TransCoder

This repo is providing the official implementation of TransCoder framefork from paper "TransCoder: A Neural-Enhancement Framework for Channel Codes".

Proposed TransCoder Framework

Running training for TransCoder for LDPC code (121, 60)

python main.py --outer_code_type=LDPC_nn --K=121 --padd_symbol_len=2 --num_iters=2 --low_bp=3 --high_bp=6 --train=1 --d_k_trx=16 --device=cuda --K_in=60

Running the TransCoder for other codes can be found in run_train.sh

License

Codes are available only for non-commercial research purposes.

Acknowledgement

As TransCoder framework relies on the channel coding part, here we provide the references to other projects that our framework is using.

1. The channel code matrices (except LDPC codes with code length 384) and functions processing these matrices were taken from ECCT.
2. The turbo code pipeline was adapted from TinyTurbo.
3. The training loss for polar codes with SC decoding structure was adapted from npd-communication.
4. Belief propagation (BP) for LDPC and BCH codes and linear code encoding/decoding has similar implementation to kaira.