Author: Hang'rui Bi ([email protected])

Thie repository provides high-level neural network layers for constructing SOTA E3-equivariant neural networks based on e3nn. It also provides examples for several applications, such as potential energy surface modeling, dipole prediction, hamiltonian prediction, as well as score-based conformation generation. Some codes are modified from Nequip. The notebook for infer matrix transformation from data is modified from scripts written by Mario Geiger. The scripts related to score based generative modeling are modified from https://github.com/yang-song/score_sde_pytorch.

Once the environment is prepared, you can run this repository. There is no need to install this repository itself.

You may prepare the environment by running pip install -r requirements.txt. Or execute the following commands after installing python3.8.

pip install --upgrade pip wheel setuptools
pip install torch==1.11.0+cu115 -f https://download.pytorch.org/whl/torch_stable.html

TORCH=1.11.0
CUDA=11.5
pip install --upgrade --force-reinstall https://data.pyg.org/whl/torch-1.11.0%2Bcu115/torch_cluster-1.6.0-cp38-cp38-linux_x86_64.whl
pip install --upgrade --force-reinstall https://data.pyg.org/whl/torch-1.11.0%2Bcu115/torch_scatter-2.0.9-cp38-cp38-linux_x86_64.whl
pip install --upgrade --force-reinstall https://data.pyg.org/whl/torch-1.11.0%2Bcu115/torch_sparse-0.6.13-cp38-cp38-linux_x86_64.whl
pip install --upgrade --force-reinstall https://data.pyg.org/whl/torch-1.11.0%2Bcu115/torch_spline_conv-1.2.1-cp38-cp38-linux_x86_64.whl
pip install torch-geometric 

pip install e3nn wandb jupyterlab absl-py ml-collections ase h5py torch_runstats torch-ema

You may pull the image from docker hub docker pull dahubao/geometric:e3nn_py38_cu115_torch11.0, or build a docker image with the following command.

# python 3.8 cuda 11.5 pytorch 1.11.0

docker run -it --name geometric --gpus all --network host nvidia/cuda:11.5.0-cudnn8-devel-ubuntu20.04 bash

# in the container, run:
apt update
apt install --assume-yes python3 python3-pip tmux
# execute the commands in the section #using pip to install the python packages 

The image can also be used by singularity.

Notice that the command line args only contains those that do not affect the expected experiment result. For example, the address to bind and the seed for the RNG. The hyperparameters of the experiment, such as learning rate and model structure, should be managed by git, and are therefore provided in the config file instead. Before running, prepare the datasets (refer to the examples section for examples) and put them in the path specified in the config file.

# if you are using a container
# docker run -it --network host -v ~:/root --name <image> --gpus all image  bash 
# or
# singularity shell --nv <image>

python3  train.py --config <name_of_the_config>

Notice that the argument <name_of_the_config> is not a filename, you should run python3 train.py --config config_energy instead of python3 train.py --config e3_layers/configs/config_energy.py. The list of available configs are registered in e3_layers/configs/__init__.py. You may run python3 train.py --help to inspect all command line args.

python3 inference.py --config <name_of_the_config> --model_path <model_path> --output_keys <key1, key2,...> --output_path <output_path>

You can download the processed data from https://drive.google.com/drive/folders/1tXW7LabtOJapgs-AZFMhnI46do0-YS30?usp=sharing.

Run python3 train.py --config config_energy. Jointly training on energy and force is also supported. The accuracy is comparable to Nequip on QM9 (total energy MAE 5.0 meV).

First, clone https://github.com/yang-song/score_sde_pytorch and add it to your environment variable PYTHONPATH. The prior for position is randn(n_nodes, 3), so you need to set the argument r_max at least 7. You may set up data.std in the config file to scale the input by 1/data.std, such that the variance is similar before and after perturbation. Run

# small organic molecules
 python3 train.py --sde_config ../score_sde_pytorch/configs/vp/cifar10_ncsnpp_continuous.py  --workdir results --config config_diffusion    --name test   --project diffusion --wandb
 
# protein
 python3 train.py --sde_config ../score_sde_pytorch/configs/vp/cifar10_ncsnpp_continuous.py  --workdir results --config config_diffusion_protein   --project diffusion_protein  --seed 0 --name test --wandb
 

The script assumes there is a tensor named pos and reserves the key t for time and species for molecule/atom/residual type.

Run python3 train.py --config config_dipole. The accuracy is comparable to or better than this paper on the dataset it used (dipole MAE 4.1e-4 ev*A). For predicting multipoles of higher degrees, you should decompose them into irreducible representations first.

Run python3 train.py --config config_hamiltonian. Notice that this config only works for H2O computed in ORCA convention. The accuracy is comprable (MAE 1.7e-5 Hatree) to PhiSNet on the water dataset it used.

The input and output of all modules in this repository are instances of e3_layers.data.Data. A Data object is a dictionary of tensors that describes a graph or a point cloud. For each tensor data[key], there may be a correspoding data.attrs[key] = (is_per, irreps). The first item means whether the tensor is defined on graphs, edges, or nodes, while the second is the type of irreducible representation or the number of dimensions. Such annotations help to infer the number of nodes and edges, perform equivariance tests, reshape the tensor, and select one item in a batch. It is assumed that the shapes of tensors are [cat_dim, channel_and_irreps]. The first dimension, cat_dim corresponds to the number of edges/nodes/graphs. The second dimension equals to irreps.dim, e.g. a tensor of type 2x1e is of dimension 6.

