I am a beginner in tensorflow and I am trying to train a model using "mini batch". To do that I created a generator and iterate it. The problem I encounter is that, at the beginning of the epoch, the train seems fast (many batch per seconds) then the train slow down (1 batch per second) so I am wondering where I am wrong in my code but I do not find the problem.
def prepare_data(filename):
'''load file which give path and label for the data'''
f = open(filename, 'r')
data = [line.split() for line in f]
feat =[]
label=[]
for l in data:
feat.append(l[0])
label.append(l[1])
n_samples = len(feat)
shuf = list(range(n_samples))
random.shuffle(shuf)
count = Counter(label)
print(count)
feature = [feat[i] for i in shuf]
label = np.array(label, dtype=np.int)
return feature, label[shuf]
def get_specgrams(paths, nsamples=16000):
'''
Given list of paths, return specgrams.
'''
# read the wav files
wavs = [wavfile.read(x)[1] for x in paths]
# zero pad the shorter samples and cut off the long ones.
data = []
for wav in wavs:
if wav.size < 16000:
d = np.pad(wav, (nsamples - wav.size, 0), mode='constant')
else:
d = wav[0:nsamples]
data.append(d)
# get the specgram
#specgram = [signal.spectrogram(d, nperseg=256, noverlap=128)[2] for d in data]
#specgram = [s.reshape(129, 124, -1) for s in specgram]
return np.asarray(data)
def get_specgram(path, nsamples=16000):
'''
Given path, return specgrams.
'''
# read the wav files
wav = wavfile.read(path)[1]
# zero pad the shorter samples and cut off the long ones.
if wav.size < 16000:
d = np.pad(wav, (nsamples - wav.size, 0), mode='constant')
else:
d = wav[0:nsamples]
# get the specgram
#specgram = [signal.spectrogram(d, nperseg=256, noverlap=128)[2] for d in data]
#specgram = [s.reshape(129, 124, -1) for s in specgram]
return d
# multci classification binary labels
def one_hot_encode(labels, n_unique_labels=31):
n_labels = len(labels)
#print('number of unique labels:', n_unique_labels)
one_hot_encode = np.zeros((n_labels,n_unique_labels))
one_hot_encode[np.arange(n_labels), labels] = 1
return np.array(one_hot_encode, dtype=np.int)
#create_path_file('train/audio/')
def model(tr_features, tr_labels, ts_features, ts_labels):
# remove gpu device error
config = tf.ConfigProto(allow_soft_placement = True)
# parameters
BATCH_SIZE = 4
number_loop = math.ceil(len(tr_features)/BATCH_SIZE)
training_epochs = 10
n_dim = 16000
n_classes = 31 #len(np.unique(ts_labels))
n_hidden_units_one = 280
n_hidden_units_two = 300
sd = 1 / np.sqrt(n_dim)
learning_rate = 0.1
# get test data
ts_features, ts_labels = get_data(ts_features, ts_labels)
# Model
X = tf.placeholder(tf.float32,[None,n_dim])
Y = tf.placeholder(tf.float32,[None,n_classes])
W_1 = tf.Variable(tf.random_normal([n_dim,n_hidden_units_one], mean = 0, stddev=sd))
b_1 = tf.Variable(tf.random_normal([n_hidden_units_one], mean = 0, stddev=sd))
h_1 = tf.nn.tanh(tf.matmul(X,W_1) + b_1)
W_2 = tf.Variable(tf.random_normal([n_hidden_units_one,n_hidden_units_two], mean = 0, stddev=sd))
b_2 = tf.Variable(tf.random_normal([n_hidden_units_two], mean = 0, stddev=sd))
h_2 = tf.nn.sigmoid(tf.matmul(h_1,W_2) + b_2)
W = tf.Variable(tf.random_normal([n_hidden_units_two,n_classes], mean = 0, stddev=sd))
b = tf.Variable(tf.random_normal([n_classes], mean = 0, stddev=sd))
y_ = tf.nn.softmax(tf.matmul(h_2,W) + b)
init = tf.initialize_all_variables()
# function and optimizers
cost_function = -tf.reduce_sum(Y * tf.log(y_))
optimizer = tf.train.GradientDescentOptimizer(learning_rate).minimize(cost_function)
correct_prediction = tf.equal(tf.argmax(y_,1), tf.argmax(Y,1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
# train loop
cost_history = np.empty(shape=[1],dtype=float)
y_true, y_pred = None, None
with tf.Session(config = config) as sess:
sess.run(init)
for epoch in range(training_epochs):
print(' ## Epoch n°', epoch+1 )
batch = batch_generator(BATCH_SIZE, tr_features, tr_labels)
acc_total = 0.0
for cpt, (train_features_batch, train_labels_batch) in enumerate(batch):
_,cost = sess.run([optimizer,cost_function],feed_dict={X:train_features_batch,Y:train_labels_batch})
cost_history = np.append(cost_history,cost)
correct_prediction = tf.equal(tf.argmax(y_,1), tf.argmax(Y,1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
acc = accuracy.eval(feed_dict={X:train_features_batch,Y:train_labels_batch})
acc_total = (acc_total*cpt + acc)/(cpt+1)
print('Train accuracy : ', acc_total, '[',str(cpt+1), '/',str(number_loop), ']' ,flush=True, end='\r')
clear_output()
print('Train accuracy : ', acc_total)
y_pred = sess.run(tf.argmax(y_,1),feed_dict={X: ts_features})
y_true = sess.run(tf.argmax(ts_labels,1))
print('Test accuracy: ', round(sess.run(accuracy, feed_dict={X: ts_features, Y: ts_labels}) , 3))
fig = plt.figure(figsize=(10,8))
plt.plot(cost_history)
plt.axis([0,training_epochs,0,np.max(cost_history)])
plt.show()
p,r,f,s = precision_recall_fscore_support(y_true, y_pred, average='micro')
print("F-Score:", round(f,3))
def batch_generator(batch_size, feat_path, labels):
n_sample = len(feat_path)
ite = math.ceil(n_sample/batch_size)
for i in range(0, ite):
if i == ite-1:
label = one_hot_encode(labels[-batch_size:])
feat = get_specgrams(feat_path[-batch_size:])
yield (feat, label)
else:
label = one_hot_encode(labels[i*batch_size:i*batch_size+batch_size])
feat = get_specgrams(feat_path[i*batch_size:i*batch_size+batch_size])
yield (feat, label)
def get_data(feat_path, labels):
feat = get_specgrams(feat_path)
label = one_hot_encode(labels)
return feat, label
def __main__():
print('## Load data and shuffle')
feat_path, labels = prepare_data('data_labelised2.txt')
idx = int(len(labels)*0.8)
print("## Create Model")
model(feat_path[0:idx], labels[0:idx], feat_path[idx+1:], labels[idx+1:])
with tf.device('/gpu:0'):
__main__()