PatchCore only support 224x224 input, and 28x28 feature maps.

[samet-akcay]

No description provided.

[soad89]

Hello, can you please guide me a bit what needs to be modified to allow higher input resolutions?
Thanks

[Lucas-Steinmann]

Hi,
I solved this problem by inferring the largest feature map size from the given feature layer names and the input/tiling size.
I assumed

• the layer names to be of the same format as ResNet ("layer1", "layer2", ...) and
• the sub-sampling ratios are the same as in ResNet, i.e. 2**(1+layer_num)

I know this is kind of hard-coded, but if that is general enough for the maintainers of anomalib, then I'm happy to create a pull request.
Support for other backbones of torchvision could be added by someone motivated enough, to collect all output shapes of various backbones.

More general solutions would either require:

• forwarding a blank image and collecting the feature maps, to get their sizes, or
• inspecting the network to infer the output shapes/sub-sampling ratio. But this is difficult, since the network implementation of torchvision does not keep track of the output shapes explicitly. Some other libraries (e.g. detectron2 with their ShapeSpec) do this.

[samet-akcay]

Hi @Lucas-Steinmann, thanks for looking at the issue and proposing a solution. Would be good if you could create the PR. It would also be great if you could document/comment the code based on what you described above so that future enhancements could be built on top of this.

[samet-akcay]

Since PR #148 addressed this, I'm closing the issue.