The class e3_layers.data.Batch inherits e3_layers.data.Data. A batch object is multiple graphs considered as a single (disconnected) graph. Samples are concatenated along their first dimension. Special keys include edge_index, and face. These indices are increased or decreased when collating into or selecting from a batch. The advantage of doing so (instead of creating a batch dimension) is that no padding is ever needed. It is both dictionary-like (indexed by keys) and list-like (indexed by (list of) integer indices). The key _n_graphs, _n_nodes, _node_segement, _n_edges, and _edge_segement are reserved for separating the concatenated tensors since no padding is used. The key _rotation_matrix is reserved for equivariance tests. Do not override them with keys in your dataset.

The e3_layers.data.CondensedDataset inherits from e3_layers.data.Batch. Refer to data.ipynb for the scripts to convert various data formats to the HDF5 format that this repository uses.

To add a new config, you can add a python file in the configs directory that defines get_config() and mention it in configs/__init__.py. There are already some examples in the directory that you can refer to. The get_config function should return an instance of ml_collections.ConfigDict. You can fully customize the model architecture by specifying config.model_config. The model config contains arguments for each layer in the model, and its structure determines the network module hierarchy.

You may specify a file or some files in a directory as the dataset.

config.data_config.path = 'dataest.hdf5'

## all files in (subfolders of) data/
config.data_config.path = 'data/'

## all .hdf5 files in data/, the string after ':' is an arbitrary python regular expression.
config.data_config.path = 'data/:.+\.hdf5'

## For distributed training, equally distributed among processes
config.data_config.path = ['split1.hdf5', 'split2.hdf5']

You may set data_config.reload=True to reload the dataset. This can be useful if the model is trained using active learning and the dataset changes from on epoch to another. n_train and n_val can be integers (number of data samples) or floats (fraction within the dataset).

Two groups of hyperparamters are particularly important during training.

The first group of hyper-parameters controls the rate of network parameter update, including learning_rate, batch_size, grad_acc and loss_coeffs. A good pratice is:

1. Set the batch size as large as possible as long as it fits in your GPU memory. If the largest possible batch size is still too small (e.g. less than 32), set config.grad_acc for gradient accumulation.
2. Set the loss coeffs such that the total loss is about 1e1 to 1e3 at the begining. If the loss becomes too small during training, the gradient may become inaccurate due to float point round-off.
3. Set the learning rate to a large value (e.g. 1e-2) at the begining and use automatic learning rate decay.

The second group of hyper-parameters controls the rate of learning rate dacay, in another words, how long the model is trained. The groupd includes lr_scheduler_patience, lr_scheduler_factor and epoch_subdivision. The learning rate is multiplied by lr_scheduler_factor if the metrics have not improved in (lr_scheduler_patience+1)/epoch_subdivision epochs. You may fix lr_scheduler_factor to be 0.8 and change the lr_scheduler_patience. This might be the most important hyperparameter that needs manual tuning. The optimal value for lr_scheduler_patience depends on the learning task, the model and the size of the dataset. You may need to try multiple values to find out a proper one.

Examples for running MD simulation with torchMD and E3_layers are provided in torchMD.ipynb. You need to install moleculekit and Parmd in order to use torchMD.

The wrapper class for composing layers together into neural networks. Internally, it sequentially runs its layers and adds the output of each module to the dictionart of inputs. For most layers in this repository, the input keys and output keys can be customized during initialization, and the network topology is induced by matching keys. In this way, this class can support a network topology of arbitrary DAG.

• OneHotEncoding
• RadialBasisEdgeEncoding
• SphericalEncoding. Can be used to generate edge attributes used in convolution, or to embed node orientations if they are not rotationally symmetric.

• PointwiseLinear: A wrapper of e3nn.o3.Linear, biases are only applied to scalars It mixes channels and preserves degree and parity.
• TensorProductExpansion, or SelfMix used to recombine (“mix”) the features different degrees and parities. This does not mix the channels and requires all irreps have the same number of channels.
• ResBlock

A factorized convolution is usually much faster and parameter-economic than a fully connected one (e.g. these in TFN and SE3Transformer). It internally calls tensor product with instruction uvu instead of uvw. Linear layers are inserted before and after the tensor product to mix the channels. By default, it uses a 'self-connection' layer, which typically contributes to most of its trainable parameters. The self-connection layer performs a fully connected tensor product between the input features and the input embedding (e.g. one hot atom type encoding).

A message-passing layer consisting of a convolution, a non-linearity activation function, and optionally a residual connection.

• Pooling
• GradientOutput
• PerTypeScaleShift
• Pairwise: Constructs a pairwise representation (edge features) from pointwise representations (node features).
• TensorProductContraction: Constructs tensor product representations from irreps.

python3  train.py --config <name_of_the_config> --profiling

The results will be save as <workdir>/<name>/profiling.json. It is recommended to inspect the profile with chrome trace viewer. Saving the trace may cause a segfault, which is a known PyTorch isssue.