def spectrograms(input_type,
data_list,
labelFile,
savePath,
fft_size,
win_size,
hop_size,
duration,
data_window=100,
window_shift=10,
augment=True,
save=True,
minimum_length=1):
from audio import compute_spectrogram
spectrograms = list()
labels = list()
print('Computing the ' + input_type + ' spectrograms !!')
with open(data_list, 'r') as f:
spectrograms = [
compute_spectrogram(input_type, file.strip(), fft_size, win_size,
hop_size, duration, augment, minimum_length)
for file in f
]
# Get the labels into a list and save it along with the spectrograms
with open(labelFile, 'r') as f:
#labels = [1 if line.strip().split(' ')[1] == 'genuine' else 0 for line in f]
labels = [line.strip() for line in f]
if augment:
new_data = list()
new_labels = list()
assert (len(labels) == len(spectrograms))
print(
'Now performing augmentation using sliding window mechanism on original spectrogram .... '
)
for i in range(len(spectrograms)):
d, l = augment_data(spectrograms[i], labels[i], data_window,
input_type, window_shift)
new_data.extend(
d) # extend the list rather than adding it into a new list
new_labels.extend(l)
spectrograms = new_data
labels = new_labels
if save:
from helper import makeDirectory
makeDirectory(savePath)
outfile = savePath + '/spec'
with open(outfile, 'w') as f:
np.savez(outfile, spectrograms=spectrograms, labels=labels)
print('Finished computing spectrogram and saved inside: ', savePath)
def run_prediction(model_path,
featType,
dataType,
protocal,
inputPath,
mean_std_file,
outBase,
batch_size=100,
activation='elu',
init_type='xavier',
targets=2,
fftSize=256,
architecture=2,
duration=1,
padding=True,
n_model=None,
inputType='mag_spec',
augment=True):
# Extract Features from Training set
#print('Extracting ' + featType + ' for the ' + dataType + 'set')
print('outBase in run_prediction is: ', outBase)
data, lab = dataset.load_data(inputPath + dataType + '/')
#data = dataset.normalise_data(data,mean_std_file,'utterance')
data = dataset.normalise_data(data, mean_std_file, 'global_mv')
labels = dataset.get_labels_according_to_targets(lab, targets)
featureList = getFeatures(featType, inputType, data, labels, batch_size,
model_path, n_model, activation, init_type,
targets, fftSize, padding, architecture,
duration, augment)
if featType == 'bottleneck':
makeDirectory(outBase + '/features/') # dataType)
saveFeatures(featureList, outBase + '/features/' +
dataType) #saves as train.npz, dev.npz etc
elif featType == 'scores':
makeDirectory(outBase + '/predictions/')
write_scores_to_file(featureList,
outfile=outBase + '/predictions/' +
str(dataType) + '_prediction.txt')
else:
print('PLEASE CHOSE CORRECT PARAM !!')
def trainCNN_on_Bulbul_architecture():
#CNN Training parameters
activation = 'elu'
init_type = 'xavier'
batch_size = 32
epochs = 200 #0
# Regularizer parameters
use_lr_decay = False #set this flag for LR decay
wDecayFlag = False #whether to perform L2 weight decay or not
lossPenalty = 0.001 # Using lambda=0.001 .
applyBatchNorm = False
deviceId = "/gpu:0"
# Adam parameters
optimizer_type = 'adam'
b1 = 0.9
b2 = 0.999
epsilon = 0.1
momentum = 0.95
dropout1 = 0.6 #for input to first FC layer
dropout3 = 0.5 #for intermediate layer input
dropout2 = [0.5, 0.6]
lambdas = [0.0005, 0.001]
architectures = [
2
] # birds architecture 1, to make it unified (check model.py for definition)
trainingSize = [1] #in seconds
#lr= 0.0005
learning_rates = [0.0005, 0.0003, 0.0001, 0.005]
targets = 2
fftSize = 256
specType = 'mag_spec'
padding = True
# Used following paths since I moved the scripts in git and used link so that code are synchronised
spectrogramPath = '/homes/bc305/myphd/stage2/deeplearning.experiment1/spectrograms/'
tensorboardPath = '/homes/bc305/myphd/stage2/deeplearning.experiment1/CNN3/tensorflow_log_dir/'
modelPath = '/homes/bc305/myphd/stage2/deeplearning.experiment1/CNN3/models/'
for duration in trainingSize:
print('Now loading the data !!')
