Belief propagation with sparse matrices (scipy.sparse) in Python

Includes implementations of a few message passing algorithms and a few other decoders for LDPC codes

• Message passing (Belief Propagation) based min-sum (MSA) and sum-product (SPA) algorithms using sparse matrices (scipy.sparse)
• maximum-likedlood (ML) and linear-programming (LP) decoders (only for short length codes like Hamming(7,4)) based on Using Linear Programming to Decode Binary Linear Codes
• ADMM decoder based on Decomposition Methods for Large Scale LP Decoding

for binary erasure (BEC), binary symmetric (BSC) and binary-AWGN (biawgn) channels.

Dependencies

Tested on the following Python/package versions:

• Python version 3.5.2
• numpy version 1.12.0
• scipy version 0.18.1

Exploring pre-computed results

See all pre-compupted results in codes, output and plots. These include different codes, simulation results and plots.

---

Starting from scratch

When cloning the repository for the first time, initialize the src/utilities sub-module (this is another git repository) by executing git submodule update --init --recursive as described in this stackoverflow question.

Generating data

• Simulation output written to ~/scratch/decoders by default.
• Execute src/main.py with following or equivalent arguments:
  • python src/main.py biawgn 7_4_hamming SPA --codeword=1 --params .01 .05 .1 .5 1 2 4 6 --console
  • python src/main.py bec 1200_3_6_rand_ldpc_1 SPA --codeword=1 --console --params .5 .475 .45 .425 .4 .375 .35 .325 .3
• See the output of python src/main.py --help for descriptions of valid arguments.
• Run FILE_CODES_DIR=~/my/codes/directory python src/main.py <OTHER ARGUMENTS> to change the default directory for codes.

Generating graphs

• Execute src/graph.py with following or equivalent arguments:
  • python src/graph.py --and bec-7_4_hamming --or_ SPA ML --error wer --leg decoder
  • python src/graph.py --and bec-1200_3_6 SPA --error ber
• See the output of python src/graph.py --help for descriptions of valid arguments.

Generating optimal irregular LDPC codes

• Generate the optimal irregular distribution a given check node distribution (rho):

  python src/ldpc.py irg --count=10 --len=1200 --rho=5 --rate=.5

• Plot the density evolution:

  python src/ldpc.py plt

Luby-Transform (LT) code

Execute the following to reproduce the Figure 50.4 (histogram for LT code with length 10000) in 'Information Theory, Inference, and Learning Algorithms' by David J.C. MacKay. Replace in the first command with 0.01, 0.03, 0.1.

• python -u src/luby.py 10000 12000 <c> .5 250 --pool=4
• python src/luby_graph.py .01 .03 .1

---

Regenerating results in output and plots

Generating data

• Execute python src/codes.py 10 1200 3 6 to generate 10 random samples from LDPC(1200,3,6) ensemble.

• Run simulations using run_sims.sh {PARA} {CASE} {ARGS} command. For example,

  • ./run_sims.sh SEQL HMG executes all Hamming code related simulations sequentially.
  • ./run_sims.sh SEQL REG_ENS --data_dir=./data --console executes some regular LDPC related simulations sequentially while printing logs onto console.
  • ./run_sims.sh PARA REG_ENS --data_dir=./data executes the same in parallel.
  • Use the latter only on a dedicated server as it will take large amount of CPU.
  • See run_sims.sh for other choices of {CASE}.

• If running the simulations on Niagara cluster, need to first activate a Python environment with required packages installed.

• Execute ./run_sims.sh PARA HMG --data_dir=$SCRATCH to test if simulations run properly.

• Run all simulations in parallel within one node:

  submitjob single -E src_nia -e "cd ~/projects/decoders" "./run_sims.sh PARA HMG MAR REG_BAD REG_ENS IREG_ENS --console --data_dir ~/scratch/decoders/output" -d time=11:59:00 -D email --print

Generating graphs

• Execute ./plot_results.py HMG --data_dir=./data --plots_dir=./plots --save --silent --error=ber to silently save Hamming code related plots.
• Replace HMG with HMG REG_ENS to save both Hamming code and regular ensemble related plots.
• If running on the Niagara cluster execute ./plot_results.py HMG REG_ENS --data_dir=$SCRATCH --plots_dir=$SCRATCH --silent --agg to use the proper back-end for matplotlib.