Problem with training YOLOv2

[mrkieumy]

Hi @andy-yun
I trained normally with yoloV3, tinyV3, tinyV2. But with YoloV2 model, it raises this error:

Traceback (most recent call last):
File "train.py", line 626, in
main()
File "train.py", line 202, in main
nsamples = train(epoch)
File "train.py", line 307, in train
ol=l(output[i]['x'], target)
File "/home/kieumy/anaconda3/lib/python3.6/site-packages/torch/nn/modules/module.py", line 489, in call
result = self.forward(*input, **kwargs)
File "/home/kieumy/YOLO/pytorch_conditioning/region_layer.py", line 172, in forward
loss_cls = self.class_scale * nn.CrossEntropyLoss(reduction='sum')(cls, tcls)/nB
File "/home/kieumy/anaconda3/lib/python3.6/site-packages/torch/nn/modules/module.py", line 489, in call
result = self.forward(*input, **kwargs)
File "/home/kieumy/anaconda3/lib/python3.6/site-packages/torch/nn/modules/loss.py", line 904, in forward
ignore_index=self.ignore_index, reduction=self.reduction)
File "/home/kieumy/anaconda3/lib/python3.6/site-packages/torch/nn/functional.py", line 1970, in cross_entropy
return nll_loss(log_softmax(input, 1), target, weight, None, ignore_index, None, reduction)
File "/home/kieumy/anaconda3/lib/python3.6/site-packages/torch/nn/functional.py", line 1790, in nll_loss
ret = torch._C._nn.nll_loss(input, target, weight, _Reduction.get_enum(reduction), ignore_index)
RuntimeError: invalid argument 2: non-empty vector or matrix expected at /pytorch/aten/src/THCUNN/generic/ClassNLLCriterion.cu:32

Do you know the reason why?
Thanks.

[andy-yun]

@mrkieumy you should check the tcls value (tcls is target). I am guessing that tcls is empty. It means that there is not assigned target value.

[mrkieumy]

@andy-yun, Thanks, you mean tcls in yolo_layer.py or in region_layer?
I'm sorry I don't understand exactly those tcls value. Is there the difference between V2 and V3,tinyV3,tinyV2 to make us fix the tcls? I think if it wrong with yolov2, it must be wrong with tinyv2 too. But the code runs normally with tinyv2. Do you know exactly why?
Thanks.

[andy-yun]

@mrkieumy tcls means class information from ground truth value. if you adopt yolov2, then tcls at region_layer.py is used, else tcls of yolo_layer.py is used. tcls at region_layer.py is compared at CrossEntropyLoss, tcls at yolo_layer.py is compared at BCELoss.

[PurpleMStone]

Hi, I suffer from the same error when training on VOC dataset, and i think tcls is not empty. I need help. Thanks.

[andy-yun]

@mrkieumy @PurpleMStone Both you should check config and data files. The codes are well worked on coco and voc dataset.

[PurpleMStone]

There may be some bugs in function data_augmentation_nocrop in image.py? When i set the crop=True, things turn out be right.