search

mihaidusmanu / d2-net

D2-Net: A Trainable CNN for Joint Description and Detection of Local Features
Other
761 stars 163 forks source link

Does the model_test need to train once again? #86

Closed xiezhiqing0112 closed 2 years ago

xiezhiqing0112 commented 3 years ago

Hello.the paper seems to use diffirent model in train and test.So what's the use of train model?I can't be loaded during infertime?

07Agarg commented 2 years ago

Same doubt! @xiezhiqing0112, were you able to resolve this?

mihaidusmanu commented 2 years ago

The train model can be loaded during inference as both test and train have exactly the same layers / weights. The only changes between the two of them is that the test model uses different strides / dilation rates for the last layers which does not affect the trainable weights. This allows us to have a more light-weight training in terms of both memory usage and compute.

07Agarg commented 2 years ago

Thanks for your quick response, @mihaidusmanu !