from matplotlib.animation import FuncAnimation
import librosa
import dataLoader
import audioProcessing
import config_
sr = 16000
loader = dataLoader.DCASE2dataPrep(SR=sr, dev=True, test=False)
print(len(loader.trainData))
print(len(loader.holdoutData))
minTimeStretch, maxTimeStretch, stepsTimeStretch, minPitchShift, maxPitchShift, stepsPitchShift, minLevelAdjust, maxLevelAdjust, stepsLevelAdjust = config_.getAugmentConfig(
config=5)
print('Augment')
aug = audioProcessing.augment(loader.trainData, sr, minTimeStretch,
maxTimeStretch, stepsTimeStretch, minPitchShift,
maxPitchShift, stepsPitchShift, minLevelAdjust,
maxLevelAdjust, stepsLevelAdjust)
nBin = 99
print('Prepare Train Set')
trainSet = loader.prepareTrainSet(aug,
deriv=False,
frameEnergy=False,
N_FFT=2048,
hop=512,
n_mels=nBin,
lengthScene=10,
eventsPerScene=5,
lengthExample=10,
hopExample=10,
asFeatureVector=False,
shuffle=False)
def main():
randGen = np.random.RandomState(1337)
baseFolder = 'ESC/ESC-US/'
signalRate = 30000
lengthScene = 10
eventsPerScene = 5
numberMelBins = 80
nfft = 2048
hop = 512
layer = 2
extractionLayer = 5
hidden = 80
lr = 0.002
sigma = 0.25
log_power = True
deriv = False
frameEnergy = False
batchSize = 128
epochs = 100000
asFeatureVector = False
temporalCompression = True
loader = dataLoader.DCASE2dataPrep(SR=signalRate, dev=False, test=False)
trainData = loader.trainData
#TODO: also get the augmented trainData
print('augment train data')
augmentedTrainData = audioProcessing.augment(trainData, signalRate, 0.9,
1.1, 2, -1, 1, 2, 0, 3, 4)
lengthExample = 11
hopExample = 11
#to compute a mean and std deviation for normalization
print('prepare train set')
trainSet = loader.prepareTrainSet(augmentedTrainData,
lengthScene=lengthScene,
lengthExample=lengthExample,
hopExample=hopExample,
eventsPerScene=eventsPerScene,
N_FFT=nfft,
hop=hop,
randSeed=1337,
log_power=log_power,
deriv=deriv,
frameEnergy=frameEnergy,
asFeatureVector=asFeatureVector,
temporalCompression=temporalCompression,
n_mels=numberMelBins,
predDelay=0)
val = trainSet[0][0:10000]
zNorm = (loader.trainMean, loader.trainStd)
fullStats = (loader.trainMean, loader.trainStd, loader.trainMin,
loader.trainMax)
modelName = 'ESC_PreTrainNew6'
#1 normal full
#2 lower lr
#3 no "noise" classes, Insects, rain, sea waves, crackling fire, crickets,
#chirping birds, water drops, wind, pouring water, toilet flush, thunderstorm, clapping,
#breathing, brushing teeth, drinking sipping, washing machine, vacuum cleaner,
#helicopter, chainsaw, engine, train, airplane, fireworks, hand saw
#4 no "noise" classes, also remove choughing, laughing, door knock, keyboard typing
#5 new training variant with loading full file and chopping subsets
#6 0.25 sigma dropout on input
modelsFolder = config_.modelsFolder
logFolder = config_.logFolder
statsFileName = modelsFolder + 'ST_' + modelName
print('Saving stats', statsFileName)
np.savez(statsFileName,
mean=fullStats[0],
std=fullStats[1],
minimum=fullStats[2],
maximum=fullStats[3])
lengthExample = 11
print('Create Generator')
generator = genBatchPreLoadWhole(baseFolder, batchSize, signalRate,
lengthExample, asFeatureVector,
temporalCompression, nfft, hop, log_power,
deriv, frameEnergy, numberMelBins, zNorm)
print('Create initial fill')
testExample = next(generator)
#dataX = loadAll(baseFolder,batchSize,signalRate,lengthExample,asFeatureVector,
# temporalCompression,nfft,hop,log_power,deriv,frameEnergy,numberMelBins,zNorm)
#testExample = dataX[0:1]
#print(dataX.shape)
inDim = testExample[0].shape #(1,lengthExample* (numberMelBins+1))
