What it does

This is a collection of tutorial code samples covering classic ML homework problems: MNIST digit recognition (from basic neural nets up to 99% accuracy CNNs) and OpenAI Gym reinforcement learning (CartPole and MountainCar via Q-Learning, DQN, and Policy Gradient). Each folder maps to a blog post on geektutu.com. The repo spans both TensorFlow 1.4 and 2.0 implementations, which is either handy or confusing depending on which version you’re actually using.

The interesting bit

The reinforcement learning progression is deliberately pedagogical: start with a Q-Table, swap it for a neural network (DQN), then graduate to Policy Gradient — all in ~70 lines. It’s a rare tutorial structure that actually shows you why you’d trade one approach for another, not just the final fancy version.

Key highlights

• MNIST CNN hits 0.99 accuracy with TensorFlow 2.0
• Covers model save/load, TensorBoard visualization, and custom HDF5/npy dataset creation
• RL examples span supervised learning → Q-Learning → DQN → Policy Gradient for the same environments
• All samples paired with Chinese-language blog posts (白话, or “plain spoken” style)
• Some examples still use TensorFlow 1.4; check folder-by-folder

Caveats

• No requirements.txt or setup.py visible; you’ll be fishing for dependencies
• Mixed TF versions mean you can’t blindly run everything in one environment
• 600 stars suggests modest reach; likely a personal blog companion rather than a maintained curriculum

Verdict

Good for Chinese-speaking developers who learn by reverse-engineering working code. Skip if you need polished English documentation, dependency management, or cutting-edge TF2 patterns — this is tutorial glue, not a framework.