RuntimeError: shape '[1, 3, 29, 76, 76]' is invalid for input of size 1472880

[DongChen06]

Hello, When I started training on my own dataset, it shows error like.

log file path:log\log_2020-07-02_23-18-42.txt
2020-07-02 23:18:42,322 train.py[line:468] INFO: Using device cuda
convalution havn't activate linear
convalution havn't activate linear
convalution havn't activate linear
2020-07-02 23:18:44,634 train.py[line:318] INFO: Starting training:
        Epochs:          300
        Batch size:      16
        Subdivisions:    16
        Learning rate:   0.001
        Training size:   144
        Validation size: 16
        Checkpoints:     True
        Device:          cuda
        Images size:     608
        Optimizer:       adam
        Dataset classes: 24
        Train label path:data/train.txt
        Pretrained:
    
Epoch 1/300:   0%|       | 0/144 [00:12<?, ?img/s]
Traceback (most recent call last):
  File "C:/Users/Windows/Downloads/pytorch-YOLOv4-master/train.py", line 483, in <module>
    device=device, )
  File "C:/Users/Windows/Downloads/pytorch-YOLOv4-master/train.py", line 356, in train
    loss, loss_xy, loss_wh, loss_obj, loss_cls, loss_l2 = criterion(bboxes_pred, bboxes)
  File "C:\Softwares\Anaconda3\envs\torch\lib\site-packages\torch\nn\modules\module.py", line 550, in __call__
    result = self.forward(*input, **kwargs)
  File "C:/Users/Windows/Downloads/pytorch-YOLOv4-master/train.py", line 233, in forward
    output = output.view(batchsize, self.n_anchors, n_ch, fsize, fsize)
RuntimeError: shape '[1, 3, 29, 76, 76]' is invalid for input of size 1472880

Does anyone face the same problem?

[Tianxiaomo]

You can choose how to build the model.

pytorch-YOLOv4/cfg.py

Lines 17 to 18 in af00822

	Cfg.use_darknet_cfg = True
	Cfg.cfgfile = 'cfg/yolov4.cfg'

if you use cfg file,

Cfg.use_darknet_cfg = True
Cfg.cfgfile = 'cfg file'

and you need to change the class parameter of the yolo layer of the cfg file according to your data set.

pytorch-YOLOv4/cfg/yolov4-custom.cfg

Lines 966 to 980 in af00822

	[yolo]
	mask = 0,1,2
	anchors = 12, 16, 19, 36, 40, 28, 36, 75, 76, 55, 72, 146, 142, 110, 192, 243, 459, 401
	classes=80
	num=9
	jitter=.3
	ignore_thresh = .7
	truth_thresh = 1
	scale_x_y = 1.2
	iou_thresh=0.213
	cls_normalizer=1.0
	iou_normalizer=0.07
	iou_loss=ciou
	nms_kind=greedynms
	beta_nms=0.6

pytorch-YOLOv4/cfg/yolov4-custom.cfg

Lines 1053 to 1068 in af00822

	[yolo]
	mask = 3,4,5
	anchors = 12, 16, 19, 36, 40, 28, 36, 75, 76, 55, 72, 146, 142, 110, 192, 243, 459, 401
	classes=80
	num=9
	jitter=.3
	ignore_thresh = .7
	truth_thresh = 1
	scale_x_y = 1.1
	iou_thresh=0.213
	cls_normalizer=1.0
	iou_normalizer=0.07
	iou_loss=ciou
	nms_kind=greedynms
	beta_nms=0.6

pytorch-YOLOv4/cfg/yolov4-custom.cfg

Lines 1140 to 1156 in af00822

	[yolo]
	mask = 6,7,8
	anchors = 12, 16, 19, 36, 40, 28, 36, 75, 76, 55, 72, 146, 142, 110, 192, 243, 459, 401
	classes=80
	num=9
	jitter=.3
	ignore_thresh = .7
	truth_thresh = 1
	random=1
	scale_x_y = 1.05
	iou_thresh=0.213
	cls_normalizer=1.0
	iou_normalizer=0.07
	iou_loss=ciou
	nms_kind=greedynms
	beta_nms=0.6

[SenWang-NEU]

Hello, When I started training on my own dataset, it shows error like.

log file path:log\log_2020-07-02_23-18-42.txt
2020-07-02 23:18:42,322 train.py[line:468] INFO: Using device cuda
convalution havn't activate linear
convalution havn't activate linear
convalution havn't activate linear
2020-07-02 23:18:44,634 train.py[line:318] INFO: Starting training:
        Epochs:          300
        Batch size:      16
        Subdivisions:    16
        Learning rate:   0.001
        Training size:   144
        Validation size: 16
        Checkpoints:     True
        Device:          cuda
        Images size:     608
        Optimizer:       adam
        Dataset classes: 24
        Train label path:data/train.txt
        Pretrained:
    
