How giabao2478/thesis-master works
SwinJSCC is a learned image joint source–channel coding system: a training or evaluation run loads image batches, encodes them, optionally simulates a communication channel, decodes reconstructions, and evaluates distortion. Its model supports multiple channel conditions and nested-rate/MRL variants. The README is empty, so the map relies on the source index and shown excerpts; some training and evaluation details are not visible in the sampled portions of main.py.
Main entry points: network.py SwinJSCC model · main.py Train and evaluate · decoder.py Swin decoder · encoder.py Swin encoder
Components 12
Run orchestration
- Train and evaluate
SwinJSCC/main.py - Image datasets
SwinJSCC/data/datasets.py - Run utilities
SwinJSCC/utils.py
Image codec
- SwinJSCC model
SwinJSCC/net/network.py - Swin encoder
SwinJSCC/net/encoder.py - Swin decoder
SwinJSCC/net/decoder.py - Transformer blocks
SwinJSCC/net/modules.py
Channel and rates
- Channel simulation
SwinJSCC/net/channel.py - MRL projections
SwinJSCC/net/mrl.py
Losses and quality
- Distortion metrics
SwinJSCC/loss/distortion.py - Matryoshka loss
SwinJSCC/loss/matryoshka_loss.py
Other components
- Run initiator
Connections 13
- Run initiator → Train and evaluate: starts run
- Train and evaluate → Image datasets: loads batches
SwinJSCC/main.py - Train and evaluate → SwinJSCC model: trains and evaluates
SwinJSCC/main.py - Train and evaluate → Run utilities: uses utilities
SwinJSCC/main.py - SwinJSCC model → Swin encoder: encodes images
SwinJSCC/net/network.py - SwinJSCC model → Channel simulation: transmits latents
SwinJSCC/net/network.py - SwinJSCC model → Swin decoder: decodes latents
SwinJSCC/net/network.py - SwinJSCC model → Distortion metrics: measures distortion
SwinJSCC/net/network.py - SwinJSCC model → Matryoshka loss: computes nested loss
SwinJSCC/net/network.py - SwinJSCC model → MRL projections: projects prefixes
SwinJSCC/net/network.py - Swin encoder → Transformer blocks: uses blocks
SwinJSCC/net/encoder.py - Swin decoder → Transformer blocks: uses blocks
SwinJSCC/net/decoder.py - Swin decoder → Swin encoder: reuses transformer block
SwinJSCC/net/decoder.py