search

MehmetAygun / demistfy_correspondence

Code for the ECCV22 paper Demystifying Unsupervised Semantic Correspondence Estimation
https://mehmetaygun.github.io/demistfy.html
13 stars 0 forks source link

Pre-Trained Model #2

Closed dragonlee258079 closed 10 months ago

dragonlee258079 commented 10 months ago

Hi, I couldn't find your pre-trained model, only log files are provided, but loading log files can only load the last convolutional layer of the model.

MehmetAygun commented 10 months ago

Hi,

In all of our experiments the backbones are frozen, so we just used already pre-trained versions. We only trained one single convolutional layer, and the provided files have the weights for them.

dragonlee258079 commented 10 months ago

Hi, the last layer of the resnet50 network outputs features of 2048 channels, but the corresponding files only provided 1024 to 256 convolutional layer, and there is no model construction for this convolutional layer in the code.

MehmetAygun commented 10 months ago

Sorry my bad, the projection layer (single layer conv) is coded as the parameter w in the evaluation code, https://github.com/MehmetAygun/demistfy_correspondence/blob/main/evaluation/cub/evaluate.py#L29

Here you can see it is loaded: https://github.com/MehmetAygun/demistfy_correspondence/blob/main/evaluation/cub/evaluate.py#L144C42-L144C42

The model that you are trying might be the unsupervised Resnet model which has a different output feature dimension.

dragonlee258079 commented 10 months ago

Thank you for resolving my question.