Migrate structures to pandas DataFrame, Complete visualization, Minor fixes and improvements

classifier/feature_extraction/batch_feature_extractor.py

@@ -1,7 +1,8 @@
 from os import listdir
 from os.path import isfile, join
 import multiprocessing as mp
-from .feature_extractor import extractFeatures
+import pandas as pd
+from feature_extractor import extractFeatures
 
 class bcolors:
 	BLUE = '\033[94m'
@@ -13,18 +14,21 @@ class bcolors:
 def batchExtract(audioFilesPath, featureFilesPath, sampleRate):
 	audioFiles = [file for file in listdir(audioFilesPath) if isfile(join(audioFilesPath, file))]
 
-	# Without multithreading
-	# for file in audioFiles:
-	# 	extractFeatures(audioFilesPath + file,
-	# 		featureFilesPath + file[0:file.rfind('.')] + '.json', int(sampleRate))
+	dataframesList = [None]*len(audioFiles)
 
-	pool = mp.Pool(processes = 8)
-	for file in audioFiles:
-		pool.apply(extractFeatures,args=(audioFilesPath + file,
-			featureFilesPath + file[0:file.rfind('.')] + '.json',int(sampleRate)))
+	pool = mp.Pool()
+	for process, file in enumerate(audioFiles):
+		dataframesList[process] = pool.apply_async(extractFeatures,args=(audioFilesPath + file,
+			featureFilesPath + file[0:file.rfind('.')] + '.json',int(sampleRate))).get()
+	pool.close()
+	pool.join()
+
+	joinedDataset = pd.concat(dataframesList)
 
 	print('Batch feature extraction finished successfully.')
 
+	return joinedDataset
+
 # Prints a nice message to let the user know the module was imported
 print(bcolors.BLUE + 'batch_feature_extractor loaded' + bcolors.ENDC)
 

classifier/feature_extraction/feature_extractor.py

@@ -1,4 +1,6 @@
 import essentia
+import pandas as pd
+import numpy as np
 from essentia.standard import (MonoLoader, Windowing, Spectrum, MFCC,
 	ZeroCrossingRate, SpectralCentroidTime, RollOff, Flux, Envelope,
 	FlatnessSFX, LogAttackTime, StrongDecay, FlatnessDB, HFC,
@@ -95,6 +97,8 @@ def extractFeatures(audio, outputPath, sampleRate):
 
 	YamlOutput(filename = outputPath, format = 'json', writeVersion = False)(pool)
 
+	return pd.DataFrame(np.array([pool[i] for i in pool.descriptorNames()]).T, columns = pool.descriptorNames())
+
 def _4HzModulation(melEnergies, frameEnergy, sampleRate):
 	from scipy.signal import butter, sosfilt, sosfreqz
 	nyquist = 0.5 * sampleRate

classifier/pipeline.py

@@ -1,24 +1,26 @@
 import numpy as np
+import pandas as pd
 from feature_extraction.feature_extractor import extractFeatures
 from feature_extraction.batch_feature_extractor import batchExtract
-from preprocessing.data_preprocessing import arrayFromJSON, createSingleFeaturesArray, standardization, PCA
+from preprocessing.data_preprocessing import arrayFromJSON, standardization, PCA
 from training.model_training import simpleTrain, kFCrossValid
 
-batchExtract('../dataset/music_wav/', 'feature_extraction/music_features/', 22050)
-batchExtract('../dataset/speech_wav/', 'feature_extraction/speech_features/', 22050)
+musicFeatures = batchExtract('../dataset/music_wav/', 'feature_extraction/music_features/', 22050)
+musicFeatures = musicFeatures.assign(target=0)
+speechFeatures = batchExtract('../dataset/speech_wav/', 'feature_extraction/speech_features/', 22050)
+speechFeatures = speechFeatures.assign(target=1)
 
-dataset, target, featureKeys = createSingleFeaturesArray(
-	'feature_extraction/music_features/',
-	'feature_extraction/speech_features/')
+dataset = pd.concat([musicFeatures, speechFeatures])
+target = dataset.pop('target').values
 
 dataset = standardization(dataset)
 # dataset = PCA(dataset)
+
 print('Simple train accuracy achieved = ' + str(simpleTrain(dataset, target)))
 kFCrossValid(dataset, target, model = 'svm')
 clf = kFCrossValid(dataset, target, model = 'rndForest')
 
