How pranavv0307/rs-uep-jscc works
This research codebase studies image transmission over noisy or erased channels, with ADJSCC-Q plus Reed–Solomon unequal error protection as its principal path: load CIFAR-10, encode and quantize images, rank latent symbols by importance, protect and transmit packets, reconstruct images, and evaluate quality. Training, clean evaluation, AWGN demos, and erasure studies are separate workflows around the model and coding stages. Edges follow the shown excerpts, SOURCE INDEX references, or documented behavior; unsampled subsystem details are not inferred.
Main entry points: toy_demo.py AWGN demo · eval_rs_uep.py RS-UEP evaluation · train_adjsccq_train_split.py ADJSCC-Q training · rs_pipeline.py RS tier pipeline
Components 19
Training & Evaluation
- ADJSCC-Q training
experiments/train_adjsccq_train_split.py - Clean evaluation
experiments/eval_adjsccq_clean.py - RS-UEP evaluation
experiments/eval_rs_uep.py - Image metrics
experiments/metrics.py - AWGN demo
experiments/toy_demo.py
Image Codecs
- Quantized ADJSCC
models/adjsccq.py - ADJSCC backbone
models/adjscc.py - DeepJSCC baseline
models/deepjscc.py - Attention module
models/Attention.py - Scalar quantizer
models/quantizer.py - AWGN channel
models/channel.py
Packet Protection
- Importance ranking
coding/importance.py - RS tier pipeline
coding/rs_pipeline.py - Packetizer
coding/packetizer.py - Erasure channel
coding/erasure_channel.py - Reed–Solomon
coding/reed_solomon.py - Tier allocation
coding/tiering.py
Image Data
- CIFAR-10 loader
data/cifar10.py
Other components
- Researcher
Connections 26
- Researcher → ADJSCC-Q training: starts training
- Researcher → Clean evaluation: evaluates codec
- Researcher → RS-UEP evaluation: tests erasures
- CIFAR-10 loader → ADJSCC-Q training: provides images
experiments/train_adjsccq_train_split.py - CIFAR-10 loader → Clean evaluation: provides images
experiments/eval_adjsccq_clean.py - ADJSCC-Q training → Quantized ADJSCC: trains model
experiments/train_adjsccq_train_split.py - ADJSCC-Q training → Image metrics: computes quality
experiments/train_adjsccq_train_split.py - ADJSCC-Q training → RS tier pipeline: evaluates protection
experiments/train_adjsccq_train_split.py - Clean evaluation → Quantized ADJSCC: runs codec
experiments/eval_adjsccq_clean.py - Clean evaluation → Image metrics: scores reconstructions
experiments/eval_adjsccq_clean.py - RS-UEP evaluation → CIFAR-10 loader: loads images
experiments/eval_rs_uep.py - RS-UEP evaluation → RS tier pipeline: evaluates packet coding
experiments/eval_rs_uep.py - RS-UEP evaluation → Image metrics: scores reconstructions
experiments/eval_rs_uep.py - Quantized ADJSCC → ADJSCC backbone: wraps backbone
models/adjsccq.py - Quantized ADJSCC → Scalar quantizer: quantizes latents
models/adjsccq.py - ADJSCC backbone → Attention module: uses attention
models/adjscc.py - AWGN demo → CIFAR-10 loader: loads images
experiments/toy_demo.py - AWGN demo → ADJSCC backbone: runs encoder
experiments/toy_demo.py - AWGN demo → DeepJSCC baseline: runs baseline
experiments/toy_demo.py - AWGN demo → AWGN channel: transmits latents
experiments/toy_demo.py - AWGN demo → Image metrics: scores reconstructions
experiments/toy_demo.py - RS tier pipeline → Importance ranking: orders symbols
coding/rs_pipeline.py - RS tier pipeline → Packetizer: forms packets
coding/rs_pipeline.py - RS tier pipeline → Reed–Solomon: encodes tiers
coding/rs_pipeline.py - Erasure channel → Packetizer: erases packets
coding/erasure_channel.py - RS-UEP evaluation → Erasure channel: simulates packet loss
experiments/eval_rs_uep.py