PotentialNet edge dimension error

[and-tos]

Hi,
I want to train a model using PotentialNet. The code fails due to an error with the edge dimensions:

Traceback (most recent call last):
  File "contextlib.py", line 130, in __exit__
python-BaseException
    self.gen.throw(type, value, traceback)
  File "dgl/heterograph.py", line 5614, in local_scope
    yield
  File "dgllife/model/model_zoo/potentialnet.py", line 238, in forward
    eids = graph.edata['e'][:,i].nonzero(as_tuple=False).view(-1).type(graph.idtype)
IndexError: index 8 is out of bounds for dimension 1 with size 8

There seems to be a mismatch with the constructed graphs.
According to a comment in the code:
n_etypes=n_etypes, # num_distance_bins + 5 covalent types
n_etypes should have length 9, but it is 8.

>>>distance_bins
[1.5, 2.5, 3.5, 4.5]
>>>d_one_hot.shape
(1384, 3)
>>>complex_knn_graph.edata['e'].shape
torch.Size([2068, 8])

I'd really appreciate your help!

[mufeili]

How did you modify the code? Can you provide a minimal script to reproduce the error?

[and-tos]

I did not make any changes to the code, but want to use it to run on a private dataset. For this I defined the following configuration:

PotentialNet_inhouse = {
    'dataset': 'PDBBind',
    'task': 'Ki_inhouse', 
    'subset': '',
    'load_binding_pocket': True,
    'random_seed': 123,
    'frac_train': 0.6,
    'frac_val': 0.2,
    'frac_test': 0.2,
    'batch_size': 200,
    'shuffle': False,
    'max_num_neighbors': 5,
    'distance_bins': [1.5, 2.5, 3.5, 4.5],
    'f_in': 44,
    'f_bond': 48,  # has to be larger than f_in
    'f_gather': 48,
    'f_spatial': 48,  # better to be the same as f_gather
    'n_rows_fc': [48, 24],
    'n_bond_conv_steps': 2,
    'n_spatial_conv_steps': 1,
    'dropouts': [0.25, 0.25, 0.25],
    'lr': 0.01,
    'num_epochs': 100,
    'wd': 1e-05,
    'metrics': ['r2', 'mae'],
    'split': 'temporal'
}

Apologies, I cannot provide an actual example right now, but would appreciate any hint on what to look for.

[and-tos]

The code works if I set:
distance_bins': [1.5, 2.5, 3.5]
I suggest that

def int_2_one_hot(a):
    """Convert integer encodings on a vector to a matrix of one-hot encoding"""
    n = len(a)
    b = np.zeros((n, a.max()+1))
    b[np.arange(n), a] = 1
    return b

changes to

def int_2_one_hot(a, distance_bins):
    """Convert integer encodings on a vector to a matrix of one-hot encoding"""
    n = len(a)
    b = np.zeros((n, len(distance_bins))
    b[np.arange(n), a] = 1
    return b

[mufeili]

Sorry for the late reply. #170 should fix the issue.