SGIRMolecularPredictor¶

Semi-supervised learning for imbalanced regression targets.

from torch_molecule import SGIRMolecularPredictor

Examples · Methods

Constructor¶

class SGIRMolecularPredictor(num_anchor=10, warmup_epoch=20, labeling_interval=5, augmentation_interval=5, top_quantile=0.1, label_logscale=False, lw_aug=1, gamma=0.4, num_task=1, task_type='regression', num_layer=5, hidden_size=300, gnn_type='gin-virtual', drop_ratio=0.5, norm_layer='batch_norm', graph_pooling='sum', augmented_feature=None, batch_size=128, epochs=500, learning_rate=0.001, weight_decay=0.0, grad_clip_value=None, patience=50, use_lr_scheduler=False, scheduler_factor=0.5, scheduler_patience=5, loss_criterion=None, evaluate_criterion=None, evaluate_higher_better=None, verbose='none', device=None, model_name='SGIRMolecularPredictor')[source]¶

This predictor implements SGIR for semi-supervised graph imbalanced regression.

It trains the GREA model based on pseudo-labeling and data augmentation.

References

Semi-Supervised Graph Imbalanced Regression. https://dl.acm.org/doi/10.1145/3580305.3599497
Code: https://github.com/liugangcode/SGIR

Parameters:

num_anchor (int, default=10) – Number of anchor points used to split the label space during pseudo-labeling.
warmup_epoch (int, default=20) – Number of epochs to train before starting pseudo-labeling and data augmentation.
labeling_interval (int, default=5) – Interval (in epochs) between two pseudo-labeling steps. It controls the update frequency of pseudo-labeling.
augmentation_interval (int, default=5) – Interval (in epochs) between two data augmentation steps. It controls the update frequency of data augmentation.
top_quantile (float, default=0.1) – Quantile threshold for selecting high confidence predictions during pseudo-labeling.
label_logscale (bool, default=False) – Whether to use log scale for the label space during pseudo-labeling and data augmentation.
lw_aug (float, default=1) – Loss weight for the augmented data.
gamma (float, default=0.4) – GREA-specific parameter that penalize the size of the rationales (ratio between the number of nodes in the rationales and the number of nodes in the original graph).
num_task (int, default=1) – Number of prediction tasks. SGIR currently only supports regression tasks with 1 task.
task_type (str, default="regression") – Type of prediction task, either “regression” or “classification”.
num_layer (int, default=5) – Number of GNN layers.
hidden_size (int, default=300) – Dimension of hidden node features.
gnn_type (str, default="gin-virtual") – Type of GNN architecture to use. One of [“gin-virtual”, “gcn-virtual”, “gin”, “gcn”].
drop_ratio (float, default=0.5) – Dropout probability.
norm_layer (str, default="batch_norm") – Type of normalization layer to use. One of [“batch_norm”, “layer_norm”, “instance_norm”, “graph_norm”, “size_norm”, “pair_norm”].
graph_pooling (str, default="sum") – Method for aggregating node features to graph-level representations. One of [“sum”, “mean”, “max”].
augmented_feature (list or None, default=None) – Additional molecular fingerprints to use as features. It will be concatenated with the graph representation after pooling. Examples like [“morgan”, “maccs”] or None.
batch_size (int, default=128) – Number of samples per batch for training.
epochs (int, default=500) – Maximum number of training epochs.
loss_criterion (callable, optional) – Loss function for training.
evaluate_criterion (str or callable, optional) – Metric for model evaluation.
evaluate_higher_better (bool, optional) – Whether higher values of the evaluation metric are better.
learning_rate (float, default=0.001) – Learning rate for optimizer.
grad_clip_value (float, optional) – Maximum norm of gradients for gradient clipping.
weight_decay (float, default=0.0) – L2 regularization strength.
patience (int, default=50) – Number of epochs to wait for improvement before early stopping.
use_lr_scheduler (bool, default=False) – Whether to use learning rate scheduler.
scheduler_factor (float, default=0.5) – Factor by which to reduce learning rate when plateau is reached.
scheduler_patience (int, default=5) – Number of epochs with no improvement after which learning rate will be reduced.
verbose (str, default="none") – Whether to display progress info. Options are: “none”, “progress_bar”, “print_statement”. If any other, “none” is automatically chosen.
device (torch.device or str, optional) – Device to run the model on. If None, will auto-detect GPU or use CPU.
model_name (str, default="SGIRMolecularPredictor") – Name identifier for the model.

