search

nianticlabs / ace

[CVPR 2023 - Highlight] Accelerated Coordinate Encoding (ACE): Learning to Relocalize in Minutes using RGB and Poses
https://nianticlabs.github.io/ace
Other
360 stars 34 forks source link

Where does ACE freezes the backbone when training the head #28

Closed willdzeng closed 1 year ago

willdzeng commented 1 year ago

Hi,

I can't seem to find where does ACE freeze the backbone during training, since during test it use the same pre-trained encoder, during training it's probably better to freeze the backbone such that the backbone doesn't change, right?

or is this on purpose to not freeze the backbone?

willdzeng commented 1 year ago

NVM, I think I know why it doesn't need to freeze the backbone in the training function, it's because the training buffer is created with eval() mode and no_grad() before running the training, so when doing the actual training of MLP head, it doesn't have gradient in the encoder's parameters. Thanks, closing the issue