def make_model(number_classes):
batch = data.wave_batch_generator(batch_size=1000,
target=data.Target.speaker)
X, Y = next(batch)
# Classification
tflearn.init_graph(num_cores=8, gpu_memory_fraction=0.5)
net = tflearn.input_data(shape=[None, 3848]) #Two wave chunks
net = tflearn.fully_connected(net, 128)
net = tflearn.dropout(net, 0.5)
net = tflearn.fully_connected(net, 16)
net = tflearn.dropout(net, 0.8)
net = tflearn.fully_connected(net, 128)
net = tflearn.dropout(net, 0.5)
net = tflearn.fully_connected(net, number_classes, activation='softmax')
net = tflearn.regression(net,
optimizer='adam',
loss='categorical_crossentropy')
model = tflearn.DNN(net)
return model
def train(number_classes):
model = ml.make_model(number_classes)
batch = data.wave_batch_generator(batch_size=1000,
target=data.Target.speaker)
X, Y = next(batch)
model.fit(X, Y, n_epoch=100, show_metric=True, snapshot_step=100)
model.save('classifier')
def main():
batch = speech_data.wave_batch_generator(10000,
target=speech_data.Target.digits)
X, Y = next(batch)
Y = [numpy.hstack([y, numpy.array([0, 0, 0, 0, 0, 0])]) for y in Y]
# Y = map(lambda a: , Y)
print(type(Y))
# print (np.hstack([Y[0], np.array([0, 0, 0, 0, 0, 0])]))
number_classes = 16 # Digits
# Classification
tflearn.init_graph(num_cores=8, gpu_memory_fraction=0.5)
net = tflearn.input_data(shape=[None, 8192])
net = tflearn.fully_connected(net, 64, name='f1')
net = tflearn.dropout(net, 0.5, name='dp')
net = tflearn.fully_connected(net,
number_classes,
activation='softmax',
name='f2')
net = tflearn.regression(net,
optimizer='sgd',
loss='categorical_crossentropy')
model = tflearn.DNN(net)
model.load('pre-trained/model.tflearn.sgd_trained')
# Overfitting okay for now
totalTime = 0
totalAcc = 0
numTimes = 100
for i in range(numTimes):
t = time.time()
result = model.predict(X)
print("-------------")
result = numpy.array([numpy.argmax(r) for r in result])
answers = numpy.array([numpy.argmax(answer) for answer in Y])
print(i, ">>>", (result == answers).sum() / float(len(answers)),
"time: ",
time.time() - t)
totalAcc = totalAcc + (result == answers).sum() / float(len(answers))
totalTime = totalTime + time.time() - t
print("Avg. Acc. = ", totalAcc / numTimes)
print("Avg. time = ", totalTime / numTimes)
#!/usr/bin/env python
#!/usr/bin/env PYTHONIOENCODING="utf-8" python
import tflearn
import pyaudio
import speech_data
import numpy
import os
import tensorflow as tf
# Simple spoken digit recognition demo, with 98% accuracy in under a minute
# Training Step: 544 | total loss: 0.15866
# | Adam | epoch: 034 | loss: 0.15866 - acc: 0.9818 -- iter: 0000/1000
batch = speech_data.wave_batch_generator(10000,
target=speech_data.Target.digits)
X, Y = next(batch)
number_classes = 10 # Digits
# Classification
tflearn.init_graph(num_cores=8, gpu_memory_fraction=0.5)
net = tflearn.input_data(shape=[None, 8192])
#net = tflearn.fully_connected(net, 64)
net = tflearn.fully_connected(net, 64)
net = tflearn.dropout(net, 0.5)
net = tflearn.fully_connected(net, number_classes, activation='softmax')
net = tflearn.regression(net,
optimizer='adam',
loss='categorical_crossentropy')
import tflearn
import pyaudio
import speech_data as data
# Simple speaker recognition demo, with 99% accuracy in under a minute ( on digits sample )
# | Adam | epoch: 030 | loss: 0.05330 - acc: 0.9966 -- iter: 0000/1000
# 'predicted speaker for 9_Vicki_260 : result = ', 'Vicki'
speakers = data.get_speakers()
number_classes = len(speakers)
##print("speakers",speakers)
WORD_WAVs = "spoken_words"
batch = data.wave_batch_generator(batch_size=1000,
source=WORD_WAVs,
target=data.Target.speaker)
X, Y = next(batch)
# Classification
tflearn.init_graph(num_cores=8, gpu_memory_fraction=0.5)
net = tflearn.input_data(shape=[None, 8192]) #Two wave chunks
net = tflearn.fully_connected(net, 64)
net = tflearn.dropout(net, 0.5)
net = tflearn.fully_connected(net, number_classes, activation='softmax')
net = tflearn.regression(net,
optimizer='adam',
loss='categorical_crossentropy')
model = tflearn.DNN(net)
# | Adam | epoch: 030 | loss: 0.05330 - acc: 0.9966 -- iter: 0000/1000
# 'predicted speaker for 9_Vicki_260 : result = ', 'Vicki'
import tensorflow as tf
print("You are using tensorflow version " +
tf.__version__) #+" tflearn version "+ tflearn.version)
if tf.__version__ >= '0.12' and os.name == 'nt':
print("sorry, tflearn is not ported to tensorflow 0.12 on windows yet!(?)")
