ium_478841/scripts/mlflow_train.py

from urllib.parse import urlparse
import mlflow
import mlflow.pytorch as model_logger

import argparse
import pandas as pd
import numpy as np
from sklearn.metrics import mean_squared_error, mean_absolute_error

import torch
from torch import nn
from torch.utils import data as t_u_data


# mlflow.set_tracking_uri("http://localhost:5000")
mlflow.set_tracking_uri("http://172.17.0.1:5000")
mlflow.set_experiment("s478841")


# * Customized Dataset class (base provided by PyTorch)
class AvocadoDataset(t_u_data.Dataset):
    def __init__(self, path: str, target: str = 'AveragePrice'):
        data = pd.read_csv(path)
        y = data[target].values.astype('float32')
        self.y = y.reshape((len(y), 1))
        self.x_data = data.drop(
            [target], axis=1).values.astype('float32')
        self.x_shape = data.drop([target], axis=1).shape
        # print("Data shape is: ", self.x_data.shape)

    def __len__(self):
        return len(self.x_data)

    def __getitem__(self, idx):
        return [self.x_data[idx], self.y[idx]]

    def get_shape(self):
        return self.x_shape

    def get_splits(self, n_test=0.33):
        test_size = round(n_test * len(self.x_data))
        train_size = len(self.x_data) - test_size
        return t_u_data.random_split(self, [train_size, test_size])


class AvocadoRegressor(nn.Module):
    def __init__(self, input_dim):
        super(AvocadoRegressor, self).__init__()
        self.hidden1 = nn.Linear(input_dim, 32)
        nn.init.xavier_uniform_(self.hidden1.weight)
        self.act1 = nn.ReLU()
        self.hidden2 = nn.Linear(32, 8)
        nn.init.xavier_uniform_(self.hidden2.weight)
        self.act2 = nn.ReLU()
        self.hidden3 = nn.Linear(8, 1)
        nn.init.xavier_uniform_(self.hidden3.weight)

    def forward(self, x):
        x = self.hidden1(x)
        x = self.act1(x)
        x = self.hidden2(x)
        x = self.act2(x)
        x = self.hidden3(x)
        return x


def prepare_data(paths):
    train_dl = t_u_data.DataLoader(AvocadoDataset(
        paths[0]), batch_size=32, shuffle=True)
    validate_dl = t_u_data.DataLoader(AvocadoDataset(
        paths[1]), batch_size=128, shuffle=True)
    test_dl = t_u_data.DataLoader(AvocadoDataset(
        paths[2]), batch_size=1, shuffle=False)
    return train_dl, validate_dl, test_dl


def train_model(train_dl, model, epochs, log_step):
    criterion = nn.MSELoss()
    optimizer = torch.optim.SGD(model.parameters(), lr=0.01, momentum=0.9)
    to_compare = None
    metrics = None
    for epoch in range(1, epochs+1):
        for _, (inputs, targets) in enumerate(train_dl):
            optimizer.zero_grad()
            yhat = model(inputs)
            # * For loss value inspection
            to_compare = (yhat, targets)
            loss = criterion(yhat, targets)
            loss.backward()
            optimizer.step()
        if epoch == 1 or (epoch) % log_step == 0:
            result, target = to_compare[0].detach(
            ).numpy(), to_compare[1].detach().numpy()
            metrics = {'train.mse': mean_squared_error(target, result),
                       'train.mae': mean_absolute_error(target, result),
                       'train.rmse': mean_squared_error(target, result, squared=False)}
            # _run.log_scalar("training.RMSE", np.sqrt(mse), epoch)
            # _run.log_scalar("training.MAE", mae, epoch)
            # _run.log_scalar('training.MSE', mse, epoch)
            print(
                f"Epoch {epoch}\t→\tMSE: {metrics['train.mse']},\tRMSE: {metrics['train.rmse']},\tMAE: {metrics['train.mae']}")
    return metrics


