import tensorflow as tf
from keras import layers
from keras.models import save_model
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
import mlflow
import mlflow.keras
from urllib.parse import urlparse
import sys


def train():

    # Definicja wartości parametrów treningu
    epochs = int(sys.argv[1]) if len(sys.argv) > 1 else 100
    units = int(sys.argv[2]) if len(sys.argv) > 2 else 1
    learning_rate = float(sys.argv[3]) if len(sys.argv) > 3 else 0.1

    # Konfiguracja serwera i nazwy eksperymentu MLflow
    # mlflow.set_tracking_uri("http://172.17.0.1:5000")
    mlflow.set_experiment('s449288')

    # Podpięcie treningu do MLflow
    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))

        # Wczytanie danych
        data_train = pd.read_csv('lego_sets_clean_train.csv')
        data_test = pd.read_csv('lego_sets_clean_test.csv')

        # Wydzielenie zbiorów dla predykcji ceny zestawu na podstawie liczby klocków, którą zawiera
        train_piece_counts = np.array(data_train['piece_count'])
        train_prices = np.array(data_train['list_price'])
        test_piece_counts = np.array(data_test['piece_count'])
        test_prices = np.array(data_test['list_price'])

        # Normalizacja
        normalizer = layers.Normalization(input_shape=[1, ], axis=None)
        normalizer.adapt(train_piece_counts)

        # Inicjalizacja
        model = tf.keras.Sequential([
            normalizer,
            layers.Dense(units=units)
        ])

        # Kompilacja
        model.compile(
            optimizer=tf.optimizers.Adam(learning_rate=learning_rate),
            loss='mean_absolute_error'
        )

        # Trening
        history = model.fit(
            train_piece_counts,
            train_prices,
            epochs=epochs,
            verbose=0,
            validation_split=0.2
        )

        # Wykonanie predykcji na danych ze zbioru testującego
        y_pred = model.predict(test_piece_counts)

        # Zapis predykcji do pliku
        results = pd.DataFrame(
            {'test_set_piece_count': test_piece_counts.tolist(),
             'predicted_price': [round(a[0], 2) for a in y_pred.tolist()]})
        results.to_csv('lego_reg_results.csv', index=False, header=True)

        # Zapis modelu do pliku
        model.save('lego_reg_model')

        # Ewaluacja MAE na potrzeby MLflow (kopia z evaluate.py)
        mae = model.evaluate(
            test_piece_counts,
            test_prices, verbose=0)

        # Zapis parametrów i metryk dla MLflow
        mlflow.log_param('epochs', epochs)
        mlflow.log_param('units', units)
        mlflow.log_param('learning_rate', learning_rate)
        mlflow.log_metric("mae", mae)

        # Logowanie i zapis modelu dla Mlflow
        signature = mlflow.models.signature.infer_signature(train_piece_counts, model.predict(train_piece_counts))
        tracking_url_type_store = urlparse(mlflow.get_tracking_uri()).scheme
        if tracking_url_type_store != 'file':
            mlflow.keras.log_model(model, 'lego-model', registered_model_name='TFLegoModel',
                                   signature=signature)
        else:
            mlflow.keras.log_model(model, 'model', signature=signature, input_example=np.array(500))


if __name__ == '__main__':
    train()