【Hackathon 10th Spring No.9】NewtonNet模型复现

interatomic_potentials/configs/newtonnet/README.md

@@ -0,0 +1,48 @@
+# NewtonNet
+
+[NewtonNet: a Newtonian message passing network for deep learning of interatomic potentials and forces](https://doi.org/10.1039/D2DD00008C)
+
+## Abstract
+
+NewtonNet is a Newtonian message-passing neural network for learning interatomic potentials and forces. It operates on atomic graphs with radial Bessel basis functions and polynomial cutoff envelopes. The architecture uses a sequence of interaction layers that update both atom-level and force-level representations, enabling accurate prediction of energies and atomic forces while maintaining energy conservation.
+
+## Model
+
+NewtonNet constructs a radius graph from atomic positions and computes pairwise radial basis features. Interaction layers iteratively refine atom representations using Newtonian message passing that preserves physical symmetries. Per-atom energies are summed for the total molecular energy, and forces are obtained as negative energy gradients with respect to positions.
+
+## Training
+
+```bash
+# single-gpu training
+python interatomic_potentials/train.py -c interatomic_potentials/configs/newtonnet/newtonnet_qm9_energy.yaml
+
+# multi-gpu training
+python -m paddle.distributed.launch --gpus="0,1,2,3" interatomic_potentials/train.py -c interatomic_potentials/configs/newtonnet/newtonnet_qm9_energy.yaml
+```
+
+## Validation
+
+```bash
+python interatomic_potentials/train.py -c interatomic_potentials/configs/newtonnet/newtonnet_qm9_energy.yaml Global.do_eval=True Global.do_train=False Global.do_test=False Trainer.pretrained_model_path='your model path(*.pdparams)'
+```
+
+## Testing
+
+```bash
+python interatomic_potentials/train.py -c interatomic_potentials/configs/newtonnet/newtonnet_qm9_energy.yaml Global.do_test=True Global.do_train=False Global.do_eval=False Trainer.pretrained_model_path='your model path(*.pdparams)'
+```
+
+## Citation
+
+```
+@article{haghighatlari2022newtonnet,
+  title={NewtonNet: a Newtonian message passing network for deep learning of interatomic potentials and forces},
+  author={Haghighatlari, Mojtaba and Li, Jie and Guan, Xingyi and Zhang, Oufan and Das, Akshaya and Stein, Christopher J and Heiber, Farnaz and Liu, Tiantian and Head-Gordon, Martin and Bertels, Luke and others},
+  journal={Digital Discovery},
+  volume={1},
+  number={3},
+  pages={333--343},
+  year={2022},
+  publisher={Royal Society of Chemistry}
+}
+```

interatomic_potentials/configs/newtonnet/newtonnet_qm9_energy.yaml

@@ -0,0 +1,89 @@
+Global:
+  do_train: True
+  do_eval: False
+  do_test: False
+
+  label_names: ['energy']
+
+Trainer:
+  max_epochs: 300
+  seed: 42
+  output_dir: ./output/newtonnet_qm9_energy
+  save_freq: 50
+  log_freq: 20
+  start_eval_epoch: 1
+  eval_freq: 1
+  pretrained_model_path: null
+  pretrained_weight_name: null
+  resume_from_checkpoint: null
+  use_amp: False
+  amp_level: 'O1'
+  eval_with_no_grad: False
+  gradient_accumulation_steps: 1
+  best_metric_indicator: 'eval_metric'
+  name_for_best_metric: "energy"
+  greater_is_better: False
+  compute_metric_during_train: True
+  metric_strategy_during_eval: 'epoch'
+  use_visualdl: False
+  use_wandb: False
+  use_tensorboard: False
+
+Model:
+  __class_name__: NewtonNet
+  __init_params__:
+    cutoff: 5.0
+    n_features: 128
+    n_basis: 20
+    n_interactions: 3
+    activation: swish
+    layer_norm: False
+    loss_type: mse_loss
+    force_loss_weight: 100.0
+    property_names: ${Global.label_names}
+
+Optimizer:
+  __class_name__: Adam
+  __init_params__:
+    lr:
+      __class_name__: Cosine
+      __init_params__:
+        learning_rate: 5e-4
+        eta_min: 1e-6
+        by_epoch: False
+
+Metric:
+  energy:
+    __class_name__: paddle.nn.L1Loss
+    __init_params__: {}
+
+Dataset:
+  dataset:
+    __class_name__: QM9Dataset
+    __init_params__:
+      path: "./data/qm9"
+      property_names: ${Global.label_names}
+    num_workers: 4
+    use_shared_memory: False
+  split_dataset_ratio:
+    train: 0.8
+    val: 0.1
+    test: 0.1
+  train_sampler:
+    __class_name__: BatchSampler
+    __init_params__:
+      shuffle: True
+      drop_last: False
+      batch_size: 32
+  val_sampler:
+    __class_name__: BatchSampler
+    __init_params__:
+      shuffle: False
+      drop_last: False
+      batch_size: 32
+  test_sampler:
+    __class_name__: BatchSampler
+    __init_params__:
+      shuffle: False
+      drop_last: False
+      batch_size: 32

