ium_444507/lab08_deepLearining_mlflow.py

#!/usr/bin/python

from urllib.parse import urlparse
import numpy as np
import torch
from torch import nn
from torch.autograd import Variable
from sklearn.metrics import accuracy_score, f1_score
import torch.nn.functional as F
import pandas as pd
from sklearn import preprocessing
import sys
import logging
import mlflow
import mlflow.pytorch

logging.basicConfig(level=logging.WARN)
logger = logging.getLogger(__name__)

# mlflow.set_tracking_uri("http://172.17.0.1:5000")
mlflow.set_experiment("s444507")


class Model(nn.Module):
    def __init__(self, input_dim):
        super(Model, self).__init__()
        self.layer1 = nn.Linear(input_dim, 100)
        self.layer2 = nn.Linear(100, 60)
        self.layer3 = nn.Linear(60, 5)

    def forward(self, x):
        x = F.relu(self.layer1(x.float()))
        x = F.relu(self.layer2(x.float()))
        x = F.softmax(self.layer3(x.float()))  # To check with the loss function
        return x


def load_dataset_raw():
    """ Load data from .csv file. """
    cars = pd.read_csv('./Car_Prices_Poland_Kaggle.csv', usecols=[1, 4, 5, 6, 10], sep=',')
    return cars


def load_dataset_files():
    """ Load shuffled, splitted dev and train files from .csv files. """

    cars_dev = pd.read_csv('./Car_Prices_Poland_Kaggle_dev.csv', usecols=[1, 4, 5, 6, 10], sep=',', names= [str(i) for i in range(5)])
    cars_train = pd.read_csv('./Car_Prices_Poland_Kaggle_train.csv', usecols=[1, 4, 5, 6, 10], sep=',', names= [str(i) for i in range(5)])

    return cars_dev, cars_train


def remove_rows(data_dev, data_train):
    dev_removed_rows = data_dev.loc[(data_dev['0'] == 'audi') | (data_dev['0'] == 'bmw') | (data_dev['0'] == 'ford') | (data_dev['0'] == 'opel') | (data_dev['0'] == 'volkswagen')]
    train_removed_rows = data_train.loc[(data_train['0'] == 'audi') | (data_train['0'] == 'bmw') | (data_train['0'] == 'ford') | (data_train['0'] == 'opel') | (data_train['0'] == 'volkswagen')]

    return dev_removed_rows, train_removed_rows


def prepare_labels_features(dataset):
    """ Label make column"""
    le = preprocessing.LabelEncoder()
    mark_column = np.array(dataset[:]['0'])
    le.fit(mark_column)

    print(list(le.classes_))
    lab = le.transform(mark_column)
    feat = dataset.drop(['0'], axis=1).to_numpy()

    mm_scaler = preprocessing.MinMaxScaler()
    feat = mm_scaler.fit_transform(feat)

    return lab, feat


def my_main(epoch):
    print("Loading dataset...")
    dev, train = load_dataset_files()
    print("Dataset loaded")

    print("Preparing dataset...")
    dev, train = remove_rows(dev, train)
    labels_train, features_train = prepare_labels_features(train)
    labels_test, features_test = prepare_labels_features(dev)
    print("Dataset prepared")

    # Training
    model = Model(features_train.shape[1])
    optimizer = torch.optim.Adam(model.parameters(), lr=0.01)
    loss_fn = nn.CrossEntropyLoss()

    # number of epochs is parametrized
    try:
        epochs_n = int(epoch)
    except Exception as e:
        print(e)
        print("Setting default epochs value to 10.")
        epochs_n = 10

    print(f"Number of epochs: {epochs_n}")
    mlflow.log_param("epochs", epochs_n)

    print("Starting model training...")
    x_train, y_train = Variable(torch.from_numpy(features_train)).float(), Variable(torch.from_numpy(labels_train)).long()
    for epoch in range(1, epochs_n + 1):
        print("Epoch #", epoch)
        y_pred = model(x_train)
        loss = loss_fn(y_pred, y_train)
        print(f"The loss calculated: {loss}")

        # Zero gradients
        optimizer.zero_grad()
        loss.backward()  # Gradients
        optimizer.step()  # Update
    print("Model training finished")

    x_test = Variable(torch.from_numpy(features_test)).float()
    pred = model(x_test)
    pred = pred.detach().numpy()
    print(f"The accuracy metric is: {accuracy_score(labels_test, np.argmax(pred, axis=1))}")

    accuracy = accuracy_score(labels_test, np.argmax(pred, axis=1))
    f1 = f1_score(labels_test, np.argmax(pred, axis=1), average='weighted')

    mlflow.log_metric("accuracy", accuracy)
    mlflow.log_metric("f1", f1)

    # Infer model signature to log it
    signature = mlflow.models.signature.infer_signature(features_train, labels_train)
    tracking_url_type_store = urlparse(mlflow.get_tracking_uri()).scheme

    if tracking_url_type_store != "file":
        mlflow.pytorch.log_model(model, "model", registered_model_name="s444507", signature=signature,
                                 input_example=features_train)
    else:
        mlflow.pytorch.log_model(model, "model", signature=signature, input_example=features_train)
        mlflow.pytorch.save_model(model, "my_model", signature=signature, input_example=features_train)

    print("Saving model to file...")
    torch.save(model, "CarPrices_pytorch_model.pkl")
    print("Model saved with name: CarPrices_pytorch_model.pkl")

    saved_model = torch.load("CarPrices_pytorch_model.pkl")
    print(np.argmax(saved_model(x_test[0]).detach().numpy(), axis=0))

    pd_predictions = pd.DataFrame(pred)
    pd_predictions.to_csv("./prediction_results.csv")


try:
    epochs = int(sys.argv[1])
except Exception as e:
    print(e)
    print("Setting default epochs value to 1000.")
    epochs = 100

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))
    my_main(epochs)