reproduceMSRC.sh

#!/bin/bash

# The purpose is to reproduce SOA results on the MSRC data set.
#
# You need to do something like this:
#
#     export PYTHONPATH="/home/sherrahj/code/lib/slic-python:/home/sherrahj/code/alienMarkovNetworks/maxflow"
#
# Recommend low number of cores like 8 because there is nesting, parallel within
# parallel, and it uses too much ram.

dataPath="$1"; shift
outDir="$1"; shift
typeset -i nbCores="$1"; shift

if [ -z "$dataPath" -o -z "$outDir" ]; then
    echo "Usage Error:"
    echo "   reproduceMSRC.sh <MSRC data path> <output directory> <nbCores>"
    echo "The output dir must not exist.  The input directory contains"
    echo "the pre-generated train/val/test data sets, see createMSRCPartition.sh"
    echo "and README.md."
    exit 1
fi

set -e

# Steps are:
#   - extract features from images
#   - train classifier
#   - apply classifier to images
#   - apply MRF to classifications
#   - compute accuracy of classifier output, MRF output against GT

# 
# Parameters
# 
# In github see these issues:
#
#   https://github.com/RockStarCoders/alienMarkovNetworks/issues/27
#   https://github.com/RockStarCoders/alienMarkovNetworks/issues/10
#
slicN=400
slicC=10.0
K=0.2
ftrsBase="$outDir"/ftrsTrain
clfrFn="$outDir"/classifier.pkl
oclDir="$outDir"/classified
olabDir="$outDir"/labelled

if true; then 

# Create output directory.  It must not already exist.
mkdir "$outDir"

echo "*** Creating training features..."
./createFeatures.py --type=pkl \
    --nbSuperPixels=$slicN --superPixelCompactness=$slicC \
    "$dataPath/trainingPlusValidation" "$ftrsBase" --nbCores $nbCores
echo "  done"

if true; then
    # make special training only, validation only sets
    ./createFeatures.py --type=pkl \
	--nbSuperPixels=$slicN --superPixelCompactness=$slicC \
	"$dataPath/training" "${ftrsBase}-train" --nbCores $nbCores
    ./createFeatures.py --type=pkl \
	--nbSuperPixels=$slicN --superPixelCompactness=$slicC \
	"$dataPath/validation" "${ftrsBase}-validation" --nbCores $nbCores
fi

echo "*** Training classifier..."
./trainClassifier.py --type=randyforest \
    --paramSearchFolds=0 \
    --rf_n_estimators=500 \
    --rf_max_depth=None \
    --rf_max_features=75 \
    --rf_min_samples_leaf=5 \
    --rf_min_samples_split=10 \
    "${ftrsBase}_ftrs.pkl" "${ftrsBase}_labs.pkl" \
    --outfile="$clfrFn" --nbJobs $nbCores 
echo "  done"

echo "*** Classify test images..."
mkdir "$oclDir"
./classifyAllImages.sh "$clfrFn" "$oclDir" $slicN $slicC $nbCores \
    "$dataPath"/test/Images/*.bmp
echo "  done"


echo "*** Label test images..."
mkdir "$olabDir"
./labelAllImagesGivenProbs.sh $K "$olabDir" $nbCores \
    "$oclDir"/*.pkl
echo "  done"



echo "*** Evaluating results..."
echo "  * Classifier only: "
./evalPredictions.py "$oclDir"/evalpairs.csv "$dataPath/test" ''
echo "  * Classifier+MRF:  "
./evalPredictions.py "$olabDir"/evalpairs.csv "$dataPath/test" ''
echo "      done"

fi

#
# ... and Training
#

echo "*** Classify training images..."
mkdir "$oclDir"/training
./classifyAllImages.sh "$clfrFn" "$oclDir"/training $slicN $slicC $nbCores \
    "$dataPath"/training/Images/*.bmp
echo "*** Evaluating results..."
./evalPredictions.py "$oclDir"/training/evalpairs.csv "$dataPath/training" ''
echo "  done"

#
# Validation too
#

echo "*** Classify validation images..."
mkdir "$oclDir"/validation
./classifyAllImages.sh "$clfrFn" "$oclDir"/validation $slicN $slicC $nbCores \
    "$dataPath"/validation/Images/*.bmp
echo "*** Evaluating results..."
./evalPredictions.py "$oclDir"/validation/evalpairs.csv "$dataPath/validation" ''
echo "  done"