Deep Learning for
Python Engineers
Learn PyTorch, Transformers, and Reinforcement Learning by building real projects, not just reading theory. Every concept lands in production-quality code with tests, logging, and reproducible results.
No paywalls. No signup wall. Just structured content, starter code, and a clear roadmap from “I know Python” to “I train agents.”
Why this curriculum?
Most ML courses teach concepts in isolation. This one connects math, code, and engineering from day one.
Math you can actually use
Every equation, from Bellman optimality to GAE, is mapped directly to PyTorch code. No hand-waving.
Build before you use
You implement attention from scratch before touching Hugging Face. You code DQN before using Tianshou.
Production patterns
Type annotations, structured configs, reproducible seeds, and experiment logging are part of the learning path.
Hardware-aware
Every project explains what runs on M4, 4090, A100, or multi-GPU machines. No mystery hardware budgeting.
Three Core Tracks
Sequential or parallel, pick your path. Each track culminates in a portfolio-ready project.
PyTorch Foundations
Tensors, autograd, training loops, and GPU acceleration, everything you need to think in PyTorch before you open a research paper.
- ▸ MNIST classifier from scratch
- ▸ Custom Dataset + DataLoader
- ▸ Transfer learning on CIFAR-10
Transformer Architecture
Attention is all you need, and you will fully implement it. Scaled dot-product to multi-head attention to a full GPT in PyTorch.
- ▸ Self-attention layer in NumPy
- ▸ Build a mini-GPT
- ▸ Fine-tune BERT for classification
Reinforcement Learning
Bellman equations to PPO, derived, coded, and debugged. Train agents on CartPole, Atari Pong, and MuJoCo HalfCheetah.
- ▸ DQN for Atari Pong
- ▸ Tianshou CartPole pipeline
- ▸ PPO MuJoCo agent
Math Foundations
Linear algebra, calculus, probability, illustrated with NumPy and PyTorch instead of abstract symbolism.
Training & Inference
A bilingual overview of PyTorch, distributed training, mixed precision, vLLM, KV cache, profiling, and memory management.
Optimization Concepts
Core concepts like reward design, credit assignment, exploration vs exploitation, and off-policy vs on-policy.
Model Thinking
Understand long context, RAG, LoRA, hallucination, quantization, KV cache cost, and architecture tradeoffs.
Hardware Guide
GPU selection, cloud vs local training, mixed precision, and profiling, with hardware recommendation tables for each project.
Ready to start?
The roadmap takes you from Python competency to training MuJoCo agents in 14–20 weeks. Follow it sequentially, or jump straight to the track you need.