How MLflow helps?
In this session, we will discuss about MLflow, a tool that is to be known in MLOps concept, before we deep dive into concepts
let us understand this how this tool can be helpful with an example.
let us say, you are running a Diary Farm and you manufacture different products like curd, cheese, milk and you have a Data scientist here who tries to figure out what Model is best for you using statistics.
let us say, he has data of last 3 year sales and now he wants to use an ML Model which is best suited to read data and predict sales for next year so that you can have exact idea of how many products to be manufactured
and to achieve this your Data scientist need to train different models using different statistics, also where he can have design different models, and also train same model with different data using versioning, compare results of training from each model and you can think of it more like a playground to train models and evaluate best model, this is where Mlflow comes into picture
Mlflow helps Data scientist track, manage and evaluate different ML models
we will understand the above scenario better with an example
MLflow Demo
Now before we start MLflow, install Jupyter notebook or VScode
also make sure python and pip is installed locally
Now to install mlflow
pip install mlflow
and now to access MLflow UI
!mlflow ui
and now by default MLflow will be running on port 5000
Now before we proceed on Model deployment, install scikit-learn module which helps in training our model
pip install scikit-learn
Now let us deploy our training model in MLflow
import mlflow
import mlflow.sklearn
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
from sklearn.model_selection import train_test_split
from sklearn.linear_model import LinearRegression
from sklearn.tree import DecisionTreeRegressor
from sklearn.ensemble import RandomForestRegressor
from sklearn.metrics import mean_squared_error, mean_absolute_error, r2_score
# ---------------------------
# 1. Generate Synthetic Sales Data
# ---------------------------
np.random.seed(42)
months = np.arange(1, 37) # 36 months (3 years)
products = ["Milk", "Curd", "Cheese"]
sales_data = {}
for product in products:
trend = months * np.random.randint(3, 8) # random growth trend
seasonality = 200 * np.sin(2 * np.pi * months / 12) # seasonal pattern
noise = np.random.normal(0, 50, 36) # noise
sales = 1000 + trend + seasonality + noise
sales_data[product] = sales
df = pd.DataFrame(sales_data)
df["month_number"] = months
# ---------------------------
# 2. MLflow Setup
# ---------------------------
mlflow.set_tracking_uri("http://localhost:5000")
mlflow.set_experiment("Multi-Product Dairy Sales Forecasting")
# ---------------------------
# 3. Function to Train & Log Models
# ---------------------------
def log_model(X_train, X_test, y_train, y_test, model, model_name, product_name):
with mlflow.start_run(run_name=f"{product_name} - {model_name}"):
# Train model
model.fit(X_train, y_train)
preds = model.predict(X_test)
# Metrics
mse = mean_squared_error(y_test, preds)
rmse = np.sqrt(mse)
mae = mean_absolute_error(y_test, preds)
r2 = r2_score(y_test, preds)
mlflow.log_metric("mse", mse)
mlflow.log_metric("rmse", rmse)
mlflow.log_metric("mae", mae)
mlflow.log_metric("r2", r2)
# Log parameters (if available)
try:
mlflow.log_params(model.get_params())
except:
pass
# Plot Actual vs Predicted
plt.figure(figsize=(8,5))
plt.plot(y_test.values, label="Actual Sales", marker='o')
plt.plot(preds, label="Predicted Sales", marker='x')
plt.title(f"{product_name} - {model_name} Actual vs Predicted")
plt.xlabel("Sample")
plt.ylabel("Sales")
plt.legend()
plt.tight_layout()
plot_file = f"{product_name}_{model_name}_plot.png"
plt.savefig(plot_file)
mlflow.log_artifact(plot_file)
plt.close()
# Log model (updated to use `name` instead of deprecated `artifact_path`)
mlflow.sklearn.log_model(model, name=f"{product_name}_{model_name}")
print(f"{product_name} - {model_name} → R2: {r2:.3f}, RMSE: {rmse:.2f}, MAE: {mae:.2f}")
# ---------------------------
# 4. Train & Log Models for Each Product
# ---------------------------
for product in products:
X = df[["month_number"]]
y = df[product]
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)
log_model(X_train, X_test, y_train, y_test, LinearRegression(), "Linear Regression", product)
log_model(X_train, X_test, y_train, y_test, DecisionTreeRegressor(), "Decision Tree", product)
log_model(X_train, X_test, y_train, y_test, RandomForestRegressor(), "Random Forest", product)
print("\nAll experiments completed! Check MLflow UI for results at http://localhost:5000")
the above file here explains that we are generating random data of sales for 3 years and we are deploying three models for each product like curd, cheese, milk and training models using linear regression, tree model and random forest
and now let us run python file to understand things better
and now if you open MLflow UI, you will be able to find experiment that we just created
and now in Runs, you will be able to find same product and different Models
and if we want to check which model is best suited for us, we can check boxes and click on compare
and now let us see if we can actually use this and Data scientist would check metrics and select best suited model for the forecast required
and now also, we can deploy these models in versions, train model, evaluate model and select best suited version
and to Register model
click on Models tab
select any Model that you want to register, click on Register Model
create New Model or you can select existing
once you select Model, you even have versioning enabled
As from above, once you evaluate all the tests, you can choose best version suitable
This concludes our blog here
Comments
Post a Comment