quit() # why? works on Mac?
speakers = data.get_speakers()
number_classes = len(speakers)
print("speakers", speakers)
batch = data.wave_batch_generator(batch_size=1000,
source=data.Source.DIGIT_WAVES,
target=data.Target.speaker)
X, Y = next(batch)
# Classification
tflearn.init_graph(num_cores=8, gpu_memory_fraction=0.5)
net = tflearn.input_data(shape=[None, 8192]) #Two wave chunks
net = tflearn.fully_connected(net, 64)
net = tflearn.dropout(net, 0.5)
net = tflearn.fully_connected(net, number_classes, activation='softmax')
net = tflearn.regression(net,
optimizer='adam',
loss='categorical_crossentropy')
model = tflearn.DNN(net)
#!/usr/local/bin/python
#!/usr/bin/env PYTHONIOENCODING="utf-8" python
import os
import tflearn
import speech_data as data
import tensorflow as tf
import numpy
speakers = data.get_speakers()
number_classes = len(speakers)
print("speakers", speakers)
batch = data.wave_batch_generator(batch_size=1000, target=data.Target.speaker)
X, Y = next(batch)
# Classification
tflearn.init_graph(num_cores=8, gpu_memory_fraction=0.5)
net = tflearn.input_data(shape=[None, 3848]) #Two wave chunks
net = tflearn.fully_connected(net, 128)
net = tflearn.dropout(net, 0.5)
net = tflearn.fully_connected(net, 16)
net = tflearn.dropout(net, 0.8)
net = tflearn.fully_connected(net, 128)
net = tflearn.dropout(net, 0.5)
net = tflearn.fully_connected(net, number_classes, activation='softmax')
#!/usr/bin/env python
#!/usr/bin/env PYTHONIOENCODING="utf-8" python
import tflearn
#import pyaudio
import speech_data
import numpy
import sys
load = speech_data.wave_batch_generator(2404)
X, Y = next(load)
number_classes = 10 # Digits
tflearn.init_graph(num_cores=8, gpu_memory_fraction=0.5)
# nombre de coeur allouer et memoire gpu allouer
net = tflearn.input_data(shape=[None, 8192])
net = tflearn.fully_connected(net, 64)
net = tflearn.dropout(net, 0.5)
net = tflearn.fully_connected(net, number_classes, activation='softmax')
net = tflearn.regression(net,
optimizer='adam',
learning_rate=0.0001,
loss='categorical_crossentropy')
model = tflearn.DNN(net)
model.fit(X,
Y,
validation_set=0.2,
n_epoch=200,
show_metric=True,
#!/usr/bin/env PYTHONIOENCODING="utf-8" python import tflearn import pyaudio import speech_data import numpy # Simple spoken digit recognition demo, with 98% accuracy in under a minute # Training Step: 544 | total loss: 0.15866 # | Adam | epoch: 034 | loss: 0.15866 - acc: 0.9818 -- iter: 0000/1000 batch=speech_data.wave_batch_generator(10000,target=speech_data.Target.digits) X,Y=next(batch) number_classes=10 # Digits # Classification tflearn.init_graph(num_cores=8, gpu_memory_fraction=0.5) net = tflearn.input_data(shape=[None, 8192]) net = tflearn.fully_connected(net, 64) net = tflearn.dropout(net, 0.5) net = tflearn.fully_connected(net, number_classes, activation='softmax') net = tflearn.regression(net, optimizer='adam', loss='categorical_crossentropy') model = tflearn.DNN(net) model.fit(X, Y,n_epoch=3,show_metric=True,snapshot_step=100) # Overfitting okay for now demo_file = "5_Vicki_260.wav" demo=speech_data.load_wav_file(speech_data.path + demo_file)
#!/usr/bin/env PYTHONIOENCODING="utf-8" python
import tflearn
import pyaudio
import speech_data as data
# Simple speaker recognition demo, with 99% accuracy in under a minute ( on digits sample )
# | Adam | epoch: 030 | loss: 0.05330 - acc: 0.9966 -- iter: 0000/1000
# 'predicted speaker for 9_Vicki_260 : result = ', 'Vicki'
speakers = data.get_speakers()
