How ming053l/drct works
DRCT is an image-super-resolution research application: users train or load a model, supply low-resolution images, and receive upscaled images. Its main runtime paths are BasicSR-based training and standalone tiled/non-tiled inference; the README also documents Real-DRCT-GAN variants and visualization/saliency tools. The model excerpt supports the DRCT feature-extraction and reconstruction stages. Several model, dataset, and visualization relationships are described only at subsystem level because their callers were not excerpted.
Main entry points: DRCT_arch.py DRCT network · BackProp.py Saliency gradients · inference.py Inference runner · train.py Training entry point
Components 21
Image inference
- Inference runner
inference.py - Half-precision runner
inference_fp16.py - ONNX inference
REAL_DRCT_GAN_ONNX inference_fp16.py
SR models
- DRCT network
drct/archs/DRCT_arch.py - Dense feature blocks
drct/archs/DRCT_arch.py - SR reconstruction
drct/archs/DRCT_arch.py - GAN discriminator
drct/archs/discriminator_arch.py - SRVGG architecture
drct/archs/srvgg_arch.py
Training
- Paired dataset
drct/data/imagenet_paired_dataset.py - DRCT training model
drct/models/drct_model.py - Real DRCT MSE
drct/models/realdrctmse_model.py - Real DRCT GAN
drct/models/realdrctgan_model.py - Training configurations
options/train - Training entry point
drct/train.py
Model visualization
- Model zoo helpers
Visualization/ModelZoo/utils.py - Saliency gradients
Visualization/SaliencyModel/BackProp.py - Saliency utilities
Visualization/SaliencyModel/utils.py
Other components
- Researcher
- LR images
- Upscaled images
- Visualization user
Connections 22
- Researcher → Inference runner: runs
inference.py - LR images → Inference runner: reads
inference.py - Inference runner → DRCT network: calls
inference.py - Inference runner → Upscaled images: writes
inference.py - Researcher → Half-precision runner: runs
inference_fp16.py - Half-precision runner → DRCT network: calls
inference_fp16.py - Researcher → ONNX inference: runs
- DRCT network → Dense feature blocks: uses
drct/archs/DRCT_arch.py - DRCT network → SR reconstruction: reconstructs
drct/archs/DRCT_arch.py - Researcher → Training entry point: starts
drct/train.py - Training entry point → DRCT training model: registers
drct/train.py - Training entry point → Paired dataset: registers
drct/train.py - Training configurations → Training entry point: configures
- DRCT training model → DRCT network: trains
- Real DRCT MSE → DRCT network: trains
- Real DRCT GAN → GAN discriminator: uses
- Real DRCT MSE → Real DRCT GAN: fine-tunes
README.md - Visualization user → Model zoo helpers: explores
README.md - Visualization user → Saliency gradients: inspects
README.md - Saliency gradients → DRCT network: backpropagates
Visualization/SaliencyModel/BackProp.py - Saliency gradients → Model zoo helpers: uses helpers
Visualization/SaliencyModel/BackProp.py - Saliency gradients → Saliency utilities: uses utilities
Visualization/SaliencyModel/BackProp.py