-extractFeatures('compined.wav', 'featuresStream/tmp.json', 22050)
-values = arrayFromJSON('featuresStream/tmp.json')[1]
-values = standardization(values)
-audioClass = clf.predict(values)
+features = extractFeatures('compined.wav', 'tmp.json', 22050)
+features = standardization(features)
+audioClass = clf.predict(features)
 print(audioClass)

classifier/preprocessing/data_preprocessing.py

@@ -1,6 +1,7 @@
 from os import listdir
 from os.path import isfile, join
 import numpy as np
+import pandas as pd
 import json
 
 class bcolors:
@@ -77,12 +78,10 @@ def PCA(dataset):
 	from sklearn.decomposition import PCA
 
 	print(bcolors.YELLOW + 'Running PCA' + bcolors.ENDC)
-	pca = PCA(n_components=15, svd_solver='full', whiten = True)
+	pca = PCA(n_components=10, svd_solver='full', whiten = True)
 	transformedDataset = pca.fit(dataset).transform(dataset)
 
-	# TODO: change the return value after the values of the parameters are decided
-	# and the feature selection is complete
-	return transformedDataset
+	return pca, transformedDataset
 
 # Prints a nice message to let the user know the module was imported
 print(bcolors.BLUE + 'feature_preprocessing loaded' + bcolors.ENDC)
@@ -91,9 +90,9 @@ print(bcolors.BLUE + 'feature_preprocessing loaded' + bcolors.ENDC)
 if __name__ == "__main__":
 	import sys
 	dataset, target, featureKeys = createSingleFeaturesArray(sys.argv[1], sys.argv[2])
-	newDataset = PCA(standardization(dataset))
+	scaledDataset = standardization(dataset)
 
-	print(bcolors.GREEN + 'Saving results to files' + bcolors.ENDC)
-	np.save('dataset.npy', newDataset)
-	np.save('target.npy', target)
-	np.save('featureKeys.npy', featureKeys)
+	print(bcolors.GREEN + 'Saving scaled results to file' + bcolors.ENDC)
+	datasetFrame = pd.DataFrame(scaledDataset, columns = featureKeys)
+	datasetFrame = datasetFrame.assign(target=target)
+	datasetFrame.to_pickle("./dataset.pkl")

classifier/test/accuracyWithoutFeature.py

@@ -1,55 +1,28 @@
 import numpy as np
+import pandas as pd
 from sys import path
 path.append('..')
 from feature_extraction.batch_feature_extractor import batchExtract
-from preprocessing.data_preprocessing import createSingleFeaturesArray, standardization
-from classification_model_training.model_training import simpleTrain
+from preprocessing.data_preprocessing import standardization
+from training.model_training import simpleTrain
 
-batchExtract('../../dataset/music_wav/', '../feature_extraction/music_features/', 22050)
-batchExtract('../../dataset/speech_wav/', '../feature_extraction/speech_features/', 22050)
+musicFeatures = batchExtract('../../dataset/music_wav/', '../feature_extraction/music_features/', 22050)
+musicFeatures = musicFeatures.assign(target=0)
+speechFeatures = batchExtract('../../dataset/speech_wav/', '../feature_extraction/speech_features/', 22050)
+speechFeatures = speechFeatures.assign(target=1)
 
-dataset, target, featureKeys = createSingleFeaturesArray(
-	'../feature_extraction/music_features/',
-	'../feature_extraction/speech_features/')
+dataset = pd.concat([musicFeatures, speechFeatures])
+target = dataset.pop('target').values
 
-dataset = standardization(dataset)
+dataset = pd.DataFrame(standardization(dataset), columns = dataset.columns.values)
 
 wholeAccuracy = simpleTrain(dataset, target, 'svm')
 print('Accuracy using whole dataset = ' + str(wholeAccuracy))
 
-damages = np.zeros(featureKeys.size)
+damages = np.zeros(dataset.columns.values.size)
 