outPath = spectrogramPath + specType + '/' + str(
fftSize) + 'FFT/' + str(duration) + 'sec/'
mean_std_file = outPath + 'train/mean_std.npz'
# Load training data, labels and perform norm
tD = dataset.load_data(outPath + 'train/')
tL = dataset.get_labels_according_to_targets(trainP, targets)
if not os.path.exists(mean_std_file):
print('Computing Mean_std file ..')
dataset.compute_global_norm(tD, mean_std_file)
print('Shape of labels: ', tL.shape)
#tD = dataset.normalise_data(tD,mean_std_file,'utterance') # utterance level
tD = dataset.normalise_data(tD, mean_std_file, 'global_mv') # global
# Load dev data, labels and perform norm
devD = dataset.load_data(outPath + 'dev/')
devL = dataset.get_labels_according_to_targets(devP, targets)
#devD = dataset.normalise_data(devD,mean_std_file,'utterance')
devD = dataset.normalise_data(devD, mean_std_file, 'global_mv')
### We are training on TRAIN set and validating on DEV set
t_data = tD
t_labels = tL
v_data = devD
v_labels = devL
for dropout in dropout2:
architecture = architectures[0]
for lr in learning_rates:
hyp_str = '_cnnModel' + str(
architecture) + '_keepProb_0.6_' + str(dropout) + str(
dropout3) + 'lr' + str(lr)
log_dir = tensorboardPath + '/birdsArch_max2000epochsTEMP/' + hyp_str
model_save_path = modelPath + '/birdsArch_max2000epochsTEMP/' + hyp_str
logfile = model_save_path + '/training.log'
figDirectory = model_save_path
makeDirectory(model_save_path)
print('Training model with ' + str(duration) +
' sec data and cnnModel' + str(architecture))
tLoss, vLoss, tAcc, vAcc = model.train(
architecture, fftSize, padding, duration, t_data, t_labels,
v_data, v_labels, activation, lr, use_lr_decay, epsilon,
b1, b2, momentum, optimizer_type, dropout1, dropout,
dropout3, model_save_path, log_dir, logfile, wDecayFlag,
lossPenalty, applyBatchNorm, init_type, epochs, batch_size,
targets)
#plot_2dGraph('#Epochs', 'Avg CE Loss', tLoss,vLoss,'train_ce','val_ce', figDirectory+'/loss.png')
#plot_2dGraph('#Epochs', 'Avg accuracy', tAcc,vAcc,'train_acc','val_acc',figDirectory+'/acc.png')
plot_2dGraph('#Epochs', 'Val loss and accuracy', vLoss, vAcc,
'val_loss', 'val_acc',
figDirectory + '/v_ls_acc.png')
def trainCNN_on_trainData():
#CNN Training parameters
activation = 'mfm' #choose activation: mfm,elu, relu, mfsoftmax, tanh ?
init_type='xavier' #'truncated_normal' #'xavier' #or 'truncated_normal'
batch_size = 32
epochs = 120 #1000
# Regularizer parameters
wDecayFlag = False #whether to perform L2 weight decay or not
lossPenalty = 0.001 # Using lambda=0.001 .
applyBatchNorm = False
deviceId = "/gpu:0"
# Adam parameters
optimizer_type = 'adam'
b1=0.9
b2=0.999
epsilon=0.1
momentum=0.95
dropout1=0.1 #for input to first FC layer
dropout2=0.2 #for intermediate layer input
drops=[0.4] # 50% dropout the inputs of FC layers
lambdas = [0.0005, 0.001]
architectures = [1]
trainingSize = [1] #in seconds
use_lr_decay=True #set this flag for LR decay
learning_rates=[0.0008,0.0006,0.0004,0.0001]
#learning_rates=[0.0001, 0.001] #0.0001
#learning_rates=np.random.uniform(0.003,0.0005,10)
#learning_rates=np.random.uniform(0.01,0.0001,10)
#learning_rates=np.linspace(0.001,0.01,10)
#learning_rates=[0.0002, 0.0003, 0.0004, 0.0005, 0.0006,0.0007,0.0008, 0.0009]
targets=2
#specType='mag_spec' #lets try loading mag_spec
#inputTypes=['mel_spec'] #'mag_spec' #'cqt_spec' ## Running on Hepworth !
