Yolov5: ResNet-50 backbone config

[twangnh]

Hi! thanks for sharing the wonderful yolov5 in pytorch, I'm wondering if you could share backbone configuration for resnet50, currently config only has CSP backbone:

backbone:
# [from, number, module, args]
[[-1, 1, Focus, [64, 3]], # 0-P1/2
[-1, 1, Conv, [128, 3, 2]], # 1-P2/4
[-1, 3, BottleneckCSP, [128]],
[-1, 1, Conv, [256, 3, 2]], # 3-P3/8
[-1, 9, BottleneckCSP, [256]],
[-1, 1, Conv, [512, 3, 2]], # 5-P4/16
[-1, 9, BottleneckCSP, [512]],
[-1, 1, Conv, [1024, 3, 2]], # 7-P5/32
[-1, 1, SPP, [1024, [5, 9, 13]]],
[-1, 3, BottleneckCSP, [1024, False]], # 9
]

Thank you.

[glenn-jocher]

@twangnh I don't have time or motivation to but we are open to PRs so feel free to submit one for this!

[github-actions[bot]]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.

[Shakesbeer333]

@twangnh did you find a backbone config for resnet50?

[glenn-jocher]

@Shakesbeer333 you can build your own resnet50.yaml by using the existing ones as templates. The backbone section you'd modify is here:
https://github.com/ultralytics/yolov5/blob/187f7c2ed16b9eb9754b28e7b0aa397f908155aa/models/yolov5s.yaml#L12-L26

[Shakesbeer333]

@glenn-jocher thank you. Any other section I need to change or to be aware of?

[glenn-jocher]

@Shakesbeer333 no just the model.yaml. You want to make a resnet50.yaml basically, with a resnet50 backbone and whatever head you want essentially.

You can use standard pytorch modules to create any architecture, i.e.:

# [from, number, module, args] 
[-1, 1, nn.Conv2d, [128, 3, 2]]
[-1, 1, nn.BatchNorm2d, [128]]
...

[Shakesbeer333]

@glenn-jocher thanks once again. Can you briefly tell me what args[0], args[1], ... is or where I can find a description?

[glenn-jocher]

@Shakesbeer333 these are arguments you would pass to the module. i.e. nn.BatchNorm2d(128)

[Shakesbeer333]

@glenn-jocher my question was a little bit vague. I retry:

[-1, 1, nn.Conv2d, [128, 3, 2] becomes torch.nn.modules.conv.Conv2d(c1=32, c2=64, k=3, s=2)

I don't understand how args[0] = 128 translates into c1=32, c2=64

[glenn-jocher]

@Shakesbeer333 YOLOv5 models are compound scaled in depth and width as shown in yaml comments:
https://github.com/ultralytics/yolov5/blob/187f7c2ed16b9eb9754b28e7b0aa397f908155aa/models/yolov5s.yaml#L1-L5

[Shakesbeer333]

@glenn-jocher aweseome! Thank you so much for taking time to answer my questions! I really appreciate it

[xoryouyou]

@Shakesbeer333 any success implementing a ResNet backbone ?

[Shakesbeer333]

@xoryouyou I'm working on it atm but but due to my inexperience it takes some time. I'll let you know if I got something useful. In the meantime let me know if you got anything.

[Shakesbeer333]

@xoryouyou I used see @FrancescoSaverioZuppichini 's ResNet implementation . ResNet18 runs without error (I didn't check any performance yet) however the other ResNet versions through an error.

[xoryouyou]

@Shakesbeer333 No I haven't seen it. Thanks for the link! I'll give it a spin once my machine is finished training my current yolov5x net.

Edit: looks like pytorch vision already features all we need https://github.com/pytorch/vision/blob/master/torchvision/models/resnet.py

[Shakesbeer333]

@xoryouyou yes we got all we need! However, I was facing tensor shape problems when connecting it to the head.