How mr-eggplant/sar works
This repository evaluates test-time adaptation methods on ImageNet-C streams. The principal workflow is: configure an experiment, prepare corruption data (and optionally impose label-shift ordering), run inference with no adaptation or an adaptation wrapper, then report accuracy. SAR is the central method; Tent and EATA are documented alternatives. The graph retains the model, data-preparation, and metric layers. Main-to-method dispatch is partly unsampled; those links rely on main.py’s indexed references rather than inspected branch code.
Main entry points: main.py Experiment runner · generate_shifted_sample_indices.py Shifted ordering · sar.py SAR · selectedRotateImageFolder.py Corruption loader
Components 13
Experiment runner
- Experiment runner
main.py
Test data
- Corruption loader
dataset/selectedRotateImageFolder.py - Shifted ordering
dataset/generate_shifted_sample_indices.py - Order arrays
dataset
Adaptation methods
- No adaptation
main.py - Tent
tent.py - EATA
eata.py - SAR
sar.py - SAM optimizer
sam.py
Models and metrics
- ResNet model
models/Res.py - Accuracy metrics
utils/cli_utils.py - Experiment utilities
utils/utils.py
Other components
- Experiment user
Connections 12
- Experiment user → Experiment runner: runs experiment
README.md - Experiment runner → Corruption loader: prepares data
main.py - Experiment runner → Shifted ordering: generates ordering
- Shifted ordering → Order arrays: saves indices
dataset/generate_shifted_sample_indices.py - Experiment runner → No adaptation: evaluates directly
main.py - Experiment runner → Tent: configures and evaluates
main.py - Experiment runner → EATA: dispatches method
main.py - Experiment runner → SAR: dispatches method
main.py - Experiment runner → ResNet model: selects model
main.py - Experiment runner → Accuracy metrics: validates accuracy
main.py - Experiment runner → Experiment utilities: gets logger
main.py - SAR → SAM optimizer: uses optimizer
README.md