inputTypes=['mag_spec'] # Not Run yet
#inputTypes=['cqt_spec'] # Not Run yet
#inputTypes=['mag_spec','cqt_spec','mel_spec']
padding=True
augment = True
trainPercentage=1.0 #Each training epoch will see only 50% of the original data at random !
valPercentage=0.8
if augment:
spectrogramPath='/homes/bc305/myphd/stage2/deeplearning.experiment1/spectrograms_augmented/1sec_shift/'
else:
spectrogramPath='/homes/bc305/myphd/stage2/deeplearning.experiment1/spectrograms/'
# Used following paths since I moved the scripts in git and used link so that code are synchronised
tensorboardPath = '/homes/bc305/myphd/stage2/deeplearning.experiment1/CNN3/tensorflow_log_dir/'
modelPath = '/homes/bc305/myphd/stage2/deeplearning.experiment1/CNN3/models_augmented/'
#for duration in trainingSize:
duration=1
fftSize=512
for specType in inputTypes:
#print('Now loading the data with FFT size: ', fftSize)
outPath = spectrogramPath +specType + '/' +str(fftSize)+ 'FFT/' + str(duration)+ 'sec/'
mean_std_file = outPath+'train/mean_std.npz'
# Load training data, labels and perform norm
tD,tL = dataset.load_data(outPath+'train/')
tL = dataset.get_labels_according_to_targets(tL, targets)
assert(len(tD)==len(tL))
if not os.path.exists(mean_std_file):
print('Computing Mean_std file ..')
dataset.compute_global_norm(tD,mean_std_file)
# We will try utterance based norm later
#tD = dataset.normalise_data(tD,mean_std_file,'utterance') # utterance level
tD = dataset.normalise_data(tD,mean_std_file,'global_mv') # global
# Now take only 80% of the new augmented data to use for validation
# Just to save some time during training
devD,devL = dataset.get_random_data(outPath+'dev/',batch_size,valPercentage)
devL = dataset.get_labels_according_to_targets(devL, targets)
assert(len(devD)==len(devL))
#devD = dataset.normalise_data(devD,mean_std_file,'utterance')
devD = dataset.normalise_data(devD,mean_std_file,'global_mv')
### We are training on TRAIN set and validating on DEV set
t_data = tD
t_labels = tL
v_data = devD
v_labels = devL
print('Training model on ', specType)
for dropout in drops: # dropout 1.0 and 0.5 to all inputs of DNN
architecture = architectures[0]
penalty=0.001 #this is not used thought at the moment
for lr in learning_rates:
#for targets in target_list:
hyp_str='keep_'+str(dropout1)+'_'+str(dropout2)+'_'+str(dropout)+'_'+str(specType)+'_Lr'+str(lr)
log_dir = tensorboardPath+ '/model1_max120epochs_32batch_with_0.85Decay/' +str(specType)+ '/' + hyp_str
model_save_path = modelPath + '/model1_max120epochs_32batch_with_0.85Decay/'+str(specType) + '/' + hyp_str
logfile = model_save_path+'/training.log'
figDirectory = model_save_path
makeDirectory(model_save_path)
tLoss,vLoss,tAcc,vAcc=model.train(specType,architecture,fftSize,padding,duration,t_data,t_labels,
v_data,v_labels,activation,lr,use_lr_decay,epsilon,b1,b2,momentum,
optimizer_type,dropout1,dropout2,dropout,model_save_path,log_dir,
logfile,wDecayFlag,penalty,applyBatchNorm,init_type,epochs,batch_size,
targets,augment)#,trainPercentage,valPercentage)
def trainCNN_on_trainData():
#CNN Training parameters
activation = 'mfm' #choose activation: mfm,elu, relu, mfsoftmax, tanh ?