interatomic_potentials/configs/schnet/schnet_md17_ethanol.yaml

@@ -0,0 +1,107 @@
+Global:
+  do_train: True
+  do_eval: True
+  do_test: False
+
+  label_names: ['energy']
+
+  graph_converter:
+    __class_name__: FindPointsInSpheres
+    __init_params__:
+      cutoff: 5.0
+
+  prim_eager_enabled: True
+
+
+Trainer:
+  max_epochs: 500
+  seed: 42
+  output_dir: ./output/schnet_md17_ethanol
+  save_freq: 50
+  log_freq: 50
+
+  start_eval_epoch: 1
+  eval_freq: 5
+  pretrained_model_path: null
+  pretrained_weight_name: null
+  resume_from_checkpoint: null
+  use_amp: False
+  eval_with_no_grad: True
+  gradient_accumulation_steps: 1
+
+  best_metric_indicator: 'eval_metric'
+  name_for_best_metric: "energy"
+  greater_is_better: False
+
+
+Model:
+  __class_name__: SchNet
+  __init_params__:
+    n_atom_basis: 64
+    n_interactions: 6
+    n_filters: 64
+    cutoff: 5.0
+    n_gaussians: 25
+    max_z: 100
+    readout: "sum"
+    property_names: ${Global.label_names}
+    data_mean: 0.0
+    data_std: 1.0
+    loss_type: "l1_loss"
+    compute_forces: False
+
+
+Optimizer:
+  __class_name__: Adam
+  __init_params__:
+    lr:
+      __class_name__: Cosine
+      __init_params__:
+        learning_rate: 1e-4
+        eta_min: 1e-7
+        by_epoch: False
+
+
+Metric:
+  energy:
+    __class_name__: IgnoreNanMetricWrapper
+    __init_params__:
+      __class_name__: paddle.nn.L1Loss
+      __init_params__: {}
+
+
+Dataset:
+  train:
+    dataset:
+      __class_name__: MD17Dataset
+      __init_params__:
+        path: "./data/md17"
+        molecule: "ethanol"
+        property_names: ${Global.label_names}
+        build_graph_cfg: ${Global.graph_converter}
+        max_samples: 50000
+    num_workers: 4
+    use_shared_memory: False
+    sampler:
+      __class_name__: BatchSampler
+      __init_params__:
+        shuffle: True
+        drop_last: False
+        batch_size: 64
+  val:
+    dataset:
+      __class_name__: MD17Dataset
+      __init_params__:
+        path: "./data/md17"
+        molecule: "ethanol"
+        property_names: ${Global.label_names}
+        build_graph_cfg: ${Global.graph_converter}
+        max_samples: 10000
+    num_workers: 4
+    use_shared_memory: False
+    sampler:
+      __class_name__: BatchSampler
+      __init_params__:
+        shuffle: False
+        drop_last: False
+        batch_size: 64

interatomic_potentials/configs/schnet/schnet_qm9_U0.yaml

@@ -0,0 +1,109 @@
+Global:
+  do_train: True
+  do_eval: True
+  do_test: False
+
+  label_names: ['energy_U0']
+
+  graph_converter:
+    __class_name__: FindPointsInSpheres
+    __init_params__:
+      cutoff: 10.0
+
+  prim_eager_enabled: True
+
+
+Trainer:
+  max_epochs: 200
+  seed: 42
+  output_dir: ./output/schnet_qm9_U0
+  save_freq: 20
+  log_freq: 50
+
+  start_eval_epoch: 1
+  eval_freq: 5
+  pretrained_model_path: null
+  pretrained_weight_name: null
+  resume_from_checkpoint: null
+  use_amp: False
+  eval_with_no_grad: True
+  gradient_accumulation_steps: 1
+
+  best_metric_indicator: 'eval_metric'
+  name_for_best_metric: "energy_U0"
+  greater_is_better: False
+
+
+Model:
+  __class_name__: SchNet
+  __init_params__:
+    n_atom_basis: 128
+    n_interactions: 6
+    n_filters: 128
+    cutoff: 10.0
+    n_gaussians: 50
+    max_z: 100
+    readout: "sum"
+    property_names: ${Global.label_names}
+    data_mean: -76.1160
+    data_std: 10.3238
+    loss_type: "l1_loss"
+    compute_forces: False
+
+
+Optimizer:
+  __class_name__: Adam
+  __init_params__:
+    lr:
+      __class_name__: Cosine
+      __init_params__:
+        learning_rate: 1e-4
+        eta_min: 1e-7
+        by_epoch: False
+
+
+Metric:
+  energy_U0:
+    __class_name__: IgnoreNanMetricWrapper
+    __init_params__:
+      __class_name__: paddle.nn.L1Loss
+      __init_params__: {}
+
+
+Dataset:
+  train:
+    dataset:
+      __class_name__: QM9Dataset
+      __init_params__:
+        path: "./data/qm9"
+        property_names: ${Global.label_names}
+        build_graph_cfg: ${Global.graph_converter}
+        cache_path: "./data/qm9"
+        overwrite: False
+        filter_unvalid: True
+    num_workers: 4
+    use_shared_memory: False
+    sampler:
+      __class_name__: BatchSampler
+      __init_params__:
+        shuffle: True
+        drop_last: False
+        batch_size: 64
+  val:
+    dataset:
+      __class_name__: QM9Dataset
+      __init_params__:
+        path: "./data/qm9"
+        property_names: ${Global.label_names}
+        build_graph_cfg: ${Global.graph_converter}
+        cache_path: "./data/qm9"
+        overwrite: False
+        filter_unvalid: True
+    num_workers: 4
+    use_shared_memory: False
+    sampler:
+      __class_name__: BatchSampler
+      __init_params__:
+        shuffle: False
+        drop_last: False
+        batch_size: 64