Example¶

examples/prediction/sgir.py¶

import numpy as np
from torch_molecule import SGIRMolecularPredictor
from torch_molecule.utils.search import ParameterType, ParameterSpec
import os

def test_sgir_predictor():
    # Test data
    smiles_list = [
        'CNC[C@H]1OCc2cnnn2CCCC(=O)N([C@H](C)CO)C[C@@H]1C',
        'CNC[C@@H]1OCc2cnnn2CCCC(=O)N([C@H](C)CO)C[C@H]1C',
        'C[C@H]1CN([C@@H](C)CO)C(=O)CCCn2cc(nn2)CO[C@@H]1CN(C)C(=O)CCC(F)(F)F',
        'CC1=CC=C(C=C1)C2=CC(=NN2C3=CC=C(C=C3)S(=O)(=O)N)C(F)(F)F'
    ]
    properties = np.array([1.2, 2.3, 3.4, 4.5])  # Regression values
    unlabeled_smiles = [
        'CC1=CC=C(C=C1)C2=CC(=NN2C3=CC=C(C=C3)S(=O)(=O)N)C(F)(F)F',
        'CNC[C@H]1OCc2cnnn2CCCC(=O)N([C@H](C)CO)C[C@@H]1C'
    ]

    # Initialize model
    print("\n=== Testing SGIR model initialization ===")
    model = SGIRMolecularPredictor(
        num_task=1,
        task_type="regression",
        gamma=0.8,
        num_layer=3,
        hidden_size=128,
        batch_size=4,
        epochs=20,
        verbose="progress_bar",
        warmup_epoch=5,
        labeling_interval=1,
        augmentation_interval=1
    )

    # Test fitting with unlabeled data
    print("\n=== Testing SGIR model fitting with unlabeled data ===")
    model.fit(smiles_list[:3], properties[:3], X_unlbl=unlabeled_smiles)

    # Test prediction
    print("\n=== Testing SGIR model prediction ===")
    predictions = model.predict(smiles_list[3:])
    print(f"Prediction shape: {predictions['prediction'].shape}")
    print(f"Prediction value: {predictions['prediction']}")

    # Test auto-fitting
    search_parameters = {
        'num_layer': ParameterSpec(
            param_type=ParameterType.INTEGER,
            value_range=(2, 4)
        ),
        'hidden_size': ParameterSpec(
            param_type=ParameterType.INTEGER,
            value_range=(64, 256)
        ),
        'learning_rate': ParameterSpec(
            param_type=ParameterType.LOG_FLOAT,
            value_range=(1e-4, 1e-2)
        )
    }
    
    model_auto = SGIRMolecularPredictor(
        num_task=1,
        task_type="regression",
        epochs=3,
        verbose="progress_bar"
    )
    
    model_auto.autofit(
        smiles_list,
        properties,
        X_unlbl=unlabeled_smiles,
        search_parameters=search_parameters,
        n_trials=2
    )

    # Test save/load
    save_path = "test_sgir_model.pt"
    model.save(save_path)
    
    new_model = SGIRMolecularPredictor(
        num_task=1,
        task_type="regression"
    )
    new_model.load(save_path)

    # # Test invalid SMILES
    # try:
    #     invalid_smiles = ['CC', 'INVALID_SMILES']
    #     predictions = model.predict(invalid_smiles)
    # except Exception as e:
    #     print(f"Invalid SMILES handled with error: {str(e)}")

    # Clean up
    if os.path.exists(save_path):
        os.remove(save_path)

if __name__ == "__main__":
    test_sgir_predictor()

Methods¶

`fit`	Fit the model to training data with optional validation set.
`autofit`	Automatically find the best hyperparameters using Optuna optimization.
`predict`	Make predictions using the fitted model.

For shared checkpoint and parameter operations, see Base Modules.

fit¶

SGIRMolecularPredictor.fit(X_train, y_train, X_val=None, y_val=None, X_unlbl=None)[source]¶

Fit the model to training data with optional validation set.

Parameters:

X_train (List[str])
y_train (List | ndarray | None)
X_val (List[str] | None)
y_val (List | ndarray | None)
X_unlbl (List[str] | None)

Return type:

SGIRMolecularPredictor

autofit¶

SGIRMolecularPredictor.autofit(X_train, y_train, X_val=None, y_val=None, X_unlbl=None, search_parameters=None, n_trials=10)¶

Automatically find the best hyperparameters using Optuna optimization.

Parameters:

X_train (List[str])
y_train (List | ndarray | None)
X_val (List[str] | None)
y_val (List | ndarray | None)
X_unlbl (List[str] | None)
search_parameters (Dict[str, ParameterSpec] | None)
n_trials (int)

Return type:

GNNMolecularPredictor

predict¶

SGIRMolecularPredictor.predict(X)¶

Make predictions using the fitted model.

Parameters:

X (List[str]) – List of SMILES strings to make predictions for

Returns:

Dictionary containing:

’prediction’: Model predictions (shape: [n_samples, n_tasks])
’variance’: Prediction variances (shape: [n_samples, n_tasks])
’node_importance’: A nested list where the outer list has length n_samples and each inner list has length n_nodes for that molecule

Return type:

Dict[str, np.ndarray]