EdwardSmith1884
commented
1 year ago Yes I believe you are right. I would say best practice would be to set the mask_features to None initially, and then check if it is None before computing a second time. I don't have the cycles right now to test this, but if you want to speed up things on your end I would try this.
Hi, thanks for your help so far! I have noticed that the mask embedding, which is computed at the first deformation loop https://github.com/facebookresearch/Active-3D-Vision-and-Touch/blob/c10da8f94491d6c9bbe55121dccb2e7faa543cfd/pterotactyl/reconstruction/vision/model.py#L230 is computed again at the third deformation loop: https://github.com/facebookresearch/Active-3D-Vision-and-Touch/blob/c10da8f94491d6c9bbe55121dccb2e7faa543cfd/pterotactyl/reconstruction/vision/model.py#L274 Because the mask embedding is fixed throughout the entire training process, shouldn't we use the same embedding variable computed at the first deformation loop? I am wondering whether computing it twice results in a suboptimal optimisation of the mask embedding network during backpropagation.
I am aware that this is a minor issue, so I am just asking to make sure that I am not missing an important detail. Thanks!