Pretrained efficientnet_b0_rwrightman-3dd342df state_dict fails sha256 check

[ptrblck]

🐛 Describe the bug

Originally reported in the forum.

I am able to reproduce the issue using torchvision==0.16.0.dev20230709+cu121:

model = models.efficientnet_b0(pretrained=True)

/home/pbialecki/miniforge3/envs/nightly_pip_cu121/lib/python3.10/site-packages/torchvision/models/_utils.py:208: UserWarning: The parameter 'pretrained' is deprecated since 0.13 and may be removed in the future, please use 'weights' instead.
  warnings.warn(
/home/pbialecki/miniforge3/envs/nightly_pip_cu121/lib/python3.10/site-packages/torchvision/models/_utils.py:223: UserWarning: Arguments other than a weight enum or `None` for 'weights' are deprecated since 0.13 and may be removed in the future. The current behavior is equivalent to passing `weights=EfficientNet_B0_Weights.IMAGENET1K_V1`. You can also use `weights=EfficientNet_B0_Weights.DEFAULT` to get the most up-to-date weights.
  warnings.warn(msg)
Downloading: "https://download.pytorch.org/models/efficientnet_b0_rwightman-3dd342df.pth" to /home/pbialecki/.cache/torch/hub/checkpoints/efficientnet_b0_rwightman-3dd342df.pth
100%|██████████| 20.5M/20.5M [00:00<00:00, 52.8MB/s]
Traceback (most recent call last):

  Cell In[75], line 1
    model = models.efficientnet_b0(pretrained=True)

  File ~/miniforge3/envs/nightly_pip_cu121/lib/python3.10/site-packages/torchvision/models/_utils.py:142 in wrapper
    return fn(*args, **kwargs)

  File ~/miniforge3/envs/nightly_pip_cu121/lib/python3.10/site-packages/torchvision/models/_utils.py:228 in inner_wrapper
    return builder(*args, **kwargs)

  File ~/miniforge3/envs/nightly_pip_cu121/lib/python3.10/site-packages/torchvision/models/efficientnet.py:770 in efficientnet_b0
    return _efficientnet(

  File ~/miniforge3/envs/nightly_pip_cu121/lib/python3.10/site-packages/torchvision/models/efficientnet.py:360 in _efficientnet
    model.load_state_dict(weights.get_state_dict(progress=progress, check_hash=True))

  File ~/miniforge3/envs/nightly_pip_cu121/lib/python3.10/site-packages/torchvision/models/_api.py:90 in get_state_dict
    return load_state_dict_from_url(self.url, *args, **kwargs)

  File ~/miniforge3/envs/nightly_pip_cu121/lib/python3.10/site-packages/torch/hub.py:757 in load_state_dict_from_url
    download_url_to_file(url, cached_file, hash_prefix, progress=progress)

  File ~/miniforge3/envs/nightly_pip_cu121/lib/python3.10/site-packages/torch/hub.py:653 in download_url_to_file
    raise RuntimeError('invalid hash value (expected "{}", got "{}")'

RuntimeError: invalid hash value (expected "3dd342df", got "7f5810bc96def8f7552d5b7e68d53c4786f81167d28291b21c0d90e1fca14934")

Downloading the state_dict manually also confirms the sha256 checksum:

wget https://download.pytorch.org/models/efficientnet_b0_rwightman-3dd342df.pth
sha256sum efficientnet_b0_rwightman-3dd342df.pth 
7f5810bc96def8f7552d5b7e68d53c4786f81167d28291b21c0d90e1fca14934  efficientnet_b0_rwightman-3dd342df.pth

which does not match the one stored in the file name.

Versions

torchvision==0.16.0.dev20230709+cu121

[NicolasHug]

Thanks for the report @ptrblck . I'll take a look on the S3 bucket and see if the file has changed. It would be strange though, I suspect / hope it was just a typo we introduced originally

[atalman]

hi @NicolasHug any progress on this ?

[malfet]

Hmm, it was last modified in Aug 2021:

% aws s3 ls s3://pytorch/models/efficientnet_b0_rwightman-3dd342df.pth
2021-08-25 06:11:27   21444401 efficientnet_b0_rwightman-3dd342df.pth
% shasum -a 256 ~/.cache/torch/hub/checkpoints/efficientnet_b0_rwightman-3dd342df.pth
7f5810bc96def8f7552d5b7e68d53c4786f81167d28291b21c0d90e1fca14934  /Users/nshulga/.cache/torch/hub/checkpoints/efficientnet_b0_rwightman-3dd342df.pth

[Edit] Torchvision checked out at 2925df7 works, but latest trunk 29418e3 fails

[Edit2] #7219 added sha-sum checks, but I guess we've never validated that it is correct.

[ptrblck]

Torchvision checked out at 2925df7 works, but latest trunk 29418e3 fails

Is it because #7219 landed between these commits?

[malfet]

@ptrblck bingo :) So it was like that for a while, but it was not checked before... Running the script to validate published checksums right now...

