inference_polymers_gnn.py

import os
import sys
from pathlib import Path

sys.path.insert(0, str(Path(__file__).resolve().parents[1]))
import torch
import pandas as pd
from polymerlearn.utils import GraphDataset, get_Tg_add, get_IV_add
from polymerlearn.models.gnn import PolymerGNN_Tg, PolymerGNN_IV
from polymerlearn.utils import make_like_batch
import numpy as np


# def convert_to_graphdataset(df, acid_positions = (0, 12), glycol_positions = (13, 25)):
#     """
#     Takes a dataframe with the input data and converts it to a graph dataset for the prediction

#     For the Tg the model takes the proportions of acid and glycols and the log of the Mw (PS) (no other parameters)
#     """
#     targets = ["Tg", "IV"]
#     # Need to pass the target column to do the preprocessing - not sure why it is needed - must be clarified
#     for col in targets:
#         df[col] = 0
#     df_duplicated = pd.concat([df, df]) # Because must specify a test set - this must be changed in the code
#     add_duplicated = get_Tg_add(df_duplicated)
#     print(f"Inference to be done on data of size {df.shape}")

#     graph_data = GraphDataset(
#         data = df_duplicated,
#         structure_dir = './Structures/AG/xyz',
#         Y_target=targets,
#         test_size = 0.5,
#         add_features = add_duplicated,
#         ac = acid_positions,
#         gc = glycol_positions
#     )

#     test_data, Ytest, add_test = graph_data.get_test()
#     print(f"After data preprocessing, inference on {len(test_data)}")
#     n_predictions = len(Ytest)
#     return test_data, add_test, n_predictions


def predict(df, model_path="../polymerlearn/data_models/", acid_positions = (0, 12), glycol_positions = (13, 25)):

    targets = ["Tg", "IV"]
    # Need to pass the target column to do the preprocessing - not sure why it is needed - must be clarified
    for col in targets:
        df[col] = 0
    df_duplicated = pd.concat([df, df]) # Because must specify a test set - this must be changed in the code
    
    add_features = {"Tg": get_Tg_add(df_duplicated), "IV": get_IV_add(df_duplicated)}

    print(f"Inference to be done on data of size {df.shape}")

    print("Tg", add_features["Tg"].shape)
    print("Tg", add_features["IV"].shape)

    pred_all = []

    for pred in targets:
        graph_data = GraphDataset(
            data = df_duplicated,
            structure_dir = './Structures/AG/xyz',
            Y_target=targets,
            test_size = 0.5,
            add_features = add_features[pred],
            ac = acid_positions,
            gc = glycol_positions
        )

        test_data, Ytest, add_test = graph_data.get_test()
        print(f"After data preprocessing, inference on {len(test_data)}")
        n_predictions = len(Ytest)

        if pred == "Tg":
            model = PolymerGNN_Tg(
                input_feat= 6,         # How many input features on each node; don't change this
                hidden_channels= 32,   # How many intermediate dimensions to use in model
                                        # Can change this ^^
                num_additional= add_features[pred].shape[1]     # How many additional resin properties to include in the prediction
                                        # Corresponds to the number in get_IV_add
            )
        else:
            model = PolymerGNN_IV(
                input_feat= 6,         # How many input features on each node; don't change this
                hidden_channels= 32,   # How many intermediate dimensions to use in model
                                        # Can change this ^^
                num_additional= add_features[pred].shape[1]     # How many additional resin properties to include in the prediction
                                        # Corresponds to the number in get_IV_add
            )

        device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
        state = torch.load(os.path.join(model_path, f"{pred}_model.pth"), map_location=device)
        model.load_state_dict(state)
        model.eval()
        predictions = []
        print(f"Prediction done on {n_predictions}")
        with torch.no_grad():
            for i in range(n_predictions):
                batch_like_tup = make_like_batch(test_data[i])
                pred = np.round(model(*batch_like_tup, add_test[i]).item(), 1)
                predictions.append(pred)
        pred_all.append(predictions)
    
    return pred_all


# def predict(df, acid_positions = (0, 12), glycol_positions = (13, 25), model_path="../polymerlearn/data_models/tg_model_test.pth"):
#     test_data, add_test, n_predictions = convert_to_graphdataset(df, acid_positions=acid_positions, glycol_positions=glycol_positions)

#     predictions = predict_from_graph(test_data, add_test, n_predictions, model_path=model_path)

#     return predictions