Yuan Lab袁静课题组

Stanford's classic CV course — the full deep-learning vision stack from image classification to detection and segmentation, with public lecture notes.

An open interactive deep-learning textbook with runnable code, adopted by 500+ universities; available in multiple languages.

The classic top-down free course — you train a working model in lesson one.

MIT's annually updated one-semester intro covering deep-learning fundamentals through generative models; all videos and slides public.

Google's free crash course with visual exercises for building core ML concepts quickly.

Harvard's most popular course — solid programming and CS fundamentals for absolute beginners.

Official free courses on Transformers, LLM fine-tuning and agents — directly relevant to LLM research.

A beginner's guide to isolated environments with conda, choosing the right PyTorch build for your GPU, and verifying the install.

A widely-adopted one-stage object detector with polished engineering, thorough docs and a low reproduction barrier — a good entry point for detection projects.

OpenMMLab's modular detection toolbox supporting a large family of detection and segmentation algorithms — suited for systematically reproducing and benchmarking methods.

An attention-based network for P/S-wave arrival picking, with released pretrained weights, sample data and a Google Colab notebook for end-to-end reproduction.

ETH's deep-learning toolkit for seismology, providing a unified interface to pretrained models and benchmark datasets for tasks like detection and phase picking.

Official PyTorch example suite covering image classification, DCGAN, reinforcement learning and more — concise reference implementations for reproducing core workflows.

A unified interface to a large family of NLP and multimodal pretrained models, with thorough docs and examples for reproducing Transformers and fine-tuning workflows.

How to use the 20 practice projects (9 beginner + 11 intermediate): reading the guides, downloading, troubleshooting, and a suggested learning order. Spend 10 minutes here before starting.

A from-scratch perceptron with three hand-crafted features for fire/normal image classification — NumPy only, the lightest first step.

A pure-NumPy MLP with hand-written forward/backward passes — no frameworks, ideal for understanding how neural networks really work.

Build a CNN in PyTorch to classify fire vs. normal images; ships with a trained model and a one-click script.

An RNN learns from a full year of hourly water-level records to forecast future levels — meet networks with memory.

Same water-level task with an LSTM — pair it with the RNN project to see what gating really buys you.

A mini Transformer with a self-built vocabulary classifies emergency text into six categories, plus a local web demo.

An autoencoder learns the hazy-to-clear mapping; haze is synthesized on the fly, so no paired dataset is needed.

Generator vs. discriminator adversarial training for dehazing — compare the results with the autoencoder version.

A conditional U-Net DDPM removes synthesized rain — experience the noise-and-denoise idea behind generative models.

Cleans and fuses 7 public data sources (weather, earthquakes, FEMA, CDC SVI, etc.) into unified tables, with a Streamlit dashboard for AI risk scoring and anomaly alerts. Great for learning data engineering.

Extracts 49 earthquake-preparedness documents into an entity-relation knowledge graph with global/local/gap search, published as an interactive web atlas. Great for learning cutting-edge GraphRAG.

Detects and repairs 7 types of rain gauge anomalies (stuck, spikes, negatives...) using physics rules plus ML, with ground-truth evaluation (F1 98.7%). The best first project for data thinking.

Learns 18 years of real river-gauge data with LSTM/BiLSTM to forecast the next 24 hours of water levels with risk grading. A classic deep-learning time-series starter.

A Flask web app that generates personalized earthquake warning texts via a rule engine, magnitude grading, and haversine distance. The best first web project.

BERT+CRF extracts location/person/disaster/need entities from distress texts, geocoded into an interactive rescue map (entity F1 0.95). The most substantial NLP project.

Solves the multi-warehouse, multi-victim, multi-supply 5-objective allocation problem with an enhanced genetic algorithm, visualized in a Vue3+ECharts dashboard.

Overlays historical disaster buffers on the real Beijing road network and compares shortest vs hazard-aware routes via Dijkstra. Only one third-party dependency—the quickest starter.

A FastAPI+Vue3 command dashboard: public data ingestion, map/chart visualization, and a 7-step AI decision workflow (auto-degrades to local simulation without an API key). Full-stack engineering practice.

Fuses USGS earthquake catalog, NOAA losses, and BEA state GDP into a panel, predicting recovery years with Random Forest vs GBDT. The full tabular-ML workflow.

Decrypts the data, reads 12 GIS raster layers, and runs AHP susceptibility overlay, 4-level rainfall warning simulation, and village-to-shelter evacuation planning. GIS + crypto licensing.

MIT's open archive of full courses — lecture notes and exams freely available.

Official tutorials from tensors to deployment — the first stop for both API reference and patterns.

The famed interactive ML journal (archived, fully readable) — several articles on attention and visualization are classic primers.