Where to get the trained model on Instre and imageNet?

[zhangliukun]

Thanks for your great work. The released v1-train ,v2-init,v2-train models are trained on the grozi3.2k dataset. How could we get these models trained on the Instre or the Imagenet-LOC dataset? Training one model costs too much time.

[aosokin]

Hi, I've managed to dig out some models trained on INSTRE: gdrive However, I'm not sure whether those are the correct ones and I dot not have capacity to test them. If everything is correct they should be compatible with the script launcher_instre_eval.py when unpacked in the root of OS2D (note that there are more models programmed to for evaluation whereas I've uploaded only the two best ones).

Unfortunately, I have not been able to find any models trained there.

Best, Anton

[zhangliukun]

Thanks for your reply. I have trained the resnet50 model on INSTRE-all and evaluate on the iINSTRE-s2-val, However, the GPU memery-Usage is increasing gradually and the program crash because of CUDA out of memory when the Iter is 60000(200000). My total GPU-memory is 32G and training batch 4 ,class 10. Have you encountered the same problem?

[aosokin]

I do not remember any problems like that. I did very long training runs on INSTRE on 32GB GPU and never had any memory leaks.

[aosokin]

Please also note that maybe you do not need that many iterations as the quality on validation may already be stale

[zhangliukun]

I trained os2d using the INSTRE-all and evaluate on INSTRE-S2-val，and Iter is 60000(200000). Resnet50 loss 0.0282 loc_smoothL1 1.63 cls_RLL 0.0282 cls_RLL_pos 0.02 cls_RLL_neg 0.0079 mAP0.5 0.8082 mAPw0.5 0.8208 recall 0.94 APjointclasses 0.78 the mAP0.5 0.80 is slight higher than the paper 0.777, then the program did crash again. I will try to figure it out why it is.

[aosokin]

Have you tried ladoing and evaluating the trained models I've shared above?

[zhangliukun]

I evaluate these two resnet101 weights on INSTRE-s2-val using the main.py. the mAP 0.5 are 0.8511 and 0.8751. I add a new merge NMS to deal with the large object because the large object always has lots of small bboxes. So I merge these which has a high iou bboxes and it looks good with I use the default weights trained on Grozi3.2k dataset.

And I run the demo with the resnet101 weights on some images taken by my phone. The both models may get more false positive.