines now it can load lightning movenet model i try your training code only using COCO dataset

training sample : 64115 validation sample : 5000

when i try finetuning published lightning movenet, the loss will be like: --->train|loss 4.476185 | hm_loss 1.625346 | hp_loss 1.318379 | hm_hp_loss 1.305676 | hp_offset_loss 0.226784 |​ --->valid| loss 3.907756 | hm_loss 1.076794 | hp_loss 0.682829 | hm_hp_loss 2.044044 | hp_offset_loss 0.104088 |

when i train the lightning movenet from scratch, the loss will be like: --->train|loss 8.478377 | hm_loss 2.583637 | hp_loss 2.864822 | hm_hp_loss 2.782381 | hp_offset_loss 0.247538 | ​ --->valid| loss 7.517864 | hm_loss 1.706246 | hp_loss 1.409458 | hm_hp_loss 4.288480 | hp_offset_loss 0.113680 |​ is it good?​

then i try show demo results with both models, the model trained from scratch can not even detect key points.

[lee-man]

hi, @lucaskyle This repo is not longer maintained. Sorry that I cannot offer valuable feedback to you.

[lucaskyle]

@lee-man understood, thank u