def evaluate_model(test_dl, model):
    predictions, actuals = list(), list()
    for _, (inputs, targets) in enumerate(test_dl):
        yhat = model(inputs)
        # * retrieve numpy array
        yhat = yhat.detach().numpy()
        actual = targets.numpy()
        actual = actual.reshape((len(actual), 1))
        # * store predictions
        predictions.append(yhat)
        actuals.append(actual)
    predictions, actuals = np.vstack(predictions), np.vstack(actuals)
    # * return MSE value
    mse = mean_squared_error(actuals, predictions)
    rmse = mean_squared_error(actuals, predictions, squared=False)
    mae = mean_absolute_error(actuals, predictions)
    return mse, rmse, mae


def predict(row, model):
    row = row[0].flatten()
    yhat = model(row)
    yhat = yhat.detach().numpy()
    return yhat


def main(epochs, save_model, log_step):
    print(
        f"Your model will be trained for {epochs} epochs, logging every {log_step} steps. Trained model will {'not ' if save_model else ''}be saved.")

    # * Paths to data
    avocado_data = ['./data/avocado.data.train',
                    './data/avocado.data.valid',
                    './data/avocado.data.test']

    # * Data preparation
    train_dl, validate_dl, test_dl = prepare_data(paths=avocado_data)
    print(f"""
          Train set size: {len(train_dl.dataset)},
          Validate set size: {len(validate_dl.dataset)}
          Test set size: {len(test_dl.dataset)}
          """)

    # * Model definition
    # ! 66 - in case only regions and type are used (among all the categorical vals)
    model = AvocadoRegressor(235)

    # * Train model
    print("Let's start the training, mate!")
    with mlflow.start_run() as run:
        print("MLflow run experiment_id: {0}".format(run.info.experiment_id))
        print("MLflow run artifact_uri: {0}".format(run.info.artifact_uri))
        metrics = train_model(train_dl=train_dl, model=model,
                              epochs=epochs, log_step=log_step)
        mlflow.log_param('epochs', epochs)
        mlflow.log_metrics(metrics)

        # * Evaluate model
        val_metrics = {key: val for key, val in zip(
            ['validate.mse', 'validate.rmse', 'validate.mae'], evaluate_model(validate_dl, model))}
        print(
            f"\nEvaluation on VALIDATION set\t→\tMSE: {val_metrics['validate.mse']}, RMSE: {val_metrics['validate.rmse']}, MAE: {val_metrics['validate.mae']}")
        mlflow.log_metrics(val_metrics)

        test_loss = {key: val for key, val in zip(
            ['test.mse', 'test.rmse', 'test.mae'], evaluate_model(test_dl, model))}
        print(
            f"\nEvaluation on TEST set\t→\tMSE: {test_loss['test.mse']}, RMSE: {test_loss['test.rmse']}, MAE: {test_loss['test.mae']}")
        mlflow.log_metrics(test_loss)

        # tracking_url_type_store = urlparse(mlflow.get_tracking_uri()).scheme

        # if tracking_url_type_store != 'file':
        #     print('First option')
        #     model_logger.log_model(
        #         model, "avocados-model", registered_model_name="AvocadoModel_478841")
        # else:
        #     print('Second option')
        #     model_logger.log_model(model, "model")


    # * Save the trained model
    if save_model:
        print("Your model has been saved - have a nice day!")
        scripted_model = torch.jit.script(model)
        scripted_model.save('./data/model_scripted.pt')
        # ex.add_artifact('./data/model_scripted.pt')


# ex.run()
if __name__ == '__main__':
    # * Model parameters
    parser = argparse.ArgumentParser(description="Script performing logistic regression model training",
                                     formatter_class=argparse.ArgumentDefaultsHelpFormatter)
    parser.add_argument(
        "-e", "--epochs", default=100, help="Number of epochs the model will be trained for")
    parser.add_argument(
        "-s", "--step", default=10, help="Number of steps to repeat logging loss values on")
    parser.add_argument("--save", action="store_true",
                        help="Save trained model to file 'trained_model.h5'")

    args = vars(parser.parse_args())

    epochs = int(args['epochs'])
    save_model = args['save']
    log_step = int(args['step'])

    main(epochs, save_model, log_step)