How prs-eth/thera works
Thera is a JAX/Flax arbitrary-scale image super-resolution system. Its main workflow takes an image and requested scale, builds a model from a checkpoint, and predicts and saves an upscaled image; separate workflows train models and evaluate them across datasets and scales. The graph also shows the model’s hypernetwork/heat-field design and the documented EDSR and RDN backbones. Relationships for unsampled inference code follow the README and SOURCE INDEX; uncertain implementation details are omitted.
Main entry points: run_train.py Training loop · thera.py Thera builder and fields · run_eval.py Evaluation runner · super_resolve.py Super-resolve CLI
Components 20
User workflows
- Super-resolve CLI
super_resolve.py - Train/eval runner
run_train_and_eval.py
Data and evaluation
- Image datasets
data.py - Image utilities
utils.py - Evaluation options
args/eval.py - Evaluation runner
run_eval.py
Thera model
- Thera builder and fields
model/thera.py - Hypernetwork
model/hyper.py - EDSR backbone
model/edsr.py - RDN backbone
model/rdn.py - Backbone tails
model/tail.py - ConvNeXt blocks
model/convnext.py - SwinIR blocks
model/swin_ir.py
Training
- Training options
args/train.py - Training loop
run_train.py
Other components
- Image user
- Model trainer
- Image dataset
- Model checkpoint
- Weights & Biases
Connections 28
- Image user → Super-resolve CLI: requests upscale
README.md - Super-resolve CLI → Thera builder and fields: builds model
super_resolve.py - Super-resolve CLI → Model checkpoint: loads parameters
- Super-resolve CLI → Image utilities: uses helpers
super_resolve.py - Super-resolve CLI → Image user: saves result
README.md - Model trainer → Train/eval runner: starts training
README.md - Train/eval runner → Training loop: runs training
run_train_and_eval.py - Train/eval runner → Evaluation runner: runs evaluation
run_train_and_eval.py - Training loop → Training options: reads options
run_train.py - Training loop → Image datasets: loads samples
run_train.py - Training loop → Thera builder and fields: builds model
run_train.py - Training loop → Image utilities: uses metrics
run_train.py - Training loop → Model checkpoint: writes parameters
run_train.py - Training loop → Weights & Biases: logs metrics
run_train.py - Image datasets → Image dataset: reads images
- Image datasets → Image utilities: uses grid and resize
data.py - Evaluation runner → Evaluation options: reads options
run_eval.py - Evaluation runner → Image datasets: loads evaluation images
run_eval.py - Evaluation runner → Model checkpoint: loads parameters
run_eval.py - Evaluation runner → Thera builder and fields: builds model
run_eval.py - Evaluation runner → Image utilities: computes metrics
run_eval.py - Thera builder and fields → Hypernetwork: constructs hypernetwork
model/thera.py - Thera builder and fields → EDSR backbone: selects backbone
model/thera.py - Thera builder and fields → RDN backbone: selects backbone
model/thera.py - Thera builder and fields → Backbone tails: builds tail
model/thera.py - Hypernetwork → Image utilities: interpolates grid
model/hyper.py - Backbone tails → ConvNeXt blocks: uses blocks
model/tail.py - Backbone tails → SwinIR blocks: uses blocks
model/tail.py