init_type='xavier' #'truncated_normal' #'xavier' #or 'truncated_normal'
batch_size = 32
epochs = 2000
# Regularizer parameters
use_lr_decay=False #set this flag for LR decay
wDecayFlag = False #whether to perform L2 weight decay or not
lossPenalty = 0.001 # Using lambda=0.001 .
applyBatchNorm = False
deviceId = "/gpu:0"
# Adam parameters
optimizer_type = 'adam'
b1=0.9
b2=0.999
epsilon=0.1
momentum=0.95
#dropout1=1.0 #for input to first FC layer
#dropout2=1.0 #for intermediate layer input
drops=[1.0] # No dropout
lambdas = [0.001]
architectures = [1]
trainingSize = [1] #in seconds
lr= 0.0001
targets=2
#specType='mag_spec' #lets try loading mag_spec
#inputTypes=['mel_spec'] #'mag_spec' #'cqt_spec' ## Running on Hepworth !
#inputTypes=['mag_spec'] # Not Run yet
#inputTypes=['cqt_spec'] # Not Run yet
inputTypes=['mel_spec','mag_spec','cqt_spec'] #CQT may throw error during scoring in model.py
padding=True
augment = True
trainPercentage=0.8 #Each training epoch will see only 80% of the original data at random !
valPercentage=0.3 #30% only used at random for validation
if augment:
spectrogramPath='/homes/bc305/myphd/stage2/deeplearning.experiment1/spectrograms_augmented/1sec_shift/'
else:
spectrogramPath='/homes/bc305/myphd/stage2/deeplearning.experiment1/spectrograms/'
# Used following paths since I moved the scripts in git and used link so that code are synchronised
tensorboardPath = '/homes/bc305/myphd/stage2/deeplearning.experiment1/CNN3/tensorflow_log_dir/'
modelPath = '/homes/bc305/myphd/stage2/deeplearning.experiment1/CNN3/models_augmented/'
#for duration in trainingSize:
duration=1
for specType in inputTypes:
if specType == 'mel_spec' or specType == 'cqt_spec':
fftSize=512
else:
fftSize=256
print('Now loading the data with FFT size: ', fftSize)
outPath = spectrogramPath +specType + '/' +str(fftSize)+ 'FFT/' + str(duration)+ 'sec/'
mean_std_file = outPath+'train/mean_std.npz'
# Load training data, labels and perform norm
tD,tL = dataset.load_data(outPath+'train/')
tL = dataset.get_labels_according_to_targets(tL, targets)
assert(len(tD)==len(tL))
if not os.path.exists(mean_std_file):
print('Computing Mean_std file ..')
dataset.compute_global_norm(tD,mean_std_file)
tD = dataset.normalise_data(tD,mean_std_file,'utterance') # utterance level
tD = dataset.normalise_data(tD,mean_std_file,'global_mv') # global
# Load dev data, labels and perform norm
devD,devL = dataset.load_data(outPath+'dev/')
devL = dataset.get_labels_according_to_targets(devL, targets)
assert(len(devD)==len(devL))
devD = dataset.normalise_data(devD,mean_std_file,'utterance')
devD = dataset.normalise_data(devD,mean_std_file,'global_mv')
### We are training on TRAIN set and validating on DEV set
t_data = tD
t_labels = tL
v_data = devD
v_labels = devL
print('Training model on ', specType)
for dropout in drops: # dropout 1.0 and 0.5 to all inputs of DNN
architecture = architectures[0]
#penalty=0.001 #this is not used thought at the moment
for penalty in lambdas:
#for targets in target_list:
hyp_str='arch'+str(architecture)+'_keep'+str(dropout)+'_'+str(specType)+'_targets'+str(targets)
log_dir = tensorboardPath+ '/model1_max2000epochs_1/'+ hyp_str
model_save_path = modelPath + '/model1_max2000epochs_1/'+ hyp_str
logfile = model_save_path+'/training.log'
figDirectory = model_save_path
makeDirectory(model_save_path)
tLoss,vLoss,tAcc,vAcc=model.train(specType,architecture,fftSize,padding,duration,t_data,t_labels,
v_data,v_labels,activation,lr,use_lr_decay,epsilon,b1,b2,momentum,
optimizer_type,dropout,dropout,dropout,model_save_path,log_dir,
logfile,wDecayFlag,penalty,applyBatchNorm,init_type,epochs,batch_size,
targets,augment,trainPercentage,valPercentage)
#plot_2dGraph('#Epochs', 'Avg CE Loss', tLoss,vLoss,'train_ce','val_ce', figDirectory+'/loss.png')
#plot_2dGraph('#Epochs', 'Avg accuracy', tAcc,vAcc,'train_acc','val_acc',figDirectory+'/acc.png')
plot_2dGraph('#Epochs', 'Val loss and accuracy', vLoss,vAcc,'val_loss','val_acc',figDirectory+'/v_ls_acc.png')
def trainCNN_on_handcrafted_features():
#CNN Training parameters
activation = 'mfm' #choose activation: mfm,elu, relu, mfsoftmax, tanh ?
