Tools & Frameworks

The Software Stack

🔄 Lessons 3-5 covered algorithms, data, and evaluation. Now let’s look at the tools that turn these concepts into working systems. The ML software stack has clear layers, each with a specific job.

The Foundation: Python

Python dominates ML. Not because it’s the fastest language (it isn’t), but because its ecosystem of libraries makes ML practical:

Every ML project uses NumPy and pandas. The ML framework you add on top depends on your problem type.

pandas: Data Preparation

Every ML project starts with pandas. It’s the tool that loads, explores, cleans, and transforms data before any algorithm touches it.

What pandas does:

• Load data from CSV, Excel, SQL databases, JSON
• Explore data: column types, missing values, summary statistics
• Clean data: handle missing values, remove duplicates, fix formats
• Engineer features: create new columns, encode categories, aggregate
• Export clean data for ML training

The workflow:

Load CSV → Explore (describe, info) → Clean (fillna, dropna) →
Engineer features → Export for ML

pandas handles the 80% of ML work that isn’t algorithms. Data scientists joke that they spend 80% of their time on data preparation and 20% on actual modeling. pandas is the 80% tool.

scikit-learn: Traditional Machine Learning

scikit-learn is the gold standard for traditional ML algorithms. If your data lives in a spreadsheet and you’re not using neural networks, scikit-learn is almost certainly the right choice.

What it covers:

• Classification: random forests, SVM, logistic regression, k-nearest neighbors
• Regression: linear regression, decision trees, gradient boosting
• Clustering: K-means, DBSCAN, hierarchical clustering
• Preprocessing: scaling, encoding, imputation
• Evaluation: accuracy, precision, recall, cross-validation
• Model selection: grid search, hyperparameter tuning

Why ML practitioners love it: Consistent API. Every algorithm follows the same pattern: create the model, .fit() on training data, .predict() on new data. Learn the pattern once, apply it to any algorithm.

When to use scikit-learn: Structured/tabular data (spreadsheets, databases, CSV files), traditional algorithms (not deep learning), projects where interpretability matters, quick prototyping.

✅ Quick Check: You need to build a spam classifier. Your data is a CSV with 15 columns (word counts, sender info, email metadata). Would you use scikit-learn or PyTorch? scikit-learn — this is structured tabular data with 15 features. A random forest or logistic regression in scikit-learn handles this perfectly in a few lines of code. PyTorch is built for neural networks on unstructured data. Using it here adds complexity without adding value.

PyTorch: Deep Learning for Research

PyTorch is the dominant framework for deep learning research and education. 60%+ of beginners choose it first, and most new ML research papers use it.

Key strength: dynamic computation graphs. PyTorch code runs like regular Python — you can print values, set breakpoints, and step through execution line by line. This makes debugging intuitive, which is crucial when you’re learning or experimenting.

When to use PyTorch:

• Neural networks (CNNs for images, transformers for text, RNNs for sequences)
• Research and experimentation
• Learning deep learning concepts
• Projects where you need flexibility to customize architecture

TensorFlow: Deep Learning for Production

TensorFlow powers ML systems at Google and many large enterprises. It’s optimized for production deployment — serving models to millions of users, running on mobile devices, and scaling across data centers.

Key strength: production ecosystem. TF Serving deploys models as APIs. TF Lite runs models on mobile and edge devices. TF.js runs in browsers. The ecosystem is built for taking a trained model and putting it in front of users.

When to use TensorFlow:

• Production deployment at scale
• Mobile or edge device ML
• Enterprise ML infrastructure
• When your organization already uses TensorFlow

Keras: The Easy Button

Keras is a high-level interface for building neural networks. It abstracts away the complex details, letting you build and train models in a few lines.

What makes it beginner-friendly: Instead of defining every matrix multiplication and gradient calculation, you describe the network structure in plain terms: “A layer with 128 neurons, then a layer with 64 neurons, then an output layer with 10 classes.”

Keras runs on top of TensorFlow (it’s built into TensorFlow as tf.keras). Think of it as a simplified control panel for TensorFlow’s engine.

When to use Keras: Your first deep learning project, quick prototyping, when you want results fast without deep framework knowledge.

Choosing Your Framework

✅ Quick Check: A startup wants to build a recommendation system. They’ll prototype quickly, iterate fast, and eventually deploy to production. What’s their framework path? Start with scikit-learn for a baseline (collaborative filtering is traditional ML). If performance needs deep learning, prototype in PyTorch (faster iteration). Deploy the final model with TensorFlow Serving (production-ready). This prototype→deploy pattern — PyTorch for research, TensorFlow for production — is common in industry.

Key Takeaways

• Python is the language of ML — NumPy, pandas, matplotlib form the foundation
• pandas handles data preparation (80% of ML work) — loading, cleaning, engineering features
• scikit-learn is the gold standard for traditional ML on structured data — consistent API, broad algorithm coverage
• PyTorch dominates research and learning — dynamic graphs, intuitive debugging, 60% beginner adoption
• TensorFlow dominates production — model serving, mobile deployment, enterprise scale
• Keras simplifies deep learning — high-level interface on top of TensorFlow
• Choose by problem type: structured data → scikit-learn, neural networks → PyTorch/TensorFlow

Up Next

You understand the concepts, algorithms, data, evaluation, and tools. Lesson 7 puts it all in context — real-world ML applications across industries, plus the ethical challenges of algorithmic bias, fairness, and accountability.