chengqingshui
commented
6 years ago Hello, everyone!
I run the project, and found some problem, and change code as bellow, then it works well. 1, about the mfcc_batch_generator funtion, which generate a batch of sound feature data and labels, but in the trainning step, the data is not updated by the next() in the loop. So I add a new function mfcc_batch_generatorEx simaliar to mfcc_batch_generator in speech_data.py file:
def mfcc_batch_generatorEx(batch_size=10, source=Source.DIGIT_WAVES, target=Target.digits): maybe_download(source, DATA_DIR) if target == Target.speaker: speakers = get_speakers() batch_features = [] labels = [] files = os.listdir(path)
print("loaded batch of %d files" % len(files))
shuffle(files)
for wav in files:
if not wav.endswith(".wav"):
continue
wave, sr = librosa.load(path+wav, mono=True)
if target==Target.speaker:
label=one_hot_from_item(speaker(wav), speakers)
elif target==Target.digits:
label=dense_to_one_hot(int(wav[0]),10)
elif target==Target.first_letter:
label=dense_to_one_hot((ord(wav[0]) - 48) % 32,32)
else:
raise Exception("todo : labels for Target!")
labels.append(label)
mfcc = librosa.feature.mfcc(wave, sr)
# print(np.array(mfcc).shape)
mfcc = np.pad(mfcc,((0,0),(0,80-len(mfcc[0]))), mode='constant', constant_values=0)
batch_features.append(np.array(mfcc))
return batch_features, labels 2、 in demo.py file, generate all sound features and labels by using bellow
X, Y = speech_data.mfcc_batch_generatorEx(batch_size)
3、 in the training step, using code bellow:
with tf.Session() as sess: model.fit(trainX, trainY, n_epoch=training_iters)#, validation_set=(testX, testY), show_metric=True,batch_size=batch_size) _y = model.predict(X) YY = [x.tolist() for x in Y] corrent_prediction = tf.equal(tf.arg_max(_y,1), tf.arg_max(YY,1)) accuracy = tf.reduce_mean(tf.cast(corrent_prediction, tf.float32)) print("\n\ncorrent_prediction = " , sess.run(accuracy) )
model.save("tflearn.lstm.model")
Sorry to the bad format, Thanks for all !
AiJunXian
raffaelbd
Hello, everyone! I run the project, and found some problem, and change code as bellow, then it works well. 1, