polymerlearn/utils/train_graphs.py

import os
import torch
import numpy as np
import pandas as pd
from typing import List

from torch_geometric.data import Batch, Data
from sklearn.metrics import r2_score, mean_squared_error, mean_absolute_error

from polymerlearn.utils import GraphDataset

def get_vector(
        data: pd.DataFrame, 
        prop: str = 'Mw (PS)', 
        fill_value: float = None,
        use_log: bool = True):
    '''
    Get vector to be added as sample-wide feature in model

    Args:
        data (pd.DataFrame): Base dataframe from which to extract the data.
        prop (str, optional): Name of column (property) for which to get the
            vector. (:default: :obj:`Mw (PS)`)
        fill_value (float, optional): Value with which to fill missing values 
            in the column. If `None`, will fill missing values with median from
            the column. (:default: :obj:`None`) 
        use_log (bool): Log transforms the values if true
        standard_scale (bool): 

    :rtype: pd.Series
    '''
    if fill_value is None:
        to_fill = sorted(data[prop].loc[data[prop].notna()])[int(sum(data[prop].notna())/2)]
    else:
        to_fill = fill_value
    
    if prop != '%TMP' and use_log:
        vec = np.log(data[prop].fillna(to_fill))
    else:
        # Don't log transform TMP
        vec = data[prop].fillna(to_fill)

    return vec

def get_IV_add(data):
    '''
    Return the standard IV additional data (i.e. resin properties) used
        in the paper.
    Args:
        data (pd.Dataframe)

    No arguments
    '''

    mw_vector = get_vector(data, prop = 'Mw (PS)').to_numpy()
    an_vector = get_vector(data, prop = 'AN').to_numpy()
    ohn_vector = get_vector(data, prop = 'OHN').to_numpy()
    tmp_vector = get_vector(data, prop = '%TMP', fill_value=0).to_numpy()

    add = np.stack([mw_vector, an_vector, ohn_vector, tmp_vector]).T

    return add

def get_IV_add_nolog(data):
    mw_vector = get_vector(data, prop = 'Mw (PS)', use_log = False).to_numpy()
    an_vector = get_vector(data, prop = 'AN', use_log = False).to_numpy()
    ohn_vector = get_vector(data, prop = 'OHN', use_log = False).to_numpy()
    tmp_vector = get_vector(data, prop = '%TMP', fill_value=0).to_numpy()

    add = np.stack([mw_vector, an_vector, ohn_vector, tmp_vector]).T
    return add

def get_Tg_add(data):
    mw_vector = get_vector(data, prop = 'Mw (PS)', use_log = True).to_numpy()

    add = np.stack([mw_vector]).T

    return add

def get_Tg_add_nolog(data):
    mw_vector = get_vector(data, prop = 'Mw (PS)', use_log = False).to_numpy()

    add = np.stack([mw_vector]).T

    return add

def get_add_properties(data: pd.DataFrame, prop_names: List[str], use_log: List[bool] = None):
    '''
    Gets properties to add to the model given data, names of properties, 
        and whether to log transform them.
    '''

    add_vectors = []

    if use_log is None:
        use_log = [True] * len(prop_names)

    for p,l in zip(prop_names, use_log):
        fill_value = 0 if p=='%TMP' else None

        add_vectors.append(
            get_vector(data, prop=p, use_log=l, fill_value=fill_value)
        )

    return np.stack(add_vectors).T

def make_like_batch(batch: tuple):
    '''
    Decomposes a batch of acid/glycol into tensors to be fed into model

    Args: 
        batch (tuple): Must be of length 2 and contain (Acid_data, Glycol_data).

    :type: tuple[`torch.geometric.data.Batch`, `torch.geometric.data.Batch`]
    '''
    Adata, Gdata = batch

    Abatch = Batch().from_data_list(Adata)
    Gbatch = Batch().from_data_list(Gdata)

    return Abatch, Gbatch 

def check_early_stop(loss_list, delay = 100):
    '''
    Checks early stopping criterion for training procedure
    Check max, see if <delay> epochs have passed since the max
    '''

    largest = np.argmin(loss_list)

    # Can enforce some smoothness condition:
    low = max(largest - 5, 0)
    up = largest + 6

    # Check if the difference between average around it and itself is different enough
    minloss = loss_list[largest]
    around_min = np.concatenate([loss_list[low:largest], loss_list[(largest+1):up]])
    smooth = np.abs(np.mean(around_min) - minloss) < np.abs(minloss * 0.25)

    return ((len(loss_list) - largest) > delay) and smooth

def train(
        model: torch.nn.Module, 
        optimizer, 
        criterion, 
        dataset: GraphDataset, 
        batch_size: int = 64, 
        epochs: int = 1000
    ):
    '''
    Args:
        model: Neural network to train
        optimizer: Optimizer to use when training the model.
        criterion: Loss function.
        dataset: Dataset class.
        batch_size: Number of samples on which to optimize at each iteration. See
            the description in CV_Eval
        epochs: Number of iterations to train on the data.
    '''

    for e in range(epochs):
        
        # Batch:
        batch, Y, add_features = dataset.get_train_batch(size = batch_size)
        test_batch, Ytest, add_test = dataset.get_test()

        train_predictions = []
        cum_loss = 0

        for i in range(batch_size):

            # Predictions:
            #predictions = torch.tensor([model(*make_like_batch(batch[i])) for i in range(batch_size)], requires_grad = True).float()
            af = None if add_features is None else torch.tensor(add_features[i]).float()
            train_prediction = model(*make_like_batch(batch[i]), af)
            train_predictions.append(train_prediction.clone().detach().item())
            #print(predictions)

            # Compute and backprop loss
            loss = criterion(train_prediction, torch.tensor([Y[i]]))
            optimizer.zero_grad()
            loss.backward()
            cum_loss += loss.item()
            optimizer.step()

        # Test:
        test_preds = []
        with torch.no_grad():
            for i in range(Ytest.shape[0]):
                at = None if add_test is None else add_test[i].clone().detach()
                test_preds.append(model(*make_like_batch(test_batch[i]), at).clone().detach().item())

        r2_test = r2_score(Ytest.numpy(), test_preds)
        mse_test = mean_squared_error(Ytest.numpy(), test_preds)

        if e % 10 == 0:
            print(f'Epoch: {e}, \t Train r2: {r2_score(Y, train_predictions):.4f} \t Train Loss: {cum_loss:.4f} \t Test r2: {r2_test:.4f} \t Test Loss {mse_test:.4f}')


def CV_eval(
        dataset: GraphDataset,
        model_generator: torch.nn.Module,
        optimizer_generator,
        criterion,
        model_generator_kwargs: dict = {},
        optimizer_kwargs: dict = {},
        batch_size = 64,
        verbose = 1,
        epochs = 1000,
        use_val = False,
        val_size = 0.1,
        stop_option = 0,
        early_stop_delay = 100,
        save_state_dicts = False,
        get_scores = False,
        device = None):
    '''
    Args:
        dataset (GraphDataset): Preprocessed dataset matching the GraphDataset
            API.
        model_generator (torch.nn.Module): Class of the neural network/model that
            can be instantiated multiple times within the function.
        optimizer_generator: Optimizer that can be instantiated multiple times within
            the function.
        criterion: Loss function that can be instantiated multiple times within
            the function.
        model_generator_kwargs (dict): Dictionary of keyword arguments to be passed
            to the model for every instantiation.
        optimizer_kwargs (dict): Dictionary of keyword arguments to be passed
            to the optimizer for every instantiation.
        batch_size (int): Number of samples to be optimized on for each step. Note
            this works differently than batch size in stochastic gradient descent.
            Here, the higher value for the argument denotes more samples to be
            trained on per epoch (usually vice versa is standard).
        verbose (int): Level at which to print. Should be 0 or 1.
        epochs (int): Number of training iterations on the dataset.
        use_val (bool): If true, uses the validation set in the Dataset class.
        val_sise (float): Size of the validation set to use
        stop_option (int): Option that specifies which method to use for early
            stopping/validation saving. 0 simply performs all epochs for each fold.
            1 performs all epochs but uses model with highest validation score for 
            evaluation on test set. 2 stops early if the validation loss was at least
            `early_stop_delay` epochs ago; it loads that trial's model and evaluates
            on it.
        early_stop_delay (int): Number of epochs to wait after an early stopping condition
            is met.
        save_state_dicts (bool): If True, returns state dictionaries for the model at
            each fold. Useful for explainability.
        get_scores (bool, optional): If True, return only the average values of metrics 
            across the folds
        device (str): Device name at which to run torch calculations on. Supports GPU.
    '''

    num_folds = 2   # LB changed from 5 to 2
    fold_count = 0

    r2_test_per_fold = []
    mse_test_per_fold = []
    mae_test_per_fold = []

    all_predictions = []
    all_y = []
    all_reference_inds = []

    model_state_dicts = []

    for test_batch, Ytest, add_test, test_inds in \
            dataset.Kfold_CV(folds = num_folds, val = True, val_size = val_size):

        # Instantiate fold-level model and optimizer:
        model = model_generator(**model_generator_kwargs).to(device)
            # Move model to GPU before setting optimizer
        optimizer = optimizer_generator(model.parameters(), **optimizer_kwargs)

        fold_count += 1
        loss_list = []

        if stop_option >= 1:
            min_val_loss = 1e10
            min_val_state_dict = None

        for e in range(epochs):
            
            # Bootstrap batches:
            batch, Y, add_features = dataset.get_train_batch(size = batch_size)

            train_predictions = []
            cum_loss = 0

            for i in range(batch_size):

                # Predictions:
                af = None if add_features is None else torch.tensor(add_features[i]).float()
                if verbose > 1:
                    print('Additional it={}'.format(i), af)
                train_prediction = model(*make_like_batch(batch[i]), af)
                if verbose > 1:
                    print('pred', train_prediction.item())
                train_predictions.append(train_prediction.clone().detach().item())

                # Compute and backprop loss
                loss = criterion(train_prediction, torch.tensor([Y[i]]))
                optimizer.zero_grad()
                loss.backward()
                cum_loss += loss.item()
                optimizer.step()

            if verbose > 1:
                print('Train predictions', train_predictions)

            # Test on validation:
            if use_val:
                model.eval()
                val_batch, Yval, add_feat_val = dataset.get_validation()
                cum_val_loss = 0
                val_preds = []
                with torch.no_grad():
                    for i in range(Yval.shape[0]):
                        pred = model(*make_like_batch(val_batch[i]), add_feat_val[i])
                        val_preds.append(pred.item())
                        cum_val_loss += criterion(pred, Yval[i]).item()
                    
                loss_list.append(cum_val_loss)
                model.train() # Must switch back to train after eval

            if e % 50 == 0 and (verbose >= 1):
                print_str = f'Fold: {fold_count} \t Epoch: {e}, \
                    \t Train r2: {r2_score(Y, train_predictions):.4f} \t Train Loss: {cum_loss:.4f}' 
                if use_val:
                   print_str += f'Val r2: {r2_score(Yval, val_preds):.4f} \t Val Loss: {cum_val_loss:.4f}'
                print(print_str)

            if stop_option >= 1:
                if cum_val_loss < min_val_loss:
                    # If min val loss, store state dict
                    min_val_loss = cum_val_loss
                    min_val_state_dict = model.state_dict()

            # Check early stop if needed:
            if stop_option == 2:
                # Check criteria:
                if check_early_stop(loss_list, early_stop_delay) and e > early_stop_delay:
                    break

        
        if stop_option >= 1: # Loads the min val loss state dict even if we didn't break
            # Load in the model with min val loss
            model = model_generator(**model_generator_kwargs)
            model.load_state_dict(min_val_state_dict)

        # Test:
        test_preds = []
        with torch.no_grad():
            for i in range(Ytest.shape[0]):
                at = None if add_test is None else torch.tensor(add_test[i]).float()
                pred = model(*make_like_batch(test_batch[i]), at).clone().detach().item()
                test_preds.append(pred)
                all_predictions.append(pred)
                all_y.append(Ytest[i].item())
                all_reference_inds.append(test_inds[i])

        r2_test = r2_score(Ytest.numpy(), test_preds)
        mse_test = mean_squared_error(Ytest.numpy(), test_preds)
        mae_test = mean_absolute_error(Ytest.numpy(), test_preds)

        print(f'Fold: {fold_count} \t Test r2: {r2_test:.4f} \t Test Loss: {mse_test:.4f} \t Test MAE: {mae_test:.4f}')

        r2_test_per_fold.append(r2_test)
        mse_test_per_fold.append(mse_test)
        mae_test_per_fold.append(mae_test)

        if save_state_dicts:
            model_state_dicts.append(model.state_dict())


    print('Final avg. r2: ', np.mean(r2_test_per_fold))
    print('Final avg. MSE:', np.mean(mse_test_per_fold))
    print('Final avg. MAE:', np.mean(mae_test_per_fold))

    r2_avg = np.mean(r2_test_per_fold)
    mae_avg = np.mean(mae_test_per_fold)

    big_ret_dict = {
        'r2': r2_avg,
        'mae': mae_avg,
        'all_predictions': all_predictions,
        'all_y': all_y,
        'all_reference_inds': all_reference_inds,
        'model_state_dicts': model_state_dicts
    }

    if save_state_dicts:
        if get_scores:
            return big_ret_dict
        else:
            print("Returning model_state_dicts")
            return all_predictions, all_y, all_reference_inds, model_state_dicts

    if get_scores: # Return scores:
        return big_ret_dict


    print("Not state dict")
    return all_predictions, all_y, all_reference_inds

def train_joint(
        model, 
        optimizer, 
        criterion, 
        dataset, 
        batch_size = 64, 
        epochs = 100,
        gamma = 1e4
    ):

    for e in range(epochs):
        
        # Batch:
        batch, Y, add_features = dataset.get_train_batch(size = batch_size)
        test_batch, Ytest, add_test = dataset.get_test()

        #Y = np.log(Y)
        #Ytest = np.log(Ytest)
        # Y[:, 0] = np.log(Y[:, 0])
        # Ytest[:, 0] = np.log(Ytest[:, 0])

        train_predictions = []
        cum_loss = 0

        model.train()

        for i in range(batch_size):

            # Predictions:
            #predictions = torch.tensor([model(*make_like_batch(batch[i])) for i in range(batch_size)], requires_grad = True).float()
            af = None if add_features is None else torch.tensor(add_features[i]).float()
            train_prediction = model(*make_like_batch(batch[i]), af)
            #train_predictions.append(train_prediction.clone().detach().item())
            train_predictions.append([train_prediction[i].clone().detach().item() for i in ['IV', 'Tg']])
            #print(predictions)

            # Compute and backprop loss
            #loss = criterion(train_prediction, torch.tensor([Y[i]]))
            loss_IV = criterion(train_prediction['IV'], torch.tensor([Y[i][0]]))
            loss_Tg = criterion(train_prediction['Tg'], torch.tensor([Y[i][1]]))
            loss = gamma * loss_IV + loss_Tg
            optimizer.zero_grad()
            loss.backward()
            cum_loss += loss.item()
            optimizer.step()

        # Test:
        # model.eval()
        # test_predIV = []
        # test_predTg = []
        # with torch.no_grad():
        #     for i in range(Ytest.shape[0]):
        #         pred = model(*make_like_batch(test_batch[i]), torch.tensor(add_test[i]).float())
        #         test_predIV.append(pred['IV'].clone().detach().item())
        #         test_predTg.append(pred['Tg'].clone().detach().item())

        # r2_testIV = r2_score(Ytest[:,0].numpy(), test_predIV)
        # r2_testTg = r2_score(Ytest[:,1].numpy(), test_predTg)

        if e % 10 == 0:
            print(f'Epoch: {e}, \t Train r2: {r2_score(Y, train_predictions):.4f} \t Train Loss: {cum_loss:.4f}') #\t Test r2: {r2_testIV:.4f} \t Test r2 (Tg): {r2_testTg}')

def CV_eval_joint(
        dataset,
        model_generator: torch.nn.Module,
        optimizer_generator,
        criterion,
        model_generator_kwargs: dict = {},
        optimizer_kwargs: dict = {},
        batch_size = 64,
        verbose = 1,
        gamma = 1e4,
        epochs = 1000,
        get_scores = False,
        device = None,
        save_state_dicts = False,
        check_r2_thresh = True):
    '''
    Cross validation of the joint Tg/IV model

    Args:
        gamma (float): Weighting factor applied to IV loss. Used
            to balance the losses between IV and Tg during the joint
            training process.
    '''

    num_folds = 5
    fold_count = 0

    r2_test_per_fold = []
    r2_test_per_fold_IV = []
    r2_test_per_fold_Tg = []

    mse_test_per_fold = []
    mse_test_per_fold_IV = []
    mse_test_per_fold_Tg = []

    mae_test_per_fold = []
    mae_test_per_fold_IV = []
    mae_test_per_fold_Tg = []

    all_predictions = []
    all_y = []
    all_reference_inds = []

    model_state_dicts = []

    for test_batch, Ytest, add_test, test_inds in dataset.Kfold_CV(folds = num_folds):

        model = model_generator(**model_generator_kwargs).to(device)
        optimizer = optimizer_generator(model.parameters(), **optimizer_kwargs)

        fold_count += 1
        #Ytest = np.log(Ytest) # Log transform Ytest
        #Ytest[:, 0] = np.log(Ytest[:, 0])

        model.train()

        #for e in range(epochs):
        e = 0
        while True:
            
            # Batch:
            batch, Y, add_features = dataset.get_train_batch(size = batch_size)
            if add_features is not None:
                add_features = torch.tensor(add_features).float().to(device)

            #Y = np.log(Y) # Log transform Y
            #[:, 0] = Y[:, 0])

            train_predictions = []
            cum_loss = 0

            for i in range(batch_size):

                # Predictions:
                #predictions = torch.tensor([model(*make_like_batch(batch[i])) for i in range(batch_size)], requires_grad = True).float()
                af = None if add_features is None else add_features[i]
                A, G = make_like_batch(batch[i])
                A, G = A.to(device), G.to(device)
                train_prediction = model(A, G, af)
                #train_prediction = model(*make_like_batch(batch[i]), af)
                train_predictions.append([train_prediction[i].clone().detach().item() for i in ['IV', 'Tg']])
                #print(predictions)

                # Compute and backprop joint loss
                loss_IV = criterion(train_prediction['IV'], torch.tensor([Y[i][0]]).to(device))
                loss_Tg = criterion(train_prediction['Tg'], torch.tensor([Y[i][1]]).to(device))
                loss = gamma * loss_IV + loss_Tg # Loss is additive between the two
                optimizer.zero_grad()
                loss.backward()
                cum_loss += loss.item()
                optimizer.step()
                
            try:
                r2IV = r2_score(Y[:][0], train_predictions[0][:])
            except:
                r2IV = -1
            try:
                r2Tg = r2_score(Y[:][1], train_predictions[1][:])
            except:
                r2Tg = -1

            if e % 50 == 0:
                #print(f'Fold: {fold_count} \t Epoch: {e}, \t Train r2: {r2_score(Y, train_predictions):.4f} \t Train Loss: {cum_loss:.4f}')
                print(f'Fold: {fold_count} : {e}, Train r2 IV, Tg: {r2IV:.4f}, {r2Tg:.4f} \t Train Loss: {cum_loss:.4f}')
                
            if check_r2_thresh and (e > epochs) and (r2IV > 0.9) and (r2Tg > 0.9):
                # Check for stable learning on both IV and Tg
                # Checks traning value, not validation
                break
            
