The main contribution of this PR is to solve the problem detailed below.
In #31, we added the --post_pool_filter script argument. Later, we realized that its implementation is not optimal because the instantiated TEN model will contain lots of trainable parameters that will never be updated (because the ranks they are associated with are filtered out before the prediction).
This PR solves that problem by renaming --post_pool_filter to --visible_dims, and makes changes to the TEN model architecture to instantiate its submodules depending on this argument. 3b94a7231bdfa780fa61287d3cfb380bd8da6814, 0c6017d79c093cdb02718372d0592a909573c329, 8caa16aac30c0b752fed8b9d921d6141d84a0e6a, f896b097f60533b5f9dc202ab2ace4ea25f45aad
This PR also contains the following small improvements:
Simplify the --compile flag. Now, just setting it is enough to trigger compilation. bfe10bc20e4983e50496f7891468ca39624db6ee
Print more details about the model architecture, and do this before the wandb initialization. 44d65dc88f393ba5cb8ce8459988fc764414d907
Rename the sbatch script train_egnn_like.sh to train.sh and make some small changes. 0adb5c9f02dedc7e05e4c8508d6bafad3019c835
Remove gradient clipping to get closer to the original EGNN implementation cd690fc4339e8a917c5972e56e982d51316f46d0
In the loss function, replace reduction=sum with reduction=mean to get closer to the original EGNN implementation cd690fc4339e8a917c5972e56e982d51316f46d0
Goal
The main contribution of this PR is to solve the problem detailed below.
In #31, we added the
--post_pool_filter
script argument. Later, we realized that its implementation is not optimal because the instantiated TEN model will contain lots of trainable parameters that will never be updated (because the ranks they are associated with are filtered out before the prediction).This PR solves that problem by renaming
--post_pool_filter
to--visible_dims
, and makes changes to the TEN model architecture to instantiate its submodules depending on this argument. 3b94a7231bdfa780fa61287d3cfb380bd8da6814, 0c6017d79c093cdb02718372d0592a909573c329, 8caa16aac30c0b752fed8b9d921d6141d84a0e6a, f896b097f60533b5f9dc202ab2ace4ea25f45aadThis PR also contains the following small improvements:
--compile
flag. Now, just setting it is enough to trigger compilation. bfe10bc20e4983e50496f7891468ca39624db6eewandb
initialization. 44d65dc88f393ba5cb8ce8459988fc764414d907sbatch
scripttrain_egnn_like.sh
totrain.sh
and make some small changes. 0adb5c9f02dedc7e05e4c8508d6bafad3019c835reduction=sum
withreduction=mean
to get closer to the original EGNN implementation cd690fc4339e8a917c5972e56e982d51316f46d0