Epoch 1/300:   0%|       | 0/144 [00:12<?, ?img/s]
Traceback (most recent call last):
  File "C:/Users/Windows/Downloads/pytorch-YOLOv4-master/train.py", line 483, in <module>
    device=device, )
  File "C:/Users/Windows/Downloads/pytorch-YOLOv4-master/train.py", line 356, in train
    loss, loss_xy, loss_wh, loss_obj, loss_cls, loss_l2 = criterion(bboxes_pred, bboxes)
  File "C:\Softwares\Anaconda3\envs\torch\lib\site-packages\torch\nn\modules\module.py", line 550, in __call__
    result = self.forward(*input, **kwargs)
  File "C:/Users/Windows/Downloads/pytorch-YOLOv4-master/train.py", line 233, in forward
    output = output.view(batchsize, self.n_anchors, n_ch, fsize, fsize)
RuntimeError: shape '[1, 3, 29, 76, 76]' is invalid for input of size 1472880

Does anyone face the same problem?

Did you solve it? According to the author's reply, I still receive the error "RuntimeError: shape'[1, 3, 8, 76, 76]' is invalid for input of size 1472880"

[DongChen06]

@Tianxiaomo, Thanks. Do we need to edit the parameters like anchors when we use custom dataset?

@SenWang-NEU I solved the problem by choosing Cfg.use_darknet_cfg = False and use the Cfg configure.

I trained the model for about 10 hours by 2x 2080 Ti, The training courses look like this,

The losses reduce a lot, while still remain relatively big values like,
Message: 'Train step_46880: loss : 147.75306701660156,loss xy : 24.71426010131836,loss wh : 1.184415578842163,loss obj : 88.55791473388672，loss cls : 33.29647445678711,loss l2 : 20.337644577026367,lr : 0.001'

@Tianxiaomo I have one question, what is the acceptable loss range for us to get good inference results?

[SenWang-NEU]

@Tianxiaomo，谢谢。 使用自定义数据集时，是否需要像锚一样编辑参数？

@ SenWang-NEU我通过选择 Cfg.use_darknet_cfg = False并使用Cfg配置解决了该问题。

我用2x 2080 Ti训练了大约10个小时的模型，训练课程看起来像这样，

损失减少了很多，但仍然保持相对较大的价值，例如，
Message: 'Train step_46880: loss : 147.75306701660156,loss xy : 24.71426010131836,loss wh : 1.184415578842163,loss obj : 88.55791473388672，loss cls : 33.29647445678711,loss l2 : 20.337644577026367,lr : 0.001'

@Tianxiaomo我有一个问题，对我们而言，获得良好推断结果的可接受损失范围是多少？

Thanks, based on your reply, I solved the problem and the model has started training.

[Tianxiaomo]

@Tianxiaomo, Thanks. Do we need to edit the parameters like anchors when we use custom dataset?

@SenWang-NEU I solved the problem by choosing Cfg.use_darknet_cfg = False and use the Cfg configure.

I trained the model for about 10 hours by 2x 2080 Ti, The training courses look like this,

The losses reduce a lot, while still remain relatively big values like,
Message: 'Train step_46880: loss : 147.75306701660156,loss xy : 24.71426010131836,loss wh : 1.184415578842163,loss obj : 88.55791473388672，loss cls : 33.29647445678711,loss l2 : 20.337644577026367,lr : 0.001'

@Tianxiaomo I have one question, what is the acceptable loss range for us to get good inference results?

The validation section code has been added, and you can infer in the training is to see when it works best.

[DongChen06]

I am not so sure why I need to comment use_darknet_cfg to make the code run correctly?
` ``
Cfg.use_darknet_cfg = False
Cfg.cfgfile = 'cfg/yolov4-custom.cfg'

And change the classes in **yolov4-custom.cfg**

[Tianxiaomo]

pytorch-YOLOv4/train.py

Lines 607 to 612 in 74347ac

	if cfg.use_darknet_cfg:
	model = Darknet(cfg.cfgfile)
	else:
	model = Yolov4(cfg.pretrained, n_classes=cfg.classes)

[Juuustin]

You can choose how to build the model.

pytorch-YOLOv4/cfg.py

Lines 17 to 18 in af00822

	Cfg.use_darknet_cfg = True
	Cfg.cfgfile = 'cfg/yolov4.cfg'

if you use cfg file,

Cfg.use_darknet_cfg = True
Cfg.cfgfile = 'cfg file'

and you need to change the class parameter of the yolo layer of the cfg file according to your data set.