init_type = 'xavier' #'truncated_normal' #'xavier' #or 'truncated_normal'
batch_size = 32
epochs = 200 #500
# Regularizer parameters
wDecayFlag = False #whether to perform L2 weight decay or not
lossPenalty = 0.001 # Using lambda=0.001 .
applyBatchNorm = False
deviceId = "/gpu:0"
# Adam parameters
optimizer_type = 'adam'
b1 = 0.9
b2 = 0.999
epsilon = 0.1
momentum = 0.95 #0.95
dropout1 = 0.5 #0.3 #for input to first FC layer
dropout2 = 0.5 #0.3 #for intermediate layer input
drops = [0.5] # 50% dropout the inputs of FC layers
lambdas = [0.0005, 0.001]
targets = 2
architectures = [1]
trainingSize = [1] #in seconds
use_lr_decay = True
learning_rates = [0.0022, 0.0008, 0.005, 0.0004]
#learning_rates=np.random.uniform(0.003,0.0005,10) #Take randomly drawn lr from uniform distribution between 0.001-0.0001
#learning_rates=np.random.uniform(0.01,0.0001,10)
# Note: LR =0.01 explodes badly. We get the huge cross entropy values we used to get before.
#inputTypes=['CQCC','LFCC','LPCC','MFCC','RFCC'] #'IMFCC' # i ran this in 1 kapoor
#inputTypes=['IMFCC', 'SCMC'] # this was running on kapoor 0
inputTypes = ['RFCC', 'LFCC', 'SCMC',
'IMFCC'] # this is running on kapoor 1
padding = True
augment = True
trainPercentage = 1.0 #Each training epoch will see only 50% of the original data at random !
valPercentage = 1.0
if augment:
spectrogramPath = '/homes/bc305/myphd/stage2/deeplearning.experiment1/features_1sec_shift/'
else:
spectrogramPath = '/homes/bc305/myphd/stage2/deeplearning.experiment1/features/'
# Used following paths since I moved the scripts in git and used link so that code are synchronised
tensorboardPath = '/homes/bc305/myphd/stage2/deeplearning.experiment1/CNN3/tensorflow_log_dir/'
modelPath = '/homes/bc305/myphd/stage2/deeplearning.experiment1/CNN3/models_augmented/'
duration = 1
fftSize = 512
for specType in inputTypes:
outPath = spectrogramPath + specType + '/'
mean_std_file = outPath + 'train/mean_std.npz'
# Load training data, labels and perform norm
tD, tL = dataset.load_data(outPath + 'train/')
tL = dataset.get_labels_according_to_targets(tL, targets)
assert (len(tD) == len(tL))
if not os.path.exists(mean_std_file):
print('Computing Mean_std file ..')