outDim = inDim
#model = modelLoader.createFeedForward_AE(layer,hidden,inDim,outDim,lr,sigma)
model = modelLoader.createSeqToSeq(layer, hidden, inDim, outDim, lr, sigma)
saveCallback = ModelCheckpoint(modelsFolder + modelName,
monitor='val_loss',
save_best_only=True)
breakCallback = TerminateOnNaN()
lrCallback = ReduceLROnPlateau(monitor='loss', factor=0.1, patience=100)
logCallback = CSVLogger(logFolder + modelName, append=True)
history = model.fit_generator(
generator,
steps_per_epoch=200,
epochs=epochs,
verbose=2,
max_queue_size=30,
callbacks=[saveCallback, breakCallback, lrCallback, logCallback],
validation_data=(val, val),
workers=0)
#history = model.fit(x=dataX,y=dataX,batch_size=batchSize,epochs=epochs,verbose=2,callbacks=[saveCallback,breakCallback,lrCallback,logCallback],validation_data=(val,val))
exit(0)
valData = reducedValData
loader.devData = valData
#"""
#Augmentation
signal_rate = loader.rate
minTimeStretch = 0.9
maxTimeStretch = 1.1
stepsTimeStretch = 0
minPitchShift = -1
maxPitchShift = 1.1
stepsPitchShift = 0
minLevelAdjust = 0.0
maxLevelAdjust = 2.0
stepsLevelAdjust = 3
augmentedTrainData = audioProcessing.augment(trainData,signal_rate,minTimeStretch,maxTimeStretch,stepsTimeStretch,minPitchShift,maxPitchShift,stepsPitchShift,minLevelAdjust,maxLevelAdjust,stepsLevelAdjust)
lengthScene = 15
eventsPerScene = 5
numberMelBins = 81
nfft = 2048
hop = 512
lengthExample=40 #40 should now be roughly a second input
hopExample=20
layer = 2
extractionLayer = 4
hidden = 80
lr = 0.002
sigma = 0.0
asFeatureVector=False
lengthExample = 1
hopExample = 1
asFeatureVector = True
#TODO: augment data
data = loader.trainData
minTime = 0.9
maxTime = 1.1
stepsTime = 0
minPitch = 0.9
maxPitch = 1.1
stepsPitch = 0
minLevel = 0.0
maxLevel = 2.0
stepsLevel = 3
augmentedData = audioProcessing.augment(data, loader.rate, minTime,
maxTime, stepsTime, minPitch,
maxPitch, stepsPitch, minLevel,
maxLevel, stepsLevel)
loader.estimateStatistics(augmentedData,
20,
deriv=False,
frameEnergy=False,
n_mels=25)
datasplits = loader.splitDataByLabel(augmentedData, loader.labels)
genExample = loader.genTrainSetExample(datasplits[0],
lengthExample=lengthExample,
hopExample=hopExample,
asFeatureVector=asFeatureVector,
deriv=False,
frameEnergy=False,
n_mels=25,
N_FFT=2048,
if __name__ == "__main__":
loader = DCASE2dataPrep(SR=30000)
normalTrainData = loader.trainData
signal_rate = loader.rate
minTimeStretch = 0.9
maxTimeStretch = 1.1
stepsTimeStretch = 2
minPitchShift = -2.
maxPitchShift = 2.
stepsPitchShift = 2
minLevelAdjust = 0.0
maxLevelAdjust = 2.0
stepsLevelAdjust = 3
augmentedTrainData = augment(normalTrainData,signal_rate,minTimeStretch,maxTimeStretch,stepsTimeStretch,minPitchShift,maxPitchShift,stepsPitchShift,minLevelAdjust,maxLevelAdjust,stepsLevelAdjust)
lengthScene = 10
eventsPerScene = 5
numberMelBins = 80
nfft=2048
hop=512
lenExample = 10
hopExample = 10
trainSet = loader.prepareTrainSet(augmentedTrainData,lengthScene=lengthScene,lengthExample=lenExample,hopExample=hopExample,
eventsPerScene=eventsPerScene,N_FFT=nfft,hop=hop,randSeed=1337,
log_power=True,deriv=False,frameEnergy=False,asFeatureVector=False,temporalCompression=True,
n_mels=numberMelBins,predDelay=0,zNorm=True)
validationSet = loader.prepareValSet(lengthScene=lengthScene,lengthExample=lenExample,hopExample=hopExample,eventsPerScene=eventsPerScene,
N_FFT=nfft,hop=hop,randSeed=1337,log_power=True,asFeatureVector=False,temporalCompression=True,
deriv=False,frameEnergy=False,n_mels=numberMelBins,