✨ Add Reverse Distillation #343
Conversation
There was a problem hiding this comment.
Thanks, good to see this model being added to the repo. The torch model seems to contain some classes that were directly copied from torchvision with some minor changes. Do you think it would be possible to import the models from torchvision instead, and replace the specific parts? It would keep our code base a bit cleaner.
| inplanes (int): Number of input channels. | ||
| planes (int): Number of output channels. |
There was a problem hiding this comment.
do you know why these are called in/out planes. Is it how they are named in the paper?
There was a problem hiding this comment.
I don't know why. But the code in RD is taken from resent in torchvision so based on conv parameters, I've added this docstring https://github.com/pytorch/vision/blob/main/torchvision/models/resnet.py#L64
cosine_similarity is not supported op in onnx
| @@ -68,6 +68,8 @@ def forward(self, images: Tensor) -> Union[Tensor, Tuple[List[Tensor], List[Tens | |||
| Union[Tensor, Tuple[List[Tensor],List[Tensor]]]: Encoder and decoder features in training mode, | |||
| else anomaly maps. | |||
| """ | |||
| self.encoder.eval() | |||
There was a problem hiding this comment.
Are the benchmarking results still the same after this change?
There was a problem hiding this comment.
I am running it right now. But that was a good catch. I saw that the encoder was in training mode in the train step.
There was a problem hiding this comment.
Interestingly apart from pixel level score, image level scores are lower
There was a problem hiding this comment.
The average results also seem to be a bit lower than those reported in the original paper (98.5% image AUROC, 97.8% pixel AUROC). For now I would suggest to merge this PR, but it would be good to investigate if this difference can be explained.
There was a problem hiding this comment.
Does the official implementation give the exact same results reported in the paper?
Description
Add Reverse Distillation model based on official code - https://github.com/hq-deng/RD4AD
Fixes Add Reverse Distillation Model #263
TODO