dataset.compute_global_norm(tD, mean_std_file)
# We will try utterance based norm later
#tD = dataset.normalise_data(tD,mean_std_file,'utterance') # utterance level
tD = dataset.normalise_data(tD, mean_std_file, 'global_mv') # global
# Now take only 80% of the new augmented data to use for validation
# Just to save some time during training
#devD,devL = dataset.get_random_data(outPath+'dev/',batch_size,valPercentage)
devD, devL = dataset.load_data(outPath + 'dev/')
devL = dataset.get_labels_according_to_targets(devL, targets)
assert (len(devD) == len(devL))
#devD = dataset.normalise_data(devD,mean_std_file,'utterance')
devD = dataset.normalise_data(devD, mean_std_file, 'global_mv')
### We are training on TRAIN set and validating on DEV set
t_data = tD
t_labels = tL
v_data = devD
v_labels = devL
print('Training model on ', specType)
for dropout in drops: # Just dropout 1.0 and 0.5 to all inputs of DNN
architecture = architectures[0]
penalty = 0.001
for lr in learning_rates:
hyp_str = 'keep_' + str(dropout1) + '_' + str(
dropout2) + '_' + str(dropout) + '_' + str(
specType) + '_lr_' + str(lr)
print('Hyper-parameter string is: ', hyp_str)
log_dir = tensorboardPath + '/model1_120max_handcrafted_with_0.85Decay/' + str(
specType) + 'drops' + '/' + hyp_str
model_save_path = modelPath + '/model1_120max_handcrafted_with_0.85Decay/' + str(
specType) + 'drops' + '/' + hyp_str
logfile = model_save_path + '/training.log'
figDirectory = model_save_path
makeDirectory(model_save_path)
tLoss, vLoss, tAcc, vAcc = model.train(
specType, architecture, fftSize, padding, duration, t_data,
t_labels, v_data, v_labels, activation, lr, use_lr_decay,
epsilon, b1, b2, momentum, optimizer_type, dropout1,
dropout2, dropout, model_save_path, log_dir, logfile,
wDecayFlag, penalty, applyBatchNorm, init_type, epochs,
batch_size, targets,
augment) #,trainPercentage,valPercentage)
def trainCNN_on_Sparrow_architecture():
#CNN Training parameters
activation = 'elu' #'elu'
init_type = 'xavier'
batch_size = 32
epochs = 1000
# Regularizer parameters
use_lr_decay = False #set this flag for LR decay
wDecayFlag = False #whether to perform L2 weight decay or not
lossPenalty = 0.001 # Using lambda=0.001 .
applyBatchNorm = False
deviceId = "/gpu:0"
# Adam parameters
optimizer_type = 'adam'
b1 = 0.9
b2 = 0.999
epsilon = 0.1 #1e-08 is the default
momentum = 0.95
dropout1 = 1.0 #for input to first FC layer
dropout2 = 1.0 #for intermediate layer input
dropouts = [0.5, 0.4, 0.3, 0.2, 0.1] #,0.6]
lambdas = [0.0005, 0.001]
architectures = [
3
] # birds architecture sparrow, to make it unified (check model.py for definition)
trainingSize = [1] #in seconds
learning_rates = [0.0001, 0.00008]
targets = 2
fftSize = 256
specType = 'mag_spec'
padding = False
for duration in trainingSize:
print('Now loading the data !!')
outPath = '../../spectrograms/' + specType + '/' + str(
fftSize) + 'FFT/' + str(duration) + 'sec/'
mean_std_file = outPath + 'train/mean_std.npz'
# Load training data, labels and perform norm
tD = dataset.load_data(outPath + 'train/')
tL = dataset.get_labels_according_to_targets(trainP, targets)
dataset.compute_global_norm(tD, mean_std_file)
print('Shape of labels: ', tL.shape)
#tD = dataset.normalise_data(tD,mean_std_file,'utterance') # utterance level
tD = dataset.normalise_data(tD, mean_std_file, 'global_mv') # global
#print('Norm td: max and min are ', np.max(tD))
# Load dev data, labels and perform norm
devD = dataset.load_data(outPath + 'dev/')
devL = dataset.get_labels_according_to_targets(devP, targets)
#devD = dataset.normalise_data(devD,mean_std_file,'utterance')
#print('first Norm dev: max and min are ', np.max(devD))
devD = dataset.normalise_data(devD, mean_std_file, 'global_mv')
#print('Norm dev: max and min are ', np.max(devD))
trainSize = str(
duration) + 'sec' ##may be change this in model.py also !
