Checkpointed crossdocked_full_atom_cond atom features has no "other" feature in dict

[SanazKaz]

Hello,

When i try to train from the checkpoint of crossdocked_full_atom_cond and my data is in full atom mode using aa_encoder for crossdocked full -

   'aa_encoder': {'C': 0, 'N': 1, 'O': 2, 'S': 3, 'B': 4, 'Br': 5, 'Cl': 6, 'P': 7, 'I': 8, 'F': 9, 'others': 10}, - 11 dimensions.

The uploaded model atom_nf in the keys has 10 features ( missing 'other' in the dict)

i am wonder if the uploaded model has actually used c alpha and not full atom encodings?

please see debugging prints -

checkpoint = torch.load(
CHECKPOINT KEYS: ['epoch', 'global_step', 'pytorch-lightning_version', 'state_dict', 'loops', 'callbacks', 'optimizer_states', 'lr_schedulers', 'hparams_name', 'hyper_parameters', 'legacy_pytorch-lightning_version']

Atom encoder weight shape: torch.Size([20, 10])
Atom encoder weight shape: torch.Size([32, 20])

ss: n_dims=3, atom_nf=10
xh0_lig shape: torch.Size([417, 14])
DEBUG - noised_representation input shapes:
xh_lig shape: torch.Size([417, 14])
n_dims: 3, atom_nf: 10
Expected eps_lig shape: (417, 13)
xh_lig.shape[1]: 14
Sampled eps_lig shape: torch.Size([417, 14])