How amshaker/groupmamba works
GroupMamba is a visual state-space model project whose primary workflow trains and evaluates image-classification models. Training and evaluation configure the model and data pipeline, then run classification; the model uses grouped directional state-space blocks backed by selective-scan kernels. The README also describes distillation training and reports detection and segmentation results, but this tree exposes classification code only. Links reflect the excerpts, SOURCE INDEX references, and README; unsampled subsystem details are kept at directory level.
Main entry points: train.py Model training · eval.py Model evaluation · build.py Loader builder · selective_scan CUDA scan implementations
Components 19
Classification workflow
- Run configuration
classification/config.py - Model training
classification/train.py - Model evaluation
classification/eval.py
Image data
- Loader builder
classification/data/build.py - Folder and ZIP data
classification/data/cached_image_folder.py - ZIP reader
classification/data/zipreader.py - Specialized datasets
classification/data
GroupMamba model
- GroupMamba network
classification/models/groupmamba.py - SS2D blocks
classification/models/ss2d.py - Directional scans
classification/models/csms6s.py
Training support
- Optimizer builder
classification/utils/optimizer.py - LR scheduler
classification/utils/lr_scheduler.py - Cosine scheduler
classification/utils/cosine_lr.py - Logging
classification/utils/logger.py - Training utilities
classification/utils/utils.py
Selective scan kernels
- Scan API
kernels/selective_scan/csrc/selective_scan/selective_scan.h - CUDA scan implementations
kernels/selective_scan/csrc/selective_scan - ND-state API
kernels/selective_scan/csrc/selective_scan/cusndstate/selective_scan_ndstate.h
Other components
- Researcher
Connections 23
- Researcher → Model training: starts training
README.md - Researcher → Model evaluation: starts evaluation
README.md - Model training → Run configuration: loads settings
classification/train.py - Model evaluation → Run configuration: loads settings
classification/eval.py - Model training → Loader builder: builds loaders
classification/train.py - Model evaluation → Loader builder: builds loaders
classification/eval.py - Loader builder → Folder and ZIP data: constructs dataset
classification/data/build.py - Folder and ZIP data → ZIP reader: reads archive images
classification/data/cached_image_folder.py - Loader builder → Specialized datasets: selects dataset
classification/data/__init__.py - Model training → GroupMamba network: creates model
classification/train.py - Model evaluation → GroupMamba network: creates model
classification/eval.py - GroupMamba network → SS2D blocks: instantiates blocks
classification/models/groupmamba.py - SS2D blocks → Directional scans: uses scan functions
classification/models/ss2d.py - Directional scans → CUDA scan implementations: dispatches scan operations
classification/models/csms6s.py - CUDA scan implementations → Scan API: uses scan parameters
kernels/selective_scan/csrc/selective_scan/cus/selective_scan.cpp - CUDA scan implementations → ND-state API: uses ND-state API
kernels/selective_scan/csrc/selective_scan/cusndstate/selective_scan_ndstate.cpp - Model training → Optimizer builder: builds optimizer
classification/train.py - Model training → LR scheduler: builds scheduler
classification/train.py - LR scheduler → Cosine scheduler: uses cosine scheduler
classification/utils/lr_scheduler.py - Model training → Logging: logs progress
classification/train.py - Model training → Training utilities: uses utilities
classification/train.py - Model evaluation → Optimizer builder: builds optimizer
classification/eval.py - Model evaluation → Logging: logs results
classification/eval.py