pytorch-YOLOv4/cfg/yolov4-custom.cfg

Lines 966 to 980 in af00822

	[yolo]
	mask = 0,1,2
	anchors = 12, 16, 19, 36, 40, 28, 36, 75, 76, 55, 72, 146, 142, 110, 192, 243, 459, 401
	classes=80
	num=9
	jitter=.3
	ignore_thresh = .7
	truth_thresh = 1
	scale_x_y = 1.2
	iou_thresh=0.213
	cls_normalizer=1.0
	iou_normalizer=0.07
	iou_loss=ciou
	nms_kind=greedynms
	beta_nms=0.6

pytorch-YOLOv4/cfg/yolov4-custom.cfg

Lines 1053 to 1068 in af00822

	[yolo]
	mask = 3,4,5
	anchors = 12, 16, 19, 36, 40, 28, 36, 75, 76, 55, 72, 146, 142, 110, 192, 243, 459, 401
	classes=80
	num=9
	jitter=.3
	ignore_thresh = .7
	truth_thresh = 1
	scale_x_y = 1.1
	iou_thresh=0.213
	cls_normalizer=1.0
	iou_normalizer=0.07
	iou_loss=ciou
	nms_kind=greedynms
	beta_nms=0.6

pytorch-YOLOv4/cfg/yolov4-custom.cfg

Lines 1140 to 1156 in af00822

	[yolo]
	mask = 6,7,8
	anchors = 12, 16, 19, 36, 40, 28, 36, 75, 76, 55, 72, 146, 142, 110, 192, 243, 459, 401
	classes=80
	num=9
	jitter=.3
	ignore_thresh = .7
	truth_thresh = 1
	random=1
	scale_x_y = 1.05
	iou_thresh=0.213
	cls_normalizer=1.0
	iou_normalizer=0.07
	iou_loss=ciou
	nms_kind=greedynms
	beta_nms=0.6

Hey, I had a similar error, can you tell us how to change it according to the dataset? I have no idea how to change and what to change. Thank you very much!

[zhaoyin214]

You can choose how to build the model.

pytorch-YOLOv4/cfg.py

Lines 17 to 18 in af00822

	Cfg.use_darknet_cfg = True
	Cfg.cfgfile = 'cfg/yolov4.cfg'

if you use cfg file,

Cfg.use_darknet_cfg = True
Cfg.cfgfile = 'cfg file'

and you need to change the class parameter of the yolo layer of the cfg file according to your data set.

pytorch-YOLOv4/cfg/yolov4-custom.cfg

Lines 966 to 980 in af00822

	[yolo]
	mask = 0,1,2
	anchors = 12, 16, 19, 36, 40, 28, 36, 75, 76, 55, 72, 146, 142, 110, 192, 243, 459, 401
	classes=80
	num=9
	jitter=.3
	ignore_thresh = .7
	truth_thresh = 1
	scale_x_y = 1.2
	iou_thresh=0.213
	cls_normalizer=1.0
	iou_normalizer=0.07
	iou_loss=ciou
	nms_kind=greedynms
	beta_nms=0.6

pytorch-YOLOv4/cfg/yolov4-custom.cfg

Lines 1053 to 1068 in af00822

	[yolo]
	mask = 3,4,5
	anchors = 12, 16, 19, 36, 40, 28, 36, 75, 76, 55, 72, 146, 142, 110, 192, 243, 459, 401
	classes=80
	num=9
	jitter=.3
	ignore_thresh = .7
	truth_thresh = 1
	scale_x_y = 1.1
	iou_thresh=0.213
	cls_normalizer=1.0
	iou_normalizer=0.07
	iou_loss=ciou
	nms_kind=greedynms
	beta_nms=0.6

pytorch-YOLOv4/cfg/yolov4-custom.cfg

Lines 1140 to 1156 in af00822

	[yolo]
	mask = 6,7,8
	anchors = 12, 16, 19, 36, 40, 28, 36, 75, 76, 55, 72, 146, 142, 110, 192, 243, 459, 401
	classes=80
	num=9
	jitter=.3
	ignore_thresh = .7
	truth_thresh = 1
	random=1
	scale_x_y = 1.05
	iou_thresh=0.213
	cls_normalizer=1.0
	iou_normalizer=0.07
	iou_loss=ciou
	nms_kind=greedynms
	beta_nms=0.6

not only classes, but also the filters in [convolutional] just before [yolo],

[convolutional]
size=1
stride=1
pad=1
filters=255
activation=linear

[yolo]
mask = 3,4,5
anchors = 12, 16, 19, 36, 40, 28, 36, 75, 76, 55, 72, 146, 142, 110, 192, 243, 459, 401
classes=80
num=9
jitter=.3
ignore_thresh = .7
truth_thresh = 1
scale_x_y = 1.1
iou_thresh=0.213
cls_normalizer=1.0
iou_normalizer=0.07
iou_loss=ciou
nms_kind=greedynms
beta_nms=0.6

it should be 255 / 85 * (5 + classes)

