Home

Awesome

TorchDrug

<h1 align="center"> with <a href="https://torchprotein.ai/"> <img src="asset/torchprotein_logo_tight.svg" alt="TorchProtein" style="height:26px" /> </a> </h1>

Open in Colab Contributions License Apache-2.0 PyPI downloads TorchDrug Twitter

Docs | Tutorials | Benchmarks | Papers Implemented

TorchDrug is a PyTorch-based machine learning toolbox designed for several purposes.

Installation

TorchDrug can be installed on either Linux, Windows or macOS. It is compatible with 3.7 <= Python <= 3.10 and PyTorch >= 1.8.0.

From Conda

conda install torchdrug -c milagraph -c conda-forge -c pytorch -c pyg

From Pip

pip install torch==1.9.0
pip install torch-scatter torch-cluster -f https://pytorch-geometric.com/whl/torch-1.9.0+cu102.html
pip install torchdrug

To install torch-scatter for other PyTorch or CUDA versions, please see the instructions in https://github.com/rusty1s/pytorch_scatter

From Source

git clone https://github.com/DeepGraphLearning/torchdrug
cd torchdrug
pip install -r requirements.txt
python setup.py install

Windows (PowerShell)

We need to first install the build tools for Visual Studio. We then install the following modules in PowerShell.

Install-Module Pscx -AllowClobber
Install-Module VSSetup

Initialize Visual Studio in PowerShell with the following commands. We may setup this for all PowerShell sessions by writing it to the PowerShell profile. Change the library path according to your own case.

Import-VisualStudioVars -Architecture x64
$env:LIB += ";C:\Program Files\Python37\libs"

Apple Silicon (M1/M2 Chips)

We need PyTorch >= 1.13 to run TorchDrug on Apple silicon. For torch-scatter and torch-cluster, they can be compiled from their sources. Note TorchDrug doesn't support mps devices.

pip install torch==1.13.0
pip install git+https://github.com/rusty1s/pytorch_scatter.git
pip install git+https://github.com/rusty1s/pytorch_cluster.git
pip install torchdrug

Quick Start

TorchDrug is designed for humans and focused on graph structured data. It enables easy implementation of graph operations in machine learning models. All the operations in TorchDrug are backed by PyTorch framework, and support GPU acceleration and auto differentiation.

from torchdrug import data

edge_list = [[0, 1], [1, 2], [2, 3], [3, 4], [4, 5], [5, 0]]
graph = data.Graph(edge_list, num_node=6)
graph = graph.cuda()
# the subgraph induced by nodes 2, 3 & 4
subgraph = graph.subgraph([2, 3, 4])

Molecules are also supported in TorchDrug. You can get the desired molecule properties without any domain knowledge.

mol = data.Molecule.from_smiles("CCOC(=O)N", atom_feature="default", bond_feature="default")
print(mol.node_feature)
print(mol.atom_type)
print(mol.to_scaffold())

You may also register custom node, edge or graph attributes. They will be automatically processed during indexing operations.

with mol.edge():
	mol.is_CC_bond = (mol.edge_list[:, :2] == td.CARBON).all(dim=-1)
sub_mol = mol.subgraph(mol.atom_type != td.NITROGEN)
print(sub_mol.is_CC_bond)

TorchDrug provides a wide range of common datasets and building blocks for drug discovery. With minimal code, you can apply standard models to solve your own problem.

import torch
from torchdrug import datasets

dataset = datasets.Tox21()
dataset[0].visualize()
lengths = [int(0.8 * len(dataset)), int(0.1 * len(dataset))]
lengths += [len(dataset) - sum(lengths)]
train_set, valid_set, test_set = torch.utils.data.random_split(dataset, lengths)
from torchdrug import models, tasks

model = models.GIN(dataset.node_feature_dim, hidden_dims=[256, 256, 256, 256])
task = tasks.PropertyPrediction(model, task=dataset.tasks)

Training and inference are accelerated by multiple CPUs or GPUs. This can be seamlessly switched in TorchDrug by just a line of code.

from torchdrug import core

# Single CPU / Multiple CPUs / Distributed CPUs
solver = core.Engine(task, train_set, valid_set, test_set, optimizer)
# Single GPU
solver = core.Engine(task, train_set, valid_set, test_set, optimizer, gpus=[0])
# Multiple GPUs
solver = core.Engine(task, train_set, valid_set, test_set, optimizer, gpus=[0, 1, 2, 3])
# Distributed GPUs
solver = core.Engine(task, train_set, valid_set, test_set, optimizer, gpus=[0, 1, 2, 3, 0, 1, 2, 3])

Experiments can be easily tracked and managed through Weights & Biases platform.

solver = core.Engine(task, train_set, valid_set, test_set, optimizer, logger="wandb")

Contributing

Everyone is welcome to contribute to the development of TorchDrug. Please refer to contributing guidelines for more details.

License

TorchDrug is released under Apache-2.0 License.