Check out this work by Alexandre Milesi et al. from Nvidia. They managed to speed up training of the SE(3)-Transformer by up to 21(!) times and reduced memory consumption by up to 43 times. Code here.
This repository is the official implementation of SE(3)-Transformers: 3D Roto-Translation Equivariant Attention Networks.
Please cite us as
@inproceedings{fuchs2020se3transformers,
title={SE(3)-Transformers: 3D Roto-Translation Equivariant Attention Networks},
author={Fabian B. Fuchs and Daniel E. Worrall and Volker Fischer and Max Welling},
year={2020},
booktitle = {Advances in Neural Information Processing Systems 34 (NeurIPS)},
}
update 2020/12/22: we made some updates to support dgl >= dgl0.5.0; it is tested with the combinations dgl0.5.3 & torch 1.7 and dgl0.4.3.post2 & torch 1.4 We recommend using dgl 0.5.3 & torch 1.7
- Install this repo in your virtual environment as
pip install -e .(this is important for easier importing from parent folders) - Pytorch
- DGL
- pip install dgl-cu110
dgl 0.4.3 heads-up : this part is a bit tricky to make work correctly; for uspip install dgl-cu90==0.4.3.post2worked; if you use a different version, you might need to do some debugging (I believe expected datatypes for some interfaces changed)- to check which CUDA version pytorch is using:
python -c "import torch; print(torch.version.cuda)" - check here for compatibility of DGL with CUDA etc.
- e.g. for CUDA 9.0:
pip install dgl-cu90==0.4.3.post2 - e.g. for CUDA 10.1:
pip install dgl-cu101==0.4.3.post2 - e.g. for CUDA 10.2:
pip install dgl-cu102==0.4.3.post2
- to check which CUDA version pytorch is using:
if you get the error “libcublas.so.10: cannot open shared object file: No such file or directory”, running this command might help:pip install torch-cluster==latest+cu101 -f https://pytorch-geometric.com/whl/torch-1.4.0.html- please help us making this part more robust: tell us what you did to make it work on your specific system and we'll put it here
pip install packaging- optional: Weights & Biases
- install with
pip install wandb
- install with
- optional: install the lie learn library via:
pip install git+https://github.com/AMLab-Amsterdam/lie_learn - optional: for testing speed of different parts of the code, we used https://github.com/pyutils/line_profiler
Check requirements.txt for other dependencies
The code for experiments specific is meant to be placed in the folder experiments.
We provide an implementation for the n-body experiment, the QM9 experiments. In addition, we provide a multi-step toy-optimisation experiment losely inspired the by protein structure prediction task.
The SE(3)-transformer is built on top of the DGL in Pytorch.
###
# Define a toy model: more complex models in experiments/qm9/models.py
###
# The maximum feature type is harmonic degree 3
num_degrees = 4
# The Fiber() object is a representation of the structure of the activations.
# Its first argument is the number of degrees (0-based), so num_degrees=4 leads
# to feature types 0,1,2,3. The second argument is the number of channels (aka
# multiplicities) for each degree. It is possible to have a varying number of
# channels/multiplicities per feature type and to use arbitrary feature types,
# for this functionality check out fibers.py.
fiber_in = Fiber(1, num_features)
fiber_mid = Fiber(num_degrees, 32)
fiber_out = Fiber(1, 128)
# We build a module from:
# 1) a multihead attention block
# 2) a nonlinearity
# 3) a TFN layer (no attention)
# 4) graph max pooling
# 5) a fully connected layer -> 1 output
model = nn.ModuleList([GSE3Res(fiber_in, fiber_mid),
GNormSE3(fiber_mid),
GConvSE3(fiber_mid, fiber_out, self_interaction=True),
GMaxPooling()])
fc_layer = nn.Linear(128, 1)
###
# Run model: complete train script in experiments/qm9/run.py
###
# Before each forward pass we make a call to get_basis_and_r, which computes
# the equivariant weight basis and relative positions of all the nodes in the
# graph. Pass these variables as keyword arguments to SE(3)-transformer layers.
basis, r = get_basis_and_r(G, num_degrees-1)
# Run SE(3)-transformer layers: the activations are passed around as a dict,
# the key given as the feature type (an integer in string form) and the value
# represented as a Pytorch tensor in the DGL node feature representation.
features = {'0': G.ndata['my_features']}
for layer in model:
features = layer(features, G=G, r=r, basis=basis)
# Run non-DGL layers: we can do this because GMaxPooling has converted features
# from the DGL node feature representation to the standard Pytorch tensor rep.
output = fc_layer(features)Type issues with QM9 experiments
One user reported that they experienced issues with data types when running the QM9 experiments. For them, adding the following lines just before line 184 of qm9.py fixed the issue:
x=x.astype(np.float32)
one_hot=one_hot.astype(np.float32)
atomic_numbers=atomic_numbers.astype(np.float32)
Speed
-
Constructing equivariant layers is indeed computationally expensive, one of the bottlenecks being the computation of the spherical harmonics.
-
We did put significant effort into speeding up the computation of spherical harmonics, in part by parallelising the computation on the GPU (the paper goes into more depth about what we did). A key challenge here is that they depend on the relative distances and hence are different for each point cloud / graph.
-
Further speeding up SE3-Transformer type approaches (approaches working with irreducible representations, that is) and potentially making them more memory efficient would certainly be a big step forward. If I think about why translation-equivariance is part of so many current network architectures (whether the task actually has translational symmetry or not) and other symmetries are not, the reason is most likely computational efficiency. So, if anyone is interested in researching this direction, we can only encourage them.
Here are some ideas about speeding up the SE3-Transformer:
-
Depending on the dataset and your system requirements, it might be possible to cache the spherical harmonics / basis vectors. This could tremendously speed things up as it basically addresses all the bottlenecks at once.
-
A not-so-puristic alternative/version of the above is voxelising the input. This destroys exact equivariance but makes it much easier to cache the spherical harmonics as it reduces the number of overall evaluations.
-
Lower hanging fruit: check what part of the network you can make slimmer for a specific task. E.g., the number of degrees makes a big difference in speed & memory and the effect on performance saturates.
The code in the subfolder equivariant_attention/from_se3cnn is strongly based on https://github.com/mariogeiger/se3cnn which accompanies the paper '3D Steerable CNNs: Learning Rotationally Equivariant Features in Volumetric Data' by Weiler et al.
Please contact us at: fabian @ robots . ox . ac . uk
MIT License
Copyright (c) 2020 Fabian Fuchs and Daniel Worrall
Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.