Dimension mismatch for einsum() in ops.py

[daniel-dpq]

When using DAP, I encountered the following error:

  File "/home/pdengad/protein/fastfold/fastfold/model/fastnn/ops.py", line 153, in forward
    norm = torch.einsum('bsid,bsjd->bijd', M_mask_col, M_mask) + 1e-3
  File "/home/pdengad/anaconda3/envs/fastfold/lib/python3.8/site-packages/torch/functional.py", line 360, in einsum
    return _VF.einsum(equation, operands)  # type: ignore[attr-defined]
RuntimeError: einsum(): the number of subscripts in the equation (4) does not match the number of dimensions (5) for operand 0 and no ellipsis was given

I am running multimer mode. In my case, M_mask has 5 dimensions: [1, batch_size, N_seq, N_res, 1]. The first dimension (1) is from line 235 in fastnn.msa.py: msa_mask = msa_mask.unsqueeze(0). The last dimension is from line 149 in fastnn.ops.py: M_mask = M_mask.unsqueeze(-1). I am wondering why there should be only 4 dimensions. Here are the dimensions of some input features:

aatype torch.Size([1, 134, 1])
residue_index torch.Size([1, 134, 1])
msa torch.Size([1, 252, 134, 1])
asym_id torch.Size([1, 134, 1])
sym_id torch.Size([1, 134, 1])
entity_id torch.Size([1, 134, 1])
seq_mask torch.Size([1, 134, 1])
msa_mask torch.Size([1, 252, 134, 1])
target_feat torch.Size([1, 134, 21, 1])
extra_msa torch.Size([1, 260, 134, 1])
extra_deletion_matrix torch.Size([1, 260, 134, 1])
extra_msa_mask torch.Size([1, 260, 134, 1])
msa_feat torch.Size([1, 252, 134, 49, 1])

Thanks for any kind help.

[semal]

Maybe you should add the batch dimension to parameters too.

[Gy-Lu]

Hi, it seems that these features were not produced by FastFold preprocess. Did you load a pickle file or other things as features?

[daniel-dpq]

Thanks for your reply. Yes, I load my own features. I finally figure out that I need to remove the batch_size dimension in the input features.