[zsh4614]

You can choose how to build the model.

pytorch-YOLOv4/cfg.py

Lines 17 to 18 in af00822

	Cfg.use_darknet_cfg = True
	Cfg.cfgfile = 'cfg/yolov4.cfg'

if you use cfg file,

Cfg.use_darknet_cfg = True
Cfg.cfgfile = 'cfg file'

and you need to change the class parameter of the yolo layer of the cfg file according to your data set.

pytorch-YOLOv4/cfg/yolov4-custom.cfg

Lines 966 to 980 in af00822

	[yolo]
	mask = 0,1,2
	anchors = 12, 16, 19, 36, 40, 28, 36, 75, 76, 55, 72, 146, 142, 110, 192, 243, 459, 401
	classes=80
	num=9
	jitter=.3
	ignore_thresh = .7
	truth_thresh = 1
	scale_x_y = 1.2
	iou_thresh=0.213
	cls_normalizer=1.0
	iou_normalizer=0.07
	iou_loss=ciou
	nms_kind=greedynms
	beta_nms=0.6

pytorch-YOLOv4/cfg/yolov4-custom.cfg

Lines 1053 to 1068 in af00822

	[yolo]
	mask = 3,4,5
	anchors = 12, 16, 19, 36, 40, 28, 36, 75, 76, 55, 72, 146, 142, 110, 192, 243, 459, 401
	classes=80
	num=9
	jitter=.3
	ignore_thresh = .7
	truth_thresh = 1
	scale_x_y = 1.1
	iou_thresh=0.213
	cls_normalizer=1.0
	iou_normalizer=0.07
	iou_loss=ciou
	nms_kind=greedynms
	beta_nms=0.6

pytorch-YOLOv4/cfg/yolov4-custom.cfg

Lines 1140 to 1156 in af00822

	[yolo]
	mask = 6,7,8
	anchors = 12, 16, 19, 36, 40, 28, 36, 75, 76, 55, 72, 146, 142, 110, 192, 243, 459, 401
	classes=80
	num=9
	jitter=.3
	ignore_thresh = .7
	truth_thresh = 1
	random=1
	scale_x_y = 1.05
	iou_thresh=0.213
	cls_normalizer=1.0
	iou_normalizer=0.07
	iou_loss=ciou
	nms_kind=greedynms
	beta_nms=0.6

Hey, I had a similar error, can you tell us how to change it according to the dataset? I have no idea how to change and what to change. Thank you very much!

the layer output channels before [yolo] should be set 3*(5+classes) instead of 255!

[Andrej-sens]

Hello, When I started training on my own dataset, it shows error like.

log file path:log\log_2020-07-02_23-18-42.txt
2020-07-02 23:18:42,322 train.py[line:468] INFO: Using device cuda
convalution havn't activate linear
convalution havn't activate linear
convalution havn't activate linear
2020-07-02 23:18:44,634 train.py[line:318] INFO: Starting training:
        Epochs:          300
        Batch size:      16
        Subdivisions:    16
        Learning rate:   0.001
        Training size:   144
        Validation size: 16
        Checkpoints:     True
        Device:          cuda
        Images size:     608
        Optimizer:       adam
        Dataset classes: 24
        Train label path:data/train.txt
        Pretrained:
    
Epoch 1/300:   0%|       | 0/144 [00:12<?, ?img/s]
Traceback (most recent call last):
  File "C:/Users/Windows/Downloads/pytorch-YOLOv4-master/train.py", line 483, in <module>
    device=device, )
  File "C:/Users/Windows/Downloads/pytorch-YOLOv4-master/train.py", line 356, in train
    loss, loss_xy, loss_wh, loss_obj, loss_cls, loss_l2 = criterion(bboxes_pred, bboxes)
  File "C:\Softwares\Anaconda3\envs\torch\lib\site-packages\torch\nn\modules\module.py", line 550, in __call__
    result = self.forward(*input, **kwargs)
  File "C:/Users/Windows/Downloads/pytorch-YOLOv4-master/train.py", line 233, in forward
    output = output.view(batchsize, self.n_anchors, n_ch, fsize, fsize)
RuntimeError: shape '[1, 3, 29, 76, 76]' is invalid for input of size 1472880

Does anyone face the same problem?

I had the same issue.
After the debugging I realised that I did stupid mistake and the input image to YOLOv4 had wrong dimensions. It should be [B, CH, W, H]. If you still have a problem print the size of the output for further debugging.

[wangyuehy]

I have met same issue while deal with scale 19201080, the problem is the during upsampling, the x size is [batch, 128, 68, 80] while target_size is [batch, 256,135, 160], the 682 != 135, so for PyTorch, a possible way is to get x[:,:,:135,:].
but another problem is during PyTorch inference, in get_region_boxes, the boxes size is not same with Darknet version.