Train on own data

[Palushok]

Hello. I am trying to fit this network on my data, but hm_loss down to 9.208498 and don't improve. I tried to use different samples of dataset but result was same. After that i download tiny_coco and anyway hm_loss == 9.208498. Can you help me with this?

2020-09-30-17-53: epoch: 1 |loss 110.024204 | hm_loss 109.175058 | wh_loss 8.276505 | off_loss 0.870643 | time 0.250000 | 
2020-09-30-17-53: epoch: 2 |loss 10.162599 | hm_loss 9.092366 | wh_loss 8.377615 | off_loss 1.302705 | time 0.033333 | 
2020-09-30-17-54: epoch: 3 |loss 9.840244 | hm_loss 9.208498 | wh_loss 6.206020 | off_loss 0.642890 | time 0.033333 | 
2020-09-30-17-54: epoch: 4 |loss 9.729698 | hm_loss 9.208498 | wh_loss 6.791722 | off_loss 0.363228 | time 0.033333 | 
2020-09-30-17-54: epoch: 5 |loss 9.686012 | hm_loss 9.208498 | wh_loss 6.189690 | off_loss 0.336059 | time 0.033333 | loss 16.727993 | hm_loss 9.208498 | wh_loss 28.976759 | off_loss 12.141313 | time 0.000000 | 
2020-09-30-17-54: epoch: 6 |loss 9.660375 | hm_loss 9.208498 | wh_loss 5.394241 | off_loss 0.364330 | time 0.033333 | 
2020-09-30-17-54: epoch: 7 |loss 9.728366 | hm_loss 9.208498 | wh_loss 7.177228 | off_loss 0.322013 | time 0.033333 | 
2020-09-30-17-54: epoch: 8 |loss 9.784336 | hm_loss 9.208498 | wh_loss 7.994667 | off_loss 0.352210 | time 0.033333 | 
2020-09-30-17-54: epoch: 9 |loss 9.674562 | hm_loss 9.208498 | wh_loss 6.191436 | off_loss 0.312984 | time 0.033333 | 

[cao-nv]

I tried to train on my custom dataset which had only bounding box for objects not masks, but I couldn't. It seemed that this repo made use of instance masks annotations to refine bounding box for partly visible objects.

[lucasjinreal]

@cao-nv this repo changes nothing than original centernet, except adding a HarDNet backbone

[seungtaek94]

i met same problem when i trained my own dateaset. my hm_val_loss also stop at 9.208 :(

[cao-nv]

@cao-nv this repo changes nothing than original centernet, except adding a HarDNet backbone

When I just add the hardnet file in this repo to the original CenterNet, everything works well.