Hi @yilunliao,

Thanks for the nice codebase - I am adapting it for another purpose, and I was running into some issues when checking the outputs are actually equivariant. Are there any init flags that must be set in a certain way to guarantee equivariance?

I have a snippet equivalent to this:

import torch_geometric
import torch
from e3nn import o3
from torch_geometric.data import Data
from nets.equiformer_v2.equiformer_v2_oc20 import EquiformerV2_OC20

edge_index = torch.tensor([[0, 1, 1, 2],
                           [1, 0, 2, 1]], dtype=torch.long)
pos = torch.randn(10, 3)
data = Data(pos=pos, edge_index=edge_index)

R = torch.tensor(o3.rand_matrix())

model = EquiformerV2_OC20(
        num_layers=2,
        attn_hidden_channels=16,
        ffn_hidden_channels=16,
        sphere_channels=16,
        edge_channels=16,
        alpha_drop=0.0, # Turn off dropout for eq
        drop_path_rate=0.0, # Turn off drop path for eq
    )

energy1, forces1 = model(data)
rotated_pos = torch.matmul(pos, R)
data.pos = rotated_pos
energy2, forces2 = model(data)

assert energy1 == energy2
assert torch.allclose(forces1, torch.matmul(forces2, R), atol=1.0e-3)

and the energies are equal, but the forces do not obey equality under rotation. I've turned off all dropout and set the model to eval - just wondering if there are any other tricks to retain the genuine eq behaviour. Thanks!

Could you provide instructions on how to run experiments under the multi-node multi-gpu setting without using the submitit? For example, I have 2 nodes, each of which contains 16 gpus. How should I modify the scripts you provide to reproduce the reported results?

Thanks!

Hi, I am trying to integrate this with the e3nn package.

For the SO3Embedding class, how can I convert that to an irrep which is compatible with the convention e3nn?
My implementation (not sure this is right or not)

    def to_e3nn_embeddings(self):
        from e3nn.io import SphericalTensor
        from e3nn.o3 import Irreps
        embedding = self.embedding.reshape(self.length, -1)

        l = o3.Irreps(str(SphericalTensor(self.lmax_list[-1], 1, -1)).replace('1x', f'{self.num_channels}x'))
        # multiple channels
        return l, embedding

Hi,

I've heard of this strong model which can learn atomic coordinates. Now I want to adapt this model for my project, but I find the code is a bit complicated and hard to follow. (The paper also contains some concepts hard to understand)

I want to make sure, if this model can learn the following task:

**INPUTS**: protein_atom_pos, protein_atom_types, init_molecule_pos, init_molecule_types
**OUTPUTS**: molecule_pos, molecule_types

In another word, I want to learn the molecule given protein pocket, known as Structure-based Drug Design.

Please help me verify if this model suitable for SE(3)-invariant molecule learning, and indicate the relavant code pieces.

Thanks!

Hello, I have a question about eSCN.
I have read the section "A.3 ESCN CONVOLUTION" of your paper, and I wonder is there an implicit condition for the summation running the irreps order L_i and L_f? More specifically, assuming tensor product "1ox(0e+1o+2e)", I want to consider "1o" as output irreps. So, how does equiformerv2 distinguish the odd irreps "1o" that comes from "1ox(0e+2e)" and even irreps "1e" that comes from "1ox1o"?

Dear Developer,

I am writing to inquire about the rescale_factor = math.sqrt(length / (2 * mmax + 1)) defined in the SO3_Grid and CoefficientMappingModule classes within SO3.py. Could you please explain its purpose? Additionally, would omitting rescale_factor affect the network's equivariance?

Thank you for your assistance.

Best regards,

Yang Zhong

I wanted to train this network on the spice dataset (similar task where I want to predict forces and energy from structure). I was comparing speed of training with torchmd-net (https://github.com/torchmd/torchmd-net). For same parameter count, torchmd-net is at least thrice faster as compared to Equiformer and twice as compared to Equiformer-v2. Is this something common or a bug on my end?

