Update the documentation about running DeepQuest w/ Tensorflow as backend

[fredblain]

Installation requirement: conda install tensorflow (or tensorflow-gpu for GPU support)
For training, KERAS_BACKEND=tensorflow <training-script> <args> should be enough.

Tested with BiRNN models, similar results observed so far, but further tests need to be conducted, to make sure that there is no side-effect.

[zouharvi]

AFAIK the ready-made scripts (train-test-*) expect the backend to be Theano, since they manipulate the standard output from this backend with the commands:

awk '/^$/ {nlstack=nlstack "\n";next;} {printf "%s",nlstack; nlstack=""; print;}' log-${model_name}-prep.txt > log-${model_name}.txt
best_epoch=$(tail -1 log-${model_name}.txt | tr ":" "\n" | tr ' ' '\n' | tail -3 | head -1)

It would be greatly beneficial if there was at least a side note in the tutorial section about this breaking with Tensorflow.

[fredblain]

The documentation will be updated along with the code-base within the next few weeks.
Meanwhile, one can retrieve the best epoch as follows, when using Tensorflow as backend:
best_epoch=$(grep "Best pearson found at epoch" ${wdir}/log-${model_name}.txt | tr ":" "\n" | tr ' ' '\n' | tail -3 | head -1)

[mzeidhassan]

Any plans to update documentation

It still reads:

export KERAS_BACKEND=theano
export MKL_THREADING_LAYER=GNU
THEANO_FLAGS=device={device_name} python main.py | tee -a /tmp/deepQuest.log 2>&1 &