How zhouhaoyi/informer2020 works
Project type and purpose
• This is an open-source machine-learning library implementing the Informer Transformer architecture for long-sequence time-series forecasting.
• It provides data loaders, model code (encoder, decoder, attention), utilities, experiment scripts, and a main orchestration script (main_informer.py) to train/test and evaluate forecasting models on ETT/ECL/Weather datasets.
File structure
• Top-level files:
• Dockerfile, Makefile, environment.yml, requirements.txt: environment setup & reproducibility
• README.md, LICENSE: documentation and licensing
• main_informer.py: entry point for training, testing, prediction
• Directories:
• data/: dataset loaders (data_loader.py)
• models/: core model components (encoder.py, decoder.py, attn.py, embed.py, model.py)
• utils/: helper functions (masking, metrics, time-feature encodings, generic tools)
• exp/: example experiment flows
• scripts/: shell scripts to run experiments for different datasets
README for additional insights
• It lists dependencies (PyTorch, numpy, pandas, etc.) and shows how to run experiments via main_informer.py or Docker/Make.
a. Main components
1. Data Sources: ETT, ECL, Weather CSV files in data/ETT/, data/ECL/, etc.
2. Data Loader & Preprocessor: data/data_loader.py + utils/timefeatures.py, utils/masking.py
3. Model Core:
• Embedding Layer (models/embed.py): token & positional/time encoding
• Encoder Stack (models/encoder.py): ProbSparseAttention (models/attn.py), distillation
• Decoder Stack (models/decoder.py): generative decoder with mix attention
• Model Orchestration (models/model.py): connects encoder, decoder, final linear output
4. Training & Evaluation Engine: main_informer.py
• Argument parsing, dataset instantiation, model instantiation
• Optimizer, loss (utils/metrics.py), learning-rate schedule
• Checkpointing, logging, optional AMP, multi-GPU
5. Experiment Scripts: exp/, scripts/ driving reproducible workflows
6. External Dependencies: PyTorch, numpy, pandas, scikit-learn, matplotlib
b. Relationships & interactions
• Data files → Data Loader → Preprocessor → Data Batches → Model Core
• Model Core (encoder + decoder) ↔ Loss & Metrics → Training loop
• Training loop → Checkpoints + Logs + Evaluation → Prediction outputs & Figures
• Scripts & Makefile/Docker orchestrate data download, Docker build, environment setup, batch runs
c. Architectural patterns & design principles
• Modular ML pipeline: clear separation between data ingestion, preprocessing, model definition, training logic
• Encoder–decoder Transformer pattern with custom sparse attention (ProbSparse)
• Experiment reproducibility: containerization (Docker), scripted runs, parameter files (.sh scripts)
• Extensible: new datasets plug into data_loader.py; new attention/encoder variants extend base classes
d. Relevant technologies & frameworks
• PyTorch 1.x for deep learning
• Python 3.6, numpy/pandas for data manipulation
• scikit-learn for auxiliary tools
• Docker & Make for environment reproducibility
• matplotlib for plotting