Hi!

Very cool work! It would be nice to have the model checkpoints on the Hugging Face Hub rather than a Dropbox link

Some of the benefits of sharing your models through the Hub would be:

• versioning, commit history and diffs
• repos provide useful metadata about their tasks, languages, metrics, etc that make them discoverable
• multiple features from TensorBoard visualizations, PapersWithCode integration, and more
• wider reach of your work to the ecosystem

Creating the repos and adding new models should be a relatively straightforward process if you've used Git before. This is a step-by-step guide explaining the process in case you're interested. Please let us know if you would be interested and if you have any questions.

As the title states，How to use the v2 to train IS2RE datasets? I think I need some tutorials.

Hi , not issues but some questions.

1. Any comparison of the performace to some other SOTA equvariant nets such as MACE or Nequip or sth ?

2. Is MD simulation available ? or any developments going on?

Thanks for the nice work

I looked into codes and the paper, and it seems like equiformer predict force directly, rather than using gradient of energy. Although there is pros and cons, I'm curious whether is there any plan to making force predictions from energy gradient, which could be implemented using autograd.

I have a dataset with molcules and their atoms, positions, energy, and forces information. I wonder how to train my dataset with equiformer_v2. Is it possible to support customied dataset training or at least give some tutorial or suggestion?

Hi, thanks for your wonderful work and code! I am new to the equivariant learning area, and I am trying to understand each step of your code. I suspect this function (the function for calculating the edge rotation matrix) would break the equivariance during the edge-degree embedding layer.

In this function, you set the original edge_distance_vec as the final x_axis, and randomly select the y-axis and generate the z_axis according to the other two axes. I think this random operation is not correct.

For example, suppose I send two same molecules into the network, where the coordinates of one molecule are translated relative to the other molecule, but not rotated. If the network is SE(3)-equivariant, the final feature should be the same for these two molecules. However, during the above function, they will have different edge rotation matrices on the same edge since the randomly selected y_axis, and the corresponding Wigner-D matrix would be different, thus during the edge-degree embedding layer, the embedding of the two molecules are different since this step uses the Wigner-D matrix. This breaks the equivariance for all features with type>0.

I am not sure if I am correct. Looking forward to your opinion on this issue.

Hi, nice work! I was wondering what it would take to accommodate for systems with nodes that have additional non-scalar features. Any hints or snippets would be greatly appreciated. Thanks.

Thank you for the great work EquiformerV2. When I test its performance on MD17/22 dataset, I find it lags far behind SOTA models like VisNet. For example, in MD22_AT_AT, when VisNet val loss for E converges to 0.14, F converges to 0.17. While for EqV2 E val loss is 4.7 for E and 5.1 for F.
I follow the setting in oc20/configs/s2ef/all_md/equiformer_v2/equiformer_v2_N@8_L@4_M@2_31M.yml. Are there things I need to modify for adopting EqV2 in MD datasets? Thanks.

Hello, thank you for sharing your great work!

I was wondering if it would be possible for you to share the training logs, similar to how the logs were provided for Equiformer. I'm specifically interested in seeing the higher precision test results. For example, for mu target 0, the reported results in the paper are limited to 0.11, but the Equiformer logs show higher precision results of 0.1172.

If you're able to share the logs, I would greatly appreciate it as it would allow for a more detailed analysis of the impressive results you achieved.

Thank you again for this excellent contribution to the field. I look forward to hearing back from you @yilunliao.

Best regards,

Thanks for putting together this awesome repo!

I'm working on a VQVAE model for the backbone atoms of protein 3D structures and thinking about using Equiformer V2 as my first part the encoder. I used the original SE(3) model before, but it was super computationally heavy.

My plan is to quantize protein 3D structures residue-wise based on the number of carbon alpha atoms and only pass the coordinates to the encoder, leaving out the amino acid type info.

Do you have any suggestions or thoughts on how well this setup might work?