### We are training on TRAIN set and validating on DEV set
t_data = tD
t_labels = tL
v_data = devD
v_labels = devL
for dropout in dropouts:
architecture = architectures[0]
for lr in learning_rates:
#hyp_str ='cnn'+str(architecture)+'_keepProb_1.0_' + str(dropout)+str(dropout3)+'lr'+str(lr)
hyp_str = 'sparrow' + '_keep_' + str(
dropout) + '_' + 'lr' + str(lr) + '_' + str(
activation) + '_' + 'fft' + str(fftSize)
log_dir = '../tensorflow_log_dir/sparrowArch/' + hyp_str
model_save_path = '../models/sparrowArch/' + hyp_str
logfile = model_save_path + '/training.log'
figDirectory = model_save_path
makeDirectory(model_save_path)
print('Training model with ' + str(duration) +
' sec data and cnnModel' + str(architecture))
tLoss, vLoss, tAcc, vAcc = model.train(
architecture, fftSize, padding, trainSize, t_data,
t_labels, v_data, v_labels, activation, lr, use_lr_decay,
epsilon, b1, b2, momentum, optimizer_type, dropout,
dropout1, dropout2, model_save_path, log_dir, logfile,
wDecayFlag, lossPenalty, applyBatchNorm, init_type, epochs,
batch_size, targets)
#plot_2dGraph('#Epochs', 'Avg CE Loss', tLoss,vLoss,'train_ce','val_ce', figDirectory+'/loss.png')
#plot_2dGraph('#Epochs', 'Avg accuracy', tAcc,vAcc,'train_acc','val_acc',figDirectory+'/acc.png')
plot_2dGraph('#Epochs', 'Val loss and accuracy', vLoss, vAcc,
'val_loss', 'val_acc',
figDirectory + '/v_ls_acc.png')
def trainCNN_on_trainData():
#CNN Training parameters
activation = 'mfm' #'elu'
init_type = 'xavier'
batch_size = 32
epochs = 2000 #0
# Regularizer parameters
use_lr_decay = False #set this flag for LR decay
wDecayFlag = False #whether to perform L2 weight decay or not
lossPenalty = 0.001 # Using lambda=0.001 .
applyBatchNorm = False
deviceId = "/gpu:0"
# Adam parameters
optimizer_type = 'adam'
b1 = 0.9
b2 = 0.999
epsilon = 0.1
momentum = 0.95
dropout3 = 1.0 #In russian arch there are only FC1 and output layer, so this not needed.
#dropout1=[0.3,0.2,0.1] # we ran this originally ! we will look into this later
#dropout2=[0.7,0.5,0.4,0.3,0.2] # for dropout1=0.3, we ran all combination
dropout1 = [0.2, 0.1]
dropout2 = [0.4, 0.2]
#lambdas = [0.0005, 0.001]
#dropout1=[1.0]
#dropout2=[1.0]
#dropout3=1.0
architectures = [5] # Russian Architecture is 5
trainingSize = [4] #in seconds
lr = 0.0001
targets = 2
fftSize = 2048
specType = 'mag_spec'
padding = True
# Used following paths since I moved the scripts in git and used link so that code are synchronised
spectrogramPath = '/homes/bc305/myphd/stage2/deeplearning.experiment1/spectrograms/'
tensorboardPath = '/homes/bc305/myphd/stage2/deeplearning.experiment1/CNN3/tensorflow_log_dir/'
modelPath = '/homes/bc305/myphd/stage2/deeplearning.experiment1/CNN3/models/'
for duration in trainingSize:
print('Now loading the data !!')
outPath = spectrogramPath + specType + '/' + str(
fftSize) + 'FFT/' + str(duration) + 'sec/'
mean_std_file = outPath + 'train/mean_std.npz'
# Load training data, labels and perform norm
tD = dataset.load_data(outPath + 'train/')
tL = dataset.get_labels_according_to_targets(trainP, targets)
if not os.path.exists(mean_std_file):
print('Computing Mean_std file ..')