            e += 1


        # Test:
        model.eval()
        test_preds = []
        with torch.no_grad():
            for i in range(Ytest.shape[0]):
                #test_preds.append(model(*make_like_batch(test_batch[i]), torch.tensor(add_test[i]).float()).clone().detach().item())
                at = None if add_test is None else torch.tensor(add_test[i]).float().to(device)
                A, G = make_like_batch(test_batch[i])
                A, G = A.to(device), G.to(device)
                test_pred = model(A, G, at)
                pred = [test_pred[i].clone().detach().item() for i in ['IV', 'Tg']]
                test_preds.append(pred)
                all_predictions.append(pred)
                all_y.append(Ytest[i,:].detach().clone().tolist())
                all_reference_inds.append(test_inds[i])

        r2_test = r2_score(Ytest.cpu().numpy(), test_preds)
        r2_test_IV = r2_score(Ytest.cpu().numpy()[:, 0], np.array(test_preds)[:, 0])
        r2_test_Tg = r2_score(Ytest.cpu().numpy()[:, 1], np.array(test_preds)[:, 1])
        mse_test = mean_squared_error(Ytest.cpu().numpy(), test_preds)
        mse_test_IV = mean_squared_error(Ytest.cpu().numpy()[:, 0], np.array(test_preds)[:, 0])
        mse_test_Tg = mean_squared_error(Ytest.cpu().numpy()[:, 1], np.array(test_preds)[:, 1])
        mae_test = mean_absolute_error(Ytest.cpu().numpy(), test_preds)
        mae_test_IV = mean_absolute_error(Ytest.cpu().numpy()[:, 0], np.array(test_preds)[:, 0])
        mae_test_Tg = mean_absolute_error(Ytest.cpu().numpy()[:, 1], np.array(test_preds)[:, 1])

        print(f'Fold: {fold_count} \t Test r2: {r2_test:.4f} \t r2_IV: {r2_test_IV:.4f} \t r2_Tg: {r2_test_Tg:.4f} \t MSE: {mse_test:.4f} \t MSE_IV: {mse_test_IV:.4f} \t MSE_Tg: {mse_test_Tg:.4f} \t MAE: {mae_test:.4f} \t MAE_IV: {mae_test_IV:.4f} \t MAE_Tg: {mae_test_Tg:.4f}')

        r2_test_per_fold.append(r2_test)
        r2_test_per_fold_IV.append(r2_test_IV)
        r2_test_per_fold_Tg.append(r2_test_Tg)

        mse_test_per_fold.append(mse_test)
        mse_test_per_fold_IV.append(mse_test_IV)
        mse_test_per_fold_Tg.append(mse_test_Tg)

        mae_test_per_fold.append(mae_test)
        mae_test_per_fold_IV.append(mae_test_IV)
        mae_test_per_fold_Tg.append(mae_test_Tg)

        if save_state_dicts:
            model_state_dicts.append(model.state_dict())

    print('Final avg. r2: ', np.mean(r2_test_per_fold))
    print('Final avg. r2 IV: ', np.mean(r2_test_per_fold_IV))
    print('Final avg. r2 Tg: ', np.mean(r2_test_per_fold_Tg))
    
    print('Final avg. MSE:', np.mean(mse_test_per_fold))
    print('Final avg. MSE IV: ', np.mean(mse_test_per_fold_IV))
    print('Final avg. MSE Tg: ', np.mean(mse_test_per_fold_Tg))

    print('Final avg. MAE:', np.mean(mae_test_per_fold))
    print('Final avg. MAE IV: ', np.mean(mae_test_per_fold_IV))
    print('Final avg. MAE Tg: ', np.mean(mae_test_per_fold_Tg))

    d = {
        'IV':(np.mean(r2_test_per_fold_IV), np.mean(mae_test_per_fold_IV)),
        'Tg':(np.mean(r2_test_per_fold_Tg), np.mean(mae_test_per_fold_Tg)),
        'all_predictions': all_predictions,
        'all_y': all_y,
        'all_reference_inds': all_reference_inds,
        'model_state_dicts': model_state_dicts
    }

    if save_state_dicts:
        if get_scores:
            return d
        else:
            return all_predictions, all_y, all_reference_inds, model_state_dicts

    if get_scores:
        # Return in a dictionary
        return d

    return all_predictions, all_y, all_reference_inds