train on coco, 100 epochs and 20% mAP

[heragong]

The result is poor. Can anyone train it sucessfully and get good results? The evaluate speed is slow when using the my train model.

[zhongdajian]

Me too, 100 epochs and 18% MAP

[zhangyilalala]

I train on coco with darknet53.conv.74 get the map of 0.2,but it seems fine when i train on my own dataset with only 1 class. Has someone train on coco with fine result? or how to fix it to train on coco?

[LifeIsSoSolong]

I try to trian model using coco data. but I come across this error : RuntimeError: reduce failed to synchronize: device-side assert triggered.

I try to solve it using #209 but no use. Do you come across this error ? thankyou

[shengyuan-tang]

me to,80 epochs and 20.7% MAP

[gaopinghai]

Me too. The loss descend slowly after 7epochs with value around 9.5 and the map remains zero.

[zhangjian-neu]

Me too. About 0.17 after 150 epochs. I used the following config: batch=16 subdivisions=2

[Flova]

What was your learning rate?

[zhangjian-neu]

There are: @Flova

[net]

Testing

#batch=1 #subdivisions=1

Training

batch=16 subdivisions=2 width=416 height=416 channels=3 momentum=0.9 decay=0.0005 angle=0 saturation = 1.5 exposure = 1.5 hue=.1

learning_rate=0.001 burn_in=0 max_batches = 500200 policy=steps steps=400000,450000 scales=.1,.1

[Flova]

Lowering the learning rate should counterintuitively improve your convergence speed. I did some tests a while ago. Lower it by a factor of 10 and try again. I also suspect a bug in the batch normalization. You could lower the momentum there from 0.9 to 0.1. this needs to be done in the models.py and is not tested by me. I would be happy if you give it a try.

[zhangjian-neu]

Well... I may have a try, depending on my work schedule. If available, results will be posted.