SwinV2在MSMT17测试集的问题

[lianshengzhou]

您好，想问一下为什么用swinV2在MSMT17测试的时候出现BUG：
RuntimeError: The size of tensor a (4) must match the size of tensor b (32) at non-singleton dimension 1
发现是在这一步出现问题：Traceback (most recent call last):
File "/content/Person_reID_baseline_pytorch/test.py", line 316, in
gallery_feature = extract_feature(model,dataloaders['gallery'])
File "/content/Person_reID_baseline_pytorch/test.py", line 207, in extract_feature
outputs = model(input_img)
为什么在Market-1501数据集上不会出现这种BUG呢？

[chesianatalia]

Hi, I also have the same issue with a different dataset.
here is the screenshot, I tried months ago and found no problem like this, and suddenly this problem appears. Do you know how to fix it? Thank You

[layumi]

Hello. It seems like the input size.
Could you resize the image size as the same one during your training?

[lianshengzhou]

但是我在market数据集用Resnet-50、Densennet以及Swin Transformer V2均可以正常运行，在MSMT17数据集上Resnet-50和Densenet均可以正常运行，是两个数据集之间的差异吗？

[layumi]

@lianshengzhou
诶，msmt17应该也是可以的跑的。我怀疑是 input size不对。 您方便的话 可以 print 一下 input size对比一下。

[chesianatalia]

Hello, I've checked the input size, and found that the input size was the same during my training and I'm still getting the same error, is there any other solution for this issue? Thank You
the screenshot is attached below:

this is the input size in my training:

[layumi]

Thank you @chesianatalia

I will try it and return to you soon.

[layumi]

Thank you @chesianatalia @lianshengzhou

I found why. It is due to torch.jit.trace
You can comment out these two lines. The test should be Okay now.
https://github.com/layumi/Person_reID_baseline_pytorch/blob/master/test.py#L306-L307

But I am also curious why torch.jit can work with Market but MSMT. It is interesting.

[layumi]

Just for your reference.
You can check my test MSMT code at https://github.com/layumi/Person_reID_baseline_pytorch/blob/master/test_MSMT.py

[lianshengzhou]

@layumi @chesianatalia 试了一下，能够成功运行，结果也很好，感谢及时回复和帮助。

[chesianatalia]

Thank you for your explanation, i'll working on it

[chesianatalia]

Hi @layumi i've already comment out two lines as you mentioned earlier, turns out there are new error comes up. I've never see an error like this before, and when i undo what i did, the error is still the same, and now it affect on my other model (not only SwinV2). Kindly need your advice

the screenshot attached below:

here are the full error:

File "/content/drive/MyDrive/ChesiaGraceNatalia-TA/Person_reID_baseline_pytorch/test.py", line 279, in
model = load_network(model_structure)
File "/content/drive/MyDrive/ChesiaGraceNatalia-TA/Person_reID_baseline_pytorch/test.py", line 157, in load_network
network.load_state_dict(torch.load(save_path))
File "/usr/local/lib/python3.10/dist-packages/torch/nn/modules/module.py", line 2041, in load_state_dict
raise RuntimeError('Error(s) in loading state_dict for {}:\n\t{}'.format(
RuntimeError: Error(s) in loading state_dict for ft_net_swinv2:
Missing key(s) in state_dict: "model.layers.3.downsample.reduction.weight", "model.layers.3.downsample.norm.weight", "model.layers.3.downsample.norm.bias".
Unexpected key(s) in state_dict: "model.layers.0.downsample.reduction.weight", "model.layers.0.downsample.norm.weight", "model.layers.0.downsample.norm.bias", "model.layers.0.blocks.1.attn_mask", "model.layers.1.blocks.1.attn_mask", "model.layers.2.blocks.1.attn_mask", "model.layers.2.blocks.3.attn_mask", "model.layers.2.blocks.5.attn_mask", "model.layers.2.blocks.7.attn_mask", "model.layers.2.blocks.9.attn_mask", "model.layers.2.blocks.11.attn_mask", "model.layers.2.blocks.13.attn_mask", "model.layers.2.blocks.15.attn_mask", "model.layers.2.blocks.17.attn_mask".
size mismatch for model.layers.1.downsample.reduction.weight: copying a param with shape torch.Size([512, 1024]) from checkpoint, the shape in current model is torch.Size([256, 512]).
size mismatch for model.layers.1.downsample.norm.weight: copying a param with shape torch.Size([512]) from checkpoint, the shape in current model is torch.Size([256]).
size mismatch for model.layers.1.downsample.norm.bias: copying a param with shape torch.Size([512]) from checkpoint, the shape in current model is torch.Size([256]).
size mismatch for model.layers.2.downsample.reduction.weight: copying a param with shape torch.Size([1024, 2048]) from checkpoint, the shape in current model is torch.Size([512, 1024]).
size mismatch for model.layers.2.downsample.norm.weight: copying a param with shape torch.Size([1024]) from checkpoint, the shape in current model is torch.Size([512]).
size mismatch for model.layers.2.downsample.norm.bias: copying a param with shape torch.Size([1024]) from checkpoint, the shape in current model is torch.Size([512]).

[layumi]

@chesianatalia

1. Could I know the version of your timm?

2. Could you re-train a new SwinV2 model and test the result?

[chesianatalia]

@layumi Hi, I think you are right, I checked another issue, it was due to timm version, I downgraded it, and suddenly it worked, Thank you very much for your suggestion.
Have a great day!

[layumi]

Thank you @chesianatalia

Yes.

BTW, old timm like 0.6. 0.7 may be not a stable SwinV2.

For the latest timm, you also need python >=3.8, otherwise you may meet some ``typing'' error.