[NicolasHug]

Thanks a lot all for the reports and investigations.

I re-uploaded the weights with the proper hashes on S3 and also fixed the URL in #7898.

[ar0ck]

I can still reproduce this issue with torchvision==0.17.0.dev20231009.

@NicolasHug, running your script from #7898 I get:

Following models need to be fixed:
efficientnet_b0 EfficientNet_B0_Weights.IMAGENET1K_V1 invalid hash value (expected "3dd342df", got "7f5810bc96def8f7552d5b7e68d53c4786f81167d28291b21c0d90e1fca14934")
efficientnet_b2 EfficientNet_B2_Weights.IMAGENET1K_V1 invalid hash value (expected "bcdf34b7", got "c35c147384e385a5bab5a8eabdabbe5a3df0487ee4a554108626ae474a5bf755")
efficientnet_b3 EfficientNet_B3_Weights.IMAGENET1K_V1 invalid hash value (expected "cf984f9c", got "b3899882250c22946d0229d266049fcd133c169233530b36b9ffa7983988362f")
efficientnet_b4 EfficientNet_B4_Weights.IMAGENET1K_V1 invalid hash value (expected "7eb33cd5", got "23ab8bcd5bdbef61a7a43b91adcad81f622fd7f36fb4935a569828d77888c44e")
efficientnet_b5 EfficientNet_B5_Weights.IMAGENET1K_V1 invalid hash value (expected "b6417697", got "1a07897c0d357db7981640f6be44a63420f11deb932344a69768b62ebe272946")
efficientnet_b6 EfficientNet_B6_Weights.IMAGENET1K_V1 invalid hash value (expected "c76e70fd", got "24a108a596a00ad522bdfb9a2d98a9bd31819dfd7e37dd96bad6bb5bc00d6015")
efficientnet_b7 EfficientNet_B7_Weights.IMAGENET1K_V1 invalid hash value (expected "dcc49843", got "c5b4e57e0da1a10bb4f8a5b410dd3b7b6e7461f9b757209be29409707c3d87dd")

[NicolasHug]

Thanks for the report @ar0ck . I'm confused as to how I missed those in the past, but I can reproduce...

We'll fix those in the next [bugfix] release

[marcellobullo]

Thanks for the effort. Any workaround to fix it temporarily?

[NicolasHug]

EDIT: This is fixed in torchvision 0.17, just install "torchvision>0.16" and this should be fixed. If you're stuck with 0.16, do as described below

---

Original comment:

Overriding torchvision.models._api.WeightsEnum.get_state_dict() should do the trick, LMK if it doesn't work

from torchvision.models import efficientnet_b0, EfficientNet_B0_Weights
from torchvision.models._api import WeightsEnum
from torch.hub import load_state_dict_from_url

def get_state_dict(self, *args, **kwargs):
    kwargs.pop("check_hash")
    return load_state_dict_from_url(self.url, *args, **kwargs)
WeightsEnum.get_state_dict = get_state_dict

efficientnet_b0(weights=EfficientNet_B0_Weights.IMAGENET1K_V1)
efficientnet_b0(weights="DEFAULT")

[fafqqeew]

Overriding torchvision.models._api.WeightsEnum.get_state_dict() did it , thank you .

[dimakpa]

Guys, an error started to occur when loading the model. A week ago everything was loading well

model = efficientnet_b6(pretrained=True)

RuntimeError: invalid hash value (expected "c76e70fd", got "24a108a596a00ad522bdfb9a2d98a9bd31819dfd7e37dd96bad6bb5bc00d6015")

[NicolasHug]

@dimakpa please see #7744 (comment) until the fix is available in the next bugfix release (planned for mid Nov)

[William171409]

Thank you very much! Looking forward to the final solution.

[korablique]

return load_state_dict_from_url(self.url, *args, **kwargs)

What is self.url?

[NicolasHug]

It's an attribute of the WeightEnum class which will become relevant after doing

WeightsEnum.get_state_dict = get_state_dict

[mustious]

For anyone who might still be stuck with the same error. The major issue is from torchvision==0.16.0. An easy workaround is to install higher versions or nightly. Tested and worked on torchvision v0.16.1, v0.16.2, v0.17.0.

pip install "torchvision>0.16.0" 

[winnie128]

I had similar problem when running code in Jupyter Notebook.

I could download efficientnet_b3 model by PyCharm without any issue
model_ft = models.efficientnet_b3(weights=models.EfficientNet_B3_Weights.IMAGENET1K_V1)
And the torchvision version:0.15.2.

I installed torchvision version 0.15.2 for Jupyter Notebook and ran the same model loading. I got error message:

RuntimeError: invalid hash value (expected "cf984f9c", got "b3899882250c22946d0229d266049fcd133c169233530b36b9ffa7983988362f"

[NicolasHug]

@winnie128 please install a more up-to-date version of torchvision

[winnie128]

Problem was solved after torchvision version 0.18.0 installed. Thanks!