Running Inference for a large organic molecule #155
Comments
|
Read the error and do what it says
…On Fri, Nov 25, 2022, 07:40 Nishi Maru ***@***.***> wrote:
I'm trying to run inference on an organic molecule of 47 atoms. I'm using
the equivariant transformer pretrained on the ANI-1 dataset. I passed
atomic numbers and the coordinates as tensors and am interested to get the
energy. On running the code, it is giving the following error:
Traceback (most recent call last):
File "//test1.py", line 35, in <module>
energy, forces = model(z, pos)
File "/opt/conda/envs/torchmd-net/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1110, in _call_impl
return forward_call(*input, **kwargs)
File "/torchmdnet/models/model.py", line 171, in forward
x, v, z, pos, batch = self.representation_model(z, pos, batch, q=q, s=s)
File "/opt/conda/envs/torchmd-net/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1110, in _call_impl
return forward_call(*input, **kwargs)
File "/torchmdnet/models/torchmd_et.py", line 161, in forward
edge_index, edge_weight, edge_vec = self.distance(pos, batch)
File "/opt/conda/envs/torchmd-net/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1110, in _call_impl
return forward_call(*input, **kwargs)
File "/torchmdnet/models/utils.py", line 221, in forward
assert not (
AssertionError: The neighbor search missed some atoms due to max_num_neighbors being too low. Please increase this parameter to include the maximum number of atoms within the cutoff.
—
Reply to this email directly, view it on GitHub
<#155>, or unsubscribe
<https://github.com/notifications/unsubscribe-auth/AB3KUORZ4XGAWNEKW3RJDQLWKBNMNANCNFSM6AAAAAASLAEGSA>
.
You are receiving this because you are subscribed to this thread.Message
ID: ***@***.***>
|
Increased the max_num_neighbour value from 32 (default) to 64, 128, 256. But it does not solve the error. |
|
Are you sure your coordinates are in Angstrom and not nanometers? That would increase the density by enough to make it unusable. |
yes, The coordinates are in Angstrom. |
|
Actually this should not matter. If your molecule is 47 atoms, setting the max_num_neighbour to anything above 47 would trivially solve it anyway. Can you paste here |
|
correction: molecule has 48 atoms.
|
|
I see you are running it from a custom script. Can you show me how you set the max_num_neighbour attribute? Or the whole script even better. |
|
I'm running the inference code mentioned in the examples. I'm passing the atomic coordinates as tensor (not using random function). I changed the cut off values in the files mentioned in the error.
|
|
You are not supposed to modify the source code, changing the function argument defaults would probably not do anything if the caller is passing a specific value to it. You should set it in load_model. I guess the documentation is not very clear on that. model = load_model("ANI1-equivariant_transformer/epoch=359-val_loss=0.0004-test_loss=0.0120.ckpt", derivative=True, max_num_neighbors=50) |
|
Thanks, Stefan. It worked. |
|
Great to hear that :) |
I'm trying to run inference on an organic molecule of 47 atoms. I'm using the equivariant transformer pretrained on the ANI-1 dataset. I passed atomic numbers and the coordinates as tensors and am interested to get the energy. On running the code, it is giving the following error:
The text was updated successfully, but these errors were encountered: