search

cimm-kzn / 3D-MIL-QSAR

QSAR platform for molecular property prediction with 2D and 3D models
53 stars 17 forks source link

Docstrings for wrappers and attention_nets & refactor #11

Closed MariBerry closed 2 years ago

MariBerry commented 2 years ago

Here are several commits, but one is the most big and important. Note please, that MBSplitter was modified in the first one of these commits, but then i removed it at all, so this modification doesnt matter now. Main changes are: commit 51a85790b7d5aa92752c7cdbae8f11456e8471cd

  1. Added docs to wrappers, attention_nets
  2. Moved train_val_split() to utils. Previosly it had distinct implementations for wrappers and BaseNet, because of need to add_padding, but we moved this padding to BaseNet.fit().
  3. Removed array_to_tensor(). y reshape and cuda transfer is now done inplace in MLP.fit() and BaseNet.fit() methods.
  4. Removed MBSplitter(). All we needed can be done by TensorDataset()
  5. Moved get_mini_batches() to utils. Previosly it had distinct implementations for wrappers and BaseNet, because of need to handle variable size input, but we use now *args.
  6. ndim=None is removed from all wrappers and AtentionNet. None initialized network couldnt train.