How spiderforest/mixercseg works
MixerCSeg is a crack-image segmentation system: it loads labeled images, trains a multi-stage mixer encoder and refinement decoder, produces pixel-level predictions, and evaluates saved results. The README also describes benchmark/checkpoint inference, while the provided tree does not include the referenced test.py. Some model and orchestration source is unsampled, so the graph omits uncertain wiring rather than inferring a complete call chain.
Main entry points: engine.py Epoch trainer · MixerCSeg.py MixerCSeg · vss_block.py VSS encoder · evaluate.py Metric evaluation
Components 20
Training and evaluation
- Training runner
main.py - Epoch trainer
engine.py - Metric evaluation
eval/evaluate.py - Evaluation I/O
eval/data_io.py - Segmentation loss
main.py - Optimizer and scheduler
main.py - Prediction results
Crack data
- Dataset transforms
datasets/base_dataset.py - Crack dataset loader
datasets/crack_dataset.py - Folder image loader
datasets/image_folder.py
Segmentation model
- MixerCSeg
models/segmentor/MixerCSeg.py - VSS encoder
models/encoder/vss_block.py - Patch embedding
models/layers/patch_embed.py - Inception branches
models/inception.py - HoG edge gate
models/layers/hog_edge.py - SRF decoder
models/decoder/SRF.py
State-space compute
- VMamba scan
VMamba/models/vmamba.py - Selective scan kernels
VMamba/models/kernels/selective_scan/csrc/selective_scan/selective_scan.h
Other components
- Researcher
- Crack dataset
Connections 18
- Researcher → Training runner: configures training
- Crack dataset → Crack dataset loader: provides samples
- Crack dataset → Folder image loader: provides images
- Crack dataset loader → Dataset transforms: uses transforms
- Training runner → Epoch trainer: runs epochs
- Epoch trainer → MixerCSeg: forwards batches
- Epoch trainer → Segmentation loss: computes loss
- Epoch trainer → Optimizer and scheduler: updates weights
- MixerCSeg → VSS encoder: extracts features
- VSS encoder → Patch embedding: embeds images
- VSS encoder → HoG edge gate: gates edges
- VSS encoder → VMamba scan: runs state-space mixing
- MixerCSeg → Inception branches: mixes features
- MixerCSeg → SRF decoder: refines features
- VMamba scan → Selective scan kernels: dispatches scan
- Metric evaluation → Evaluation I/O: loads result data
- Metric evaluation → Prediction results: scores predictions
- Researcher → Metric evaluation: requests metrics