voc trainning convergence problem about mxnet-yolo

[zhreshold]

try dev branch with lr=0.0001

[vsooda]

@zhreshold thank you very much. I change the learning rate as you said. It is much better.

I didn't use pretrain model, and after 200 epoch, mAP is 0.51.

INFO:root:Epoch[200] Train-Acc=0.882000
INFO:root:Epoch[200] Train-IOU=0.685558
INFO:root:Epoch[200] Train-BG-score=0.004245
INFO:root:Epoch[200] Train-Obj-score=0.256506
INFO:root:Epoch[200] Time cost=297.076
INFO:root:Saved checkpoint to "/home/sooda/deep/object/mxnet-yolo/model/yolo2_416-0201.params"
INFO:root:Epoch[200] Validation-aeroplane=0.595343
INFO:root:Epoch[200] Validation-bicycle=0.642839
INFO:root:Epoch[200] Validation-bird=0.428441
INFO:root:Epoch[200] Validation-boat=0.413425
INFO:root:Epoch[200] Validation-bottle=0.204297
INFO:root:Epoch[200] Validation-bus=0.623591
INFO:root:Epoch[200] Validation-car=0.685286
INFO:root:Epoch[200] Validation-cat=0.599618
INFO:root:Epoch[200] Validation-chair=0.307342
INFO:root:Epoch[200] Validation-cow=0.450331
INFO:root:Epoch[200] Validation-diningtable=0.528938
INFO:root:Epoch[200] Validation-dog=0.531340
INFO:root:Epoch[200] Validation-horse=0.639920
INFO:root:Epoch[200] Validation-motorbike=0.583562
INFO:root:Epoch[200] Validation-person=0.632997
INFO:root:Epoch[200] Validation-pottedplant=0.254757
INFO:root:Epoch[200] Validation-sheep=0.497984
INFO:root:Epoch[200] Validation-sofa=0.472311
INFO:root:Epoch[200] Validation-train=0.686695
INFO:root:Epoch[200] Validation-tvmonitor=0.527394
INFO:root:Epoch[200] Validation-mAP=0.515321

[vsooda]

@zhreshold after 240 epochs, the mAP drop to 0.48. the highest mAP is 0.52. I only use one gpu. how can I get a better result? use a pretrained model?

Thank you.

[zhreshold]

A pretrained model on imagenet is prefered

[vsooda]

Thanks, I will have a try.