-for index, key in enumerate(featureKeys):
-	acc = simpleTrain(np.delete(dataset, index, axis=1), target, 'svm')
+for index, key in enumerate(dataset.columns.values):
+	acc = simpleTrain(dataset.drop(key, axis=1), target, 'svm')
 	damages[index] = 100*(wholeAccuracy-acc)
 	print('Accuracy without ' + key + '\t= ' + str(acc) +
-		',\tdamage\t= ' + "%.2f" % damages[index] + '%')
-
-# Accuracy using whole dataset = 0.951902893127681
-# Accuracy without 4HzMod	= 0.9456968148215752,	damage	= 0.62%
-# Accuracy without Flat	= 0.9523592224148946,	damage	= -0.05%
-# Accuracy without HFC	= 0.9526330199872228,	damage	= -0.07%
-# Accuracy without LAtt	= 0.9524504882723374,	damage	= -0.05%
-# Accuracy without SC	= 0.9520854248425664,	damage	= -0.02%
-# Accuracy without SComp	= 0.948160992972529,	damage	= 0.37%
-# Accuracy without SDec	= 0.9520854248425664,	damage	= -0.02%
-# Accuracy without SEFlat	= 0.9513552979830245,	damage	= 0.05%
-# Accuracy without SF	= 0.9492561832618417,	damage	= 0.26%
-# Accuracy without SFlat	= 0.9496212466916126,	damage	= 0.23%
-# Accuracy without SLAtt	= 0.9498950442639409,	damage	= 0.20%
-# Accuracy without SR	= 0.9523592224148946,	damage	= -0.05%
-# Accuracy without SSDec	= 0.9519941589851236,	damage	= -0.01%
-# Accuracy without ZCR	= 0.9500775759788264,	damage	= 0.18%
-# Accuracy without mfcc0	= 0.9502601076937118,	damage	= 0.16%
-# Accuracy without mfcc1	= 0.9510815004106964,	damage	= 0.08%
-# Accuracy without mfcc10	= 0.9503513735511545,	damage	= 0.16%
-# Accuracy without mfcc11	= 0.9492561832618417,	damage	= 0.26%
-# Accuracy without mfcc12	= 0.9482522588299717,	damage	= 0.37%
-# Accuracy without mfcc2	= 0.9446928903897052,	damage	= 0.72%
-# Accuracy without mfcc3	= 0.9465182075385599,	damage	= 0.54%
-# Accuracy without mfcc4	= 0.9470658026832162,	damage	= 0.48%
-# Accuracy without mfcc5	= 0.9463356758236744,	damage	= 0.56%
-# Accuracy without mfcc6	= 0.9452404855343616,	damage	= 0.67%
-# Accuracy without mfcc7	= 0.9462444099662316,	damage	= 0.57%
-# Accuracy without mfcc8	= 0.9490736515469563,	damage	= 0.28%
-# Accuracy without mfcc9	= 0.9472483343981016,	damage	= 0.47%
+		',\tdamage\t= ' + "%.2f" % damages[index] + '%')

classifier/visualization/visualization.py

@@ -4,14 +4,31 @@ import matplotlib.pyplot as plt
 import seaborn as sns
 sns.set()
 
-datasetArray = np.load('../preprocessing/dataset.npy')
-target = np.load('../preprocessing/target.npy')
-featureKeysVector = np.load('../preprocessing/featureKeys.npy')
+dataset = pd.read_pickle('../preprocessing/dataset.pkl')
+target = dataset.pop('target')
 
-dataset = pd.DataFrame(datasetArray)
-dataset.columns = featureKeysVector
+# Feature evaluation
+from sklearn.ensemble import ExtraTreesClassifier
+from sklearn.feature_selection import SelectFromModel
 
-sns.relplot(x="4HzMod", y="Flat", data=dataset[["4HzMod", "Flat"]], hue = target, style = target)
-sns.jointplot(x="SLAtt", y="ZCR", data=dataset[["SLAtt", "ZCR"]]);
-sns.pairplot(data=dataset[["SDec", "Flat", "mfcc2", "HFC"]]);
-plt.show()
+clf = ExtraTreesClassifier(n_estimators=1000)
+clf = clf.fit(dataset, target)
+
+model = SelectFromModel(clf, prefit=True, max_features = 6)
+print('Retaining features:')
+print(dataset.columns.values[model.get_support()])
+reducedDataset = pd.DataFrame(model.transform(dataset),
+	columns = dataset.columns.values[model.get_support()])
+
+# Every combination of the 6 best features with length equal to 4 features
+import itertools
+featureCombinations = itertools.combinations(range(6), 4)
+
+for plotIndex, subset in enumerate(featureCombinations):
+	featurePlot = sns.pairplot(data=(reducedDataset.iloc[:, list(subset)]).assign(target=target),
+		hue='target', palette='Set1', vars=reducedDataset.columns.values[list(subset)]);
+	featurePlot.fig.savefig("output/figure_" + str(plotIndex+1) + ".png")
+
+# sns.relplot(x="4HzMod", y="Flat", data=dataset[["4HzMod", "Flat"]], hue = target, style = target)
+# sns.jointplot(x="SLAtt", y="ZCR", data=dataset[["SLAtt", "ZCR"]]);
+# plt.show()