How kumarr99/gnn_pathplanning_do works
Project Type and Purpose
• This is a research‐oriented PyTorch codebase implementing Graph Neural Networks (GNNs) for decentralized multi‐robot path finding (MAPF), integrating ML training pipelines with “expert” solvers for data generation and curriculum learning.
• It is *not* a web app or microservice ensemble, but rather a modular ML framework combining data generation, transformation, training/inference, evaluation, and visualization.
Main Components
A. Data Generation & Expert Solvers
• offlineExpert: classical MAPF case generator (random maps + agent start/goal pairs) and solver modules (CBS, ECBS, SIPP)
• onlineExpert: online‐expert fallback for failed cases during training
B. Data Transformation
• offlineExpert/DataGen_Transformer.py & onlineExpert/DataTransformer_local_onlineExpert.py
• Convert solver outputs into tensors, graph shift operators (GSO), map encodings for model ingestion
C. Data Loading
• dataloader/: custom PyTorch Dataset classes (e.g. Dataloader_dcplocal_notTF_onlineExpert.py)
D. Core Model & Training
• graphs/models/decentralplanner.py: CNN→GNN→MLP pipeline
• graphs/losses: loss functions (cross‐entropy, Huber, BCE, etc.)
• agents/: high‐level abstractions for training/testing loops (decentralplannerlocal, with/without online expert)
• main.py: orchestrates config parsing, mode selection (train/test), experiment logging
E. Configuration & Scripts
• configs/*.json: hyperparameters & file‐paths for each training/inference scenario
• scripts/: shell scripts for running experiments locally or on clusters
F. Utilities
• utils/: visualization (paths, communication links), metrics collection, directory management, graph utilities, movie generation
G. Analysis & Results
• statistic_analysis/: generate figures and error‐bar/histogram plots for published results
Relationships & Data Flow
1. offlineExpert → (raw cases) → DataGen_Transformer → serialized dataset files
2. onlineExpert (optional) → augment training set with expert solutions for failure cases
3. dataloader → reads serialized datasets → yields training batches
4. agents.train loop
└─> graphs/models (forward on batch)
└─> graphs/losses (compute loss)
└─> optimizer.step()
6. statistic_analysis consumes experiment logs → plots
Architectural Patterns & Design Principles
• Modular separation: data gen, data transform, data load, model definition, agent logic, visualization, analysis
• Dependency injection via JSON configs: decouples hyperparameters and filesystem paths from code
• Clear layering:
• Layer 1: Data sources (expert solvers)
• Layer 2: Data adaptors (transformers, dataloader)
• Layer 3: Model & training orchestration (agents, graphs)
• Layer 4: Post‐processing (metrics, visualization, analysis)
• Reuse & extensibility: “agents” base class, configurable solver backends, plug‐in loss modules
Technologies & Frameworks
• PyTorch for neural networks and training loop
• tensorboardX for logging
• numpy, scipy, scikit‐image for data manipulation
• Custom graph utilities for GSO construction
• Shell scripts for HPC/cluster integration
Detailed Element Checklist
• Optional: annotate external dependencies (PyTorch, numpy) at the periphery