a little problem when using some pretrained models

[ZhiyanWXZ]

🐛 Bug

A little problem.

I used some pretrained models to do object detection.

However, when i used models whose name include others models' name, such as 'fasterrcnn_mobilenet_v3_large_320_fpn', (its name includes the name of model 'mobilenet_v3_large'), it will download the weight file of the short name models.

For example, when i used model 'fasterrcnn_mobilenet_v3_large_320_fpn', whenever the pretrained attribute is True or not, the weight file of model 'mobilenet_v3_large' will be downloaded.

This problem also happen in the models such as 'maskrcnn_resnet50_fpn' and many other models.

To Reproduce

Steps to reproduce the behavior:
it's easy to reproduce. For example:

from torchvision.models.detection import *
model = fasterrcnn_mobilenet_v3_large_320_fpn()

excute the code above, the weight file of model "mobilenet_v3_large" will be downloaded.
you can change the model name to other this kind of model names.

cc @fmassa @vfdev-5 @pmeier

[vfdev-5]

@ZhiyanWXZ you can pass pretrained_backbone=False and pretrained=False to skip any pretrained weights downloading

[fmassa]

@datumbox maybe we should only download the backbone weights if pretrained=False and pretrained_backbone=True?

[datumbox]

@ZhiyanWXZ the vast majority of the models have pretrained=False and pretrained_backbone=True as default parameters which means this is an expected behaviour. As @vfdev-5 said, if you pass pretrained_backbone=False it won't download the backbone data.

As for when pretrained=True as far as I can tell the pretrained_backbone is set to False automatically for all models. Which version of TorchVision are you using?

fasterrcnn_mobilenet_v3_large_fpn/fasterrcnn_mobilenet_v3_large_320_fpn:

vision/torchvision/models/detection/faster_rcnn.py

Lines 382 to 383 in 183a722

	if pretrained:
	pretrained_backbone = False

fasterrcnn_resnet50_fpn:

vision/torchvision/models/detection/faster_rcnn.py

Lines 364 to 366 in 183a722

	if pretrained:
	# no need to download the backbone if pretrained is set
	pretrained_backbone = False

keypointrcnn_resnet50_fpn:

vision/torchvision/models/detection/keypoint_rcnn.py

Lines 334 to 336 in 183a722

	if pretrained:
	# no need to download the backbone if pretrained is set
	pretrained_backbone = False

maskrcnn_resnet50_fpn:

vision/torchvision/models/detection/mask_rcnn.py

Lines 327 to 329 in 183a722

	if pretrained:
	# no need to download the backbone if pretrained is set
	pretrained_backbone = False

retinanet_resnet50_fpn:

vision/torchvision/models/detection/retinanet.py

Lines 616 to 618 in 183a722

	if pretrained:
	# no need to download the backbone if pretrained is set
	pretrained_backbone = False

ssd300_vgg16:

vision/torchvision/models/detection/ssd.py

Lines 574 to 576 in 183a722

	if pretrained:
	# no need to download the backbone if pretrained is set
	pretrained_backbone = False

ssdlite320_mobilenet_v3_large:

vision/torchvision/models/detection/ssdlite.py

Lines 193 to 194 in 183a722

	if pretrained:
	pretrained_backbone = False

fcn_resnet50/fcn_resnet101/deeplabv3_resnet50/deeplabv3_resnet101/deeplabv3_mobilenet_v3_large:

vision/torchvision/models/segmentation/segmentation.py

Lines 70 to 72 in 183a722

	if pretrained:
	aux_loss = True
	kwargs["pretrained_backbone"] = False

The only problematic case I found was for lraspp_mobilenet_v3_large which I'll patch in a bit.

[ZhiyanWXZ]

thanks, I get it!