ASCoT

Adaptive Semantic Communication with proportional DeepJSCC bandwidth allocation.

This repository contains the final conv-based code, selected checkpoints, experiment summaries, and figures used for the paper-ready release of the project.

Quick start

git clone https://github.com/mark000071/ASCoT.git
cd ASCoT
pip install -r requirements.txt

Set the COCO dataset as:

dataset/
  coco/
    train2017/
    val2017/
    annotations/
      instances_train2017.json
      instances_val2017.json

Then reproduce the final conv experiment matrix with:

DEVICE=cuda:0 \
SEMANTIC_ARCH=conv \
CHANNEL_TYPE=AWGN \
RUN_STAMP=march28prop \
IMAGE_SIZE=512 \
MAX_TRAIN_SAMPLES=40000 \
MAX_VAL_SAMPLES=2000 \
EPOCHS=20 \
BATCH_SIZE=2 \
TRAIN_SNRS="-10 0 10" \
TEST_SNRS="-10 0 10" \
bash scripts/run_architecture_fixed_snr_matrix.sh

What is included

• Final single-pass SemCom + DeepJSCC code
• Conv-based semantic encoder/decoder pipeline
• AWGN, Rayleigh, and Nakagami channel support
• Final three-channel experiment results
• Final ablation-study results
• Selected trained checkpoints for reproducibility
• Reproduction scripts and result-summary scripts

Repository layout

• src/
  • core model, data loading, training, evaluation, and adaptive inference code
• scripts/
  • experiment launchers and report-generation scripts
• weights/
  • final conv checkpoints for AWGN, Rayleigh, and Nakagami
• results/awgn/
  • final AWGN summaries and selected result tables
• results/channels/
  • merged AWGN/Rayleigh/Nakagami comparison tables and plots
• results/ablation/
  • final conv-only ablation summaries and plots
• figures/
  • selected figures organized by topic
• docs/
  • experiment notes and paper-writing helper material

Main model

The released main model is the single-pass pipeline:

x -> semantic encoder -> JSCC encoder -> channel -> JSCC decoder -> semantic decoder -> x_hat

The final released variant uses the conv-based semantic encoder/decoder and proportional semantic-symbol settings:

• small = (zc=16, jscc=16)
• medium = (zc=32, jscc=32)
• large = (zc=64, jscc=64)

Channels

The released code supports:

• AWGN
• Rayleigh
• Nakagami

The final experiments were run under fixed train/test SNR settings:

• -10 dB
• 0 dB
• 10 dB

Key scripts

• Training and matrix evaluation:
  • scripts/run_architecture_fixed_snr_matrix.sh
• Detection evaluation:
  • src/evaluate_yolo_detection.py
• Adaptive inference:
  • src/adaptive_semantic_inference.py
• Channel comparison report:
  • scripts/generate_channel_comparison_report.py
• Conv ablation suite:
  • scripts/run_conv_ablation_suite.sh
• Ablation summary:
  • scripts/summarize_conv_ablation.py

Environment

Install dependencies:

pip install -r requirements.txt

The original experiments used:

• Python 3.10
• PyTorch
• torchvision
• Ultralytics YOLO
• pycocotools
• lpips

Dataset

The experiments use COCO 2017 with the following layout:

dataset/
  coco/
    train2017/
    val2017/
    annotations/
      instances_train2017.json
      instances_val2017.json

Reproducing the final conv experiments

DEVICE=cuda:0 \
SEMANTIC_ARCH=conv \
CHANNEL_TYPE=AWGN \
RUN_STAMP=march28prop \
IMAGE_SIZE=512 \
MAX_TRAIN_SAMPLES=40000 \
MAX_VAL_SAMPLES=2000 \
EPOCHS=20 \
BATCH_SIZE=2 \
TRAIN_SNRS="-10 0 10" \
TEST_SNRS="-10 0 10" \
bash scripts/run_architecture_fixed_snr_matrix.sh

For Rayleigh or Nakagami, replace CHANNEL_TYPE with Rayleigh or Nakagami.

Reproducing the conv ablation suite

DEVICE=cuda:0 bash scripts/run_conv_ablation_suite.sh
python scripts/summarize_conv_ablation.py

Notes

• This release intentionally keeps only the final conv-based pipeline and the final selected outputs needed for open-source release and reproduction.
• The repository includes selected final checkpoints, not every intermediate training artifact.
• Large per-image raw logs from every experiment are not duplicated here; the released summaries and figures are the final curated results.

Reproducibility notes

• Source code lives in src/, and the provided shell scripts automatically export PYTHONPATH=src.
• The final release includes selected checkpoints under weights/ for the three channels and three train-SNR conditions.
• YOLO inference requires a local Ultralytics checkpoint, passed through YOLO_WEIGHTS.
• The release has been checked for command-line entry points and script-path consistency after the repo layout was flattened for open-source release.

Citation

If you use this repository, please cite it with the metadata in CITATION.cff.