warp-ctc requires torch
and running ./install.sh for torch requires

sudo ./clean.sh
sudo TORCH_NVCC_FLAGS="-D__CUDA_NO_HALF_OPERATORS__" ./install.sh 
and answer yes at the end

then create a new shell
rm -rf warp-ctc


git clone https://github.com/baidu-research/warp-ctc.git
cd warp-ctc

create a build directory:

mkdir build
cd build
if you have a non standard CUDA install export CUDA_BIN_PATH=/path_to_cuda so that CMake detects CUDA and to ensure Torch is detected, make sure th is in $PATH

run cmake and build:

cmake ../
make

sudo make install

edit .bashrc and add

cd ~/
install mxnet but change 
make/config.mk in incubator-mxnet directory by uncommenting:

WARPCTC_PATH = $(HOME)/warp-ctc
MXNET_PLUGINS += plugin/warpctc/warpctc.mk

cd ~/
git clone --recursive https://github.com/apache/incubator-mxnet.git mxnet 
cd mxnet
cd make
vi config.mk
# do the uncommenting mentioned above
cd ../..
make -j USE_BLAS=openblas USE_CUDA=1 USE_CUDA_PATH=/usr/local/cuda USE_CUDNN=1 USE_NCCL=1 USE_NCCL_PATH=/usr/local/nccl
cd python
sudo pip install -e .

follow install instruction from speech_recognition/Readme.md for LibriSpeech installation and creation of json files
mkdir checkpoints
mkdir logs

edit deepspeech.cfg to run on one GPU and correct training json file name

diff deepspeech.cfg deepspeech_orig.cfg 
22,23c22
< #context = gpu0,gpu1,gpu2
< context = gpu0
---
> context = gpu0,gpu1,gpu2
45c44
< train_json = ./train_corpus.json
---
> train_json = ./train_corpus_all.json



cd example/speech_recognition
edit train.py
diff train_orig.py train.py
26c26
< from tensorboard import SummaryWriter
---
> #from tensorboard import SummaryWriter
> import tensorflow as tf
136c136,137
<     summary_writer = SummaryWriter(tblog_dir)
---
>     #summary_writer = SummaryWriter(tblog_dir)
>     summary_writer = tf.summary.FileWriter(tblog_dir)

export CUDA_VISIBLE_DEVICES=0
tensorflow is needed for the summary.filewriter
export PYTHONPATH=~/tf_r19_92_714_py:$PYTHONPATH


python main.py --configfile deepspeech.cfg > deepspeech_v100_train_1.log 2>&1
the application spends about 15 minutes processing statistics on the input files through a large number
of python jobs executing on CPUs...they do go away eventually