dataset.compute_global_norm(tD, mean_std_file)
print('Shape of labels: ', tL.shape)
#tD = dataset.normalise_data(tD,mean_std_file,'utterance') # utterance level
tD = dataset.normalise_data(tD, mean_std_file, 'global_mv') # global
# Load dev data, labels and perform norm
devD = dataset.load_data(outPath + 'dev/')
devL = dataset.get_labels_according_to_targets(devP, targets)
#devD = dataset.normalise_data(devD,mean_std_file,'utterance')
devD = dataset.normalise_data(devD, mean_std_file, 'global_mv')
### We are training on TRAIN set and validating on DEV set
t_data = tD
t_labels = tL
v_data = devD
v_labels = devL
for dropout in dropout1:
architecture = architectures[0]
for drop in dropout2:
hyp_str = 'arch' + str(architecture) + '_keep_' + str(
dropout) + '_' + str(drop) + '_' + str(duration) + 'sec'
log_dir = tensorboardPath + '/rusCNN_max2000epochs/' + hyp_str
model_save_path = modelPath + '/rusCNN_max2000epochs/' + hyp_str
logfile = model_save_path + '/training.log'
figDirectory = model_save_path
makeDirectory(model_save_path)
print('Training model with ' + str(duration) +
' sec data and cnnModel' + str(architecture))
tLoss, vLoss, tAcc, vAcc = model.train(
architecture, fftSize, padding, duration, t_data, t_labels,
v_data, v_labels, activation, lr, use_lr_decay, epsilon,
b1, b2, momentum, optimizer_type, dropout, drop, dropout3,
model_save_path, log_dir, logfile, wDecayFlag, lossPenalty,
applyBatchNorm, init_type, epochs, batch_size, targets)
#plot_2dGraph('#Epochs', 'Avg CE Loss', tLoss,vLoss,'train_ce','val_ce', figDirectory+'/loss.png')
#plot_2dGraph('#Epochs', 'Avg accuracy', tAcc,vAcc,'train_acc','val_acc',figDirectory+'/acc.png')
plot_2dGraph('#Epochs', 'Val loss and accuracy', vLoss, vAcc,
'val_loss', 'val_acc',
figDirectory + '/v_ls_acc.png')
def other_features(input_type,
data_list,
labelFile,
savePath,
fft_size,
win_size,
hop_size,
duration,
data_window=100,
window_shift=10,
augment=True,
save=True,
minimum_length=1):
'''
Merge this function and spectrograms. Can replace spectrogram function #1
'''
from audio import compute_spectrogram
spectrograms = list()
labels = list()
if input_type == 'others':
#load the npz file which in this case is data_list
print('Loading the features..')
spectrograms = np.load(data_list)['features']
print('Length is: ', len(spectrograms))
# Make these features unified across time dimension
spectrograms = [
update_feature_matrix(matrix) for matrix in spectrograms
]
else:
print('Computing the ' + input_type + ' spectrograms !!')
with open(data_list, 'r') as f:
spectrograms = [
compute_spectrogram(input_type, file.strip(), fft_size,
win_size, hop_size, duration, augment,
minimum_length) for file in f
]
# Get the labels into a list and save it along with the spectrograms
with open(labelFile, 'r') as f:
labels = [line.strip() for line in f]
if augment:
new_data = list()
new_labels = list()
assert (len(labels) == len(spectrograms))
print(
'Now performing augmentation using sliding window mechanism on original specs/features .... '
)
for i in range(len(spectrograms)):
d, l = augment_data(spectrograms[i], labels[i], data_window,
input_type, window_shift)
new_data.extend(
d) # extend the list rather than adding it into a new list
new_labels.extend(l)
spectrograms = new_data
labels = new_labels
if save:
from helper import makeDirectory
makeDirectory(savePath)
outfile = savePath + '/spec'
with open(outfile, 'w') as f:
np.savez(outfile, spectrograms=spectrograms, labels=labels)
print('Finished computing features/spectrograms and saved inside: ',
savePath)
