Long Chen

Thank you for pointing out this. I read the darknet code again and think that you are right. The darknet uses anchors to match gt_boxes and pred_boxes (https://github.com/pjreddie/darknet/blob/master/src/region_layer.c#L264) but uses...

I do it later in https://github.com/longcw/yolo2-pytorch/blob/master/darknet.py#L227

Thank you for pointing out this bug. I'll fix this and try to train it later.

@ztyxd You can download it from https://cloud.tsinghua.edu.cn/f/a863f19ef5c84a59bc9b/

Of course, you can train YOLO on coco dataset. You first have to write a config file for coco just like [yolo2-pytorch/cfgs/config_voc.py](https://github.com/longcw/yolo2-pytorch/blob/master/cfgs/config_voc.py), and then implement a data loader for coco...

This may caused by `multiprocessing.Pool.imap`. You can use the dataloader from https://github.com/longcw/yolo2-pytorch/issues/71

Yes, you need to rewrite the loss function outside the model. `DataParallel` will duplicate your model to run on multiple gpus, so that you can not access a member variable...

Did you run `make.sh` successfully?

It seems `utils/build.py` cannot find you cuda lib https://github.com/longcw/yolo2-pytorch/blob/master/utils/build.py#L29 You can set it in the code or in env `CUDAHOME`.

Or you can get the scalar using `net.bbox_loss.item()` in pytorch0.4