cwdecode

A neural network research project.

This repository contains experimental code. The goal of this project is to investigate the possibility of using recurrent neural networks and connectionist temporal classification (CTC) for decoding morse code messages from raw audio data.

Setup

prepare the Python virtual environment and activate it:

$ . venv.sh

this form will not install CUDA libraries, and explicitly disable GPU devices.

To install CUDA libraries and enable the use of GPUs, source venv.sh WITH_CUDA set to a non-empty string:

$ WITH_CUDA=y . venv.sh

venv.sh also sets up LD_LIBRARY_PATH. Running the "deactivate" command will restore LD_LIBRARY_PATH.

generate_wav_samples.py

It is both a command-line tool for debugging and a module for generating data for training/validation.

For command line usage, issue:

$ ./generate_wav_samples.py --help"

tensorflow_lstm_ctc_train.py

Builds and trains a simple neural network with LSTM unit(s) using CTC as the loss function. The input is chunked raw audio. The model checkpoints and training / evaluation events are saved in a directory called "model_train".

Usage:

$ ./tensorflow_lstm_ctc_train.py

Training progress can be observed with:

$ tensorboard --log-dir=model_train

If you're particularly happy with one of the checkpoints, copy it manually into a directory called "model_use". To copy a checkpoint, you have to copy a meta, an index, and a data file. The checkpoint number and the type can be read in the files' names. A file called "checkpoint" is also needed. See the one in "model_training". The structure should be self-evident.

tensorflow_lstm_ctc_decode.py

Restores the latest checkpoint from the directory "model_use", and decodes a wav file given it's name as it's sole argument.

$ ./tensorflow_lstm_ctc_decode.py test.wav

The decoded text will appear under a pile of TF INFO messages.

Running with NVidia MPS

See: https://docs.nvidia.com/deploy/mps/latest/index.html

Start the MPS control process. To observe logs real-time, start it in the foreground:

$ nvidia-cuda-mps-control -f

Set up the environment in another terminal:

$ export CUDA_VISIBLE_DEVICES=0
$ export CUDA_MPS_PINNED_DEVICE_MEM_LIMIT=0=2G
$ ./tensorflow_lstm_ctc_train.py

You can safely run inference in an other terminal with less resources:

$ export CUDA_VISIBLE_DEVICES=0
$ export CUDA_MPS_PINNED_DEVICE_MEM_LIMIT=0=512M
$ ./tensorflow_lstm_ctc_train.py