number_classes=len(speakers)
print("speakers",speakers)
WORD_WAVs="spoken_words"
batch=data.wave_batch_generator(batch_size=1000,source=WORD_WAVs,target=data.Target.speaker)
X,Y=next(batch)
# Classification
tflearn.init_graph(num_cores=8, gpu_memory_fraction=0.5)
net = tflearn.input_data(shape=[None, 8192]) #Two wave chunks
net = tflearn.fully_connected(net, 64)
net = tflearn.dropout(net, 0.5)
net = tflearn.fully_connected(net, number_classes, activation='softmax')
net = tflearn.regression(net, optimizer='adam', loss='categorical_crossentropy')
model = tflearn.DNN(net)
model.fit(X, Y, n_epoch=100, show_metric=True, snapshot_step=100)
# Training Step: 544 | total loss: 0.15866
# | Adam | epoch: 034 | loss: 0.15866 - acc: 0.9818 -- iter: 0000/1000
# 98% Accuracy on training set in just a minute
# audio = pyaudio.PyAudio()
# # format=pyaudio.paFloat32
# format=pyaudio.paInt8
# # format=audio.get_format_from_width(f.getsampwidth())
# # out_stream = audio.open( format=format,channels = f.getnchannels(), rate=f.getframerate(), output= True)
# out_stream = audio.open( format=format,channels = 1, rate=48000, output= True)
# out_stream.start_stream()
# def play_pcm(data):
# out_stream.write(data)
batch = speech_data.wave_batch_generator(1000)
X, Y = next(batch)
# Classification
# x = tflearn.input_data(shape=[None, 8192])
# net = tflearn.fully_connected(x, 64)
# net = tflearn.dropout(net, 0.5)
# net = tflearn.fully_connected(net, 10, activation='softmax')
# net = tflearn.regression(net, optimizer='adam', loss='categorical_crossentropy')
# y = net.placeholder
# classifier = tflearn.DNN(net)
def model(net):
# type: (layer.net) -> None
# net.inputnet_data(shape=[None, 10])
import tflearn
from speech_data import wave_batch_generator, Target, load_wav_file, path
import numpy
# Simple spoken digit recognition demo, with 98% accuracy in under a minute
# Training Step: 544 | total loss: 0.15866
# | Adam | epoch: 034 | loss: 0.15866 - acc: 0.9818 -- iter: 0000/1000
if __name__ == '__main__':
batch = wave_batch_generator(10000, target=Target.digits)
X, Y = next(batch)
number_classes = 10 # Digits
# Classification
tflearn.init_graph(num_cores=8, gpu_memory_fraction=0.5)
net = tflearn.input_data(shape=[None, 8192])
net = tflearn.fully_connected(net, 64)
net = tflearn.dropout(net, 0.5)
net = tflearn.fully_connected(net, number_classes, activation='softmax')
net = tflearn.regression(net,
optimizer='adam',
loss='categorical_crossentropy')
model = tflearn.DNN(net)
model.fit(X, Y, n_epoch=3, show_metric=True, snapshot_step=100)
# Overfitting okay for now
demo_file = "5_Vicki_260.wav"
#!/usr/bin/env PYTHONIOENCODING="utf-8" python
import tflearn
import pyaudio
import speech_data as data
# Simple speaker recognition demo, with 99% accuracy in under a minute ( on digits sample )
# | Adam | epoch: 030 | loss: 0.05330 - acc: 0.9966 -- iter: 0000/1000
# 'predicted speaker for 9_Vicki_260 : result = ', 'Vicki'
speakers = data.get_speakers()
number_classes = len(speakers)
print("speakers", speakers)
target = data.Target.speaker
batch = data.wave_batch_generator(batch_size=1000, target=target)
X, Y = next(batch)
# Classification
tflearn.init_graph(num_cores=8, gpu_memory_fraction=0.5)
net = tflearn.input_data(shape=[None, 8192])
net = tflearn.fully_connected(net, 64)
net = tflearn.dropout(net, 0.5)
net = tflearn.fully_connected(net, number_classes, activation='softmax')
net = tflearn.regression(net,
optimizer='adam',
loss='categorical_crossentropy')
model = tflearn.DNN(net)
model.fit(X, Y, n_epoch=100, show_metric=True, snapshot_step=100)
import speech_data
import numpy as np
from constants import POSITIVE_PATH, NEGATIVE_PATH
learning_rate = 0.03
training_iters = 10 # steps
batch_size = 2
width = 1
height = 8192 # (max) length of utterance
classes = 2 # digits
positive_batch = word_batch = speech_data.wave_batch_generator(
batch_size, POSITIVE_PATH, [1.0, 0.0])
negative_batch = word_batch = speech_data.wave_batch_generator(
batch_size, NEGATIVE_PATH, [0.0, 1])
#positive_batch = np.reshape(positive_batch, (1, 10, 8192))
#negative_batch = np.reshape(negative_batch, (1, 10, 8192))
# Network building
'''net = tflearn.input_data(shape=[None, 8192])
net = tflearn.fully_connected(net, 64)
net = tflearn.dropout(net, 0.5)
net = tflearn.fully_connected(net, classes, activation='softmax')
net = tflearn.regression(net, optimizer='adam', loss='categorical_crossentropy')
'''
net = tflearn.input_data([None, width, height])
net = tflearn.lstm(net, 128 * 4, dropout=0.2)
# Training Step: 544 | total loss: 0.15866 # | Adam | epoch: 034 | loss: 0.15866 - acc: 0.9818 -- iter: 0000/1000 # 98% Accuracy on training set in just a minute # audio = pyaudio.PyAudio() # # format=pyaudio.paFloat32 # format=pyaudio.paInt8 # # format=audio.get_format_from_width(f.getsampwidth()) # # out_stream = audio.open( format=format,channels = f.getnchannels(), rate=f.getframerate(), output= True) # out_stream = audio.open( format=format,channels = 1, rate=48000, output= True) # out_stream.start_stream() # def play_pcm(data): # out_stream.write(data) batch=speech_data.wave_batch_generator(1000) X,Y=next(batch) # Classification # x = tflearn.input_data(shape=[None, 8192]) # net = tflearn.fully_connected(x, 64) # net = tflearn.dropout(net, 0.5) # net = tflearn.fully_connected(net, 10, activation='softmax') # net = tflearn.regression(net, optimizer='adam', loss='categorical_crossentropy') # y = net.placeholder # classifier = tflearn.DNN(net) def model(net): # type: (layer.net) -> None # net.inputnet_data(shape=[None, 10]) net.fully_connected( 64)
#!/usr/bin/env PYTHONIOENCODING="utf-8" python import tflearn import os import speech_data as data # training and testing data sets train_data = '/home/cc/Data/small-clean-train/' test_data = '/home/cc/Data/small-clean-test/' # grab the speakers from the training directory speakers = data.get_speakers(train_data) number_classes = len(speakers) # create the MFCC arrays from the data for training batch=data.wave_batch_generator(batch_size=1000,source=WORD_WAVs,target=data.Target.speaker,speakers=speakers) X,Y=next(batch) # Classification tflearn.init_graph(num_cores=8, gpu_memory_fraction=0.5) net = tflearn.input_data(shape=[None, 8192]) #Two wave chunks net = tflearn.fully_connected(net, 64) # seems like a higher dropout rate works better -- why is this?? net = tflearn.dropout(net, 0.5) net = tflearn.fully_connected(net, number_classes, activation='softmax') net = tflearn.regression(net, optimizer='adam', loss='categorical_crossentropy') model = tflearn.DNN(net) model.fit(X, Y, n_epoch=100, show_metric=True, snapshot_step=100)
