[Bug]: patchcore pytorch error

[jth3galv]

Using pathcore gives the following torch error.:

RuntimeError: expand(torch.FloatTensor{[1, 1, 1, 33, 33]}, size=[1, -1, -1, -1]): the number of sizes provided (4) must be greater or equal to the number of dimensions in the tensor (5)

See the log sections for the full error.

Dataset

Folder

Model

PatchCore

Steps to reproduce the behavior

I am trying to use patchcore as by tutorial. I am using a fresh install of anomalib 1.0.0.dev.
Running the following produce an error.

# Import the datamodule
from anomalib.data import Folder
from anomalib.data.utils import TestSplitMode
torchmetrics.functional.auc = AUROC # needed otherwise i get an import error

# Import the required modules
from anomalib.data import MVTec
from anomalib.models import Patchcore
from anomalib.engine import Engine


# Create the datamodule
datamodule = Folder(
    root='myfolder"',
    normal_dir="OK",
    abnormal_dir="KO",
    image_size=(256, 256),
    test_split_mode=TestSplitMode.SYNTHETIC,
    task="segmentation",
)

# Setup the datamodule

datamodule.setup()

model = Patchcore()

engine = Engine()

OS information

OS information:

• OS: Ubuntu 22.04
• Python version: 3.10
• Anomalib version: '1.0.0dev'
• PyTorch version: '2.2.1+cu121'
• CPU

Expected behavior

to run without errors

Screenshots

No response

Pip/GitHub

GitHub

What version/branch did you use?

No response

Configuration YAML

?

Logs

RuntimeError                              Traceback (most recent call last)
Cell In[34], line 7
      3 from anomalib.models import Patchcore
      4 from anomalib.engine import Engine
----> 7 model = Patchcore()
      9 engine = Engine()

File ~/code/anomalib/src/anomalib/models/image/patchcore/lightning_model.py:53, in Patchcore.__init__(self, backbone, layers, pre_trained, coreset_sampling_ratio, num_neighbors)
     43 def __init__(
     44     self,
     45     backbone: str = "wide_resnet50_2",
   (...)
     49     num_neighbors: int = 9,
     50 ) -> None:
     51     super().__init__()
---> 53     self.model: PatchcoreModel = PatchcoreModel(
     54         backbone=backbone,
     55         pre_trained=pre_trained,
     56         layers=layers,
     57         num_neighbors=num_neighbors,
     58     )
     59     self.coreset_sampling_ratio = coreset_sampling_ratio
     60     self.embeddings: list[torch.Tensor] = []

File ~/code/anomalib/src/anomalib/models/image/patchcore/torch_model.py:55, in PatchcoreModel.__init__(self, layers, backbone, pre_trained, num_neighbors)
     49 self.feature_extractor = TimmFeatureExtractor(
     50     backbone=self.backbone,
     51     pre_trained=pre_trained,
     52     layers=self.layers,
     53 )
     54 self.feature_pooler = torch.nn.AvgPool2d(3, 1, 1)
---> 55 self.anomaly_map_generator = AnomalyMapGenerator()
     57 self.register_buffer("memory_bank", torch.Tensor())
     58 self.memory_bank: torch.Tensor

File ~/code/anomalib/src/anomalib/models/image/patchcore/anomaly_map.py:30, in AnomalyMapGenerator.__init__(self, sigma)
     28 super().__init__()
     29 kernel_size = 2 * int(4.0 * sigma + 0.5) + 1
---> 30 self.blur = GaussianBlur2d(kernel_size=(kernel_size, kernel_size), sigma=(sigma, sigma), channels=1)

File ~/code/anomalib/src/anomalib/models/components/filters/blur.py:68, in GaussianBlur2d.__init__(self, sigma, channels, kernel_size, normalize, border_type, padding)
     66     self.kernel = normalize_kernel2d(self.kernel)
     67 self.kernel.unsqueeze_(0).unsqueeze_(0)
---> 68 self.kernel = self.kernel.expand(self.channels, -1, -1, -1)
     69 self.border_type = border_type
     70 self.padding = padding

RuntimeError: expand(torch.FloatTensor{[1, 1, 1, 33, 33]}, size=[1, -1, -1, -1]): the number of sizes provided (4) must be greater or equal to the number of dimensions in the tensor (5)

Code of Conduct

• I agree to follow this project's Code of Conduct

[jth3galv]

I downgraded torch and torchmetrics and it works without errors. I guess some changes will be needed to make it work with newer versions.

[lukasvrabel]

which versions are you using? I downgraded to torch==2.1.0 and it still gives the same error. I cannot downgrade more, because then it conflicts with torchvision being too low...

[jth3galv]

I am using:

• torch version: '2.1.2+cu121'
• torchmetrics version : '0.10.3'
• torchvision version: '0.16.2+cu121'

In general I think the quickest approach (this is what I finally did ) to make it work is to have a virtual env and install all the requirements as listed in the requirements files.

[voidmain443]

I downgraded this package and then it works

I downgraded and. then it works I can't resolve this issue just downgrade this package 0.7.0

[samet-akcay]

This is peculiar. I'm having trouble reproducing this issue on my end

[samet-akcay]

@voidmain443, on a fresh environment, can you try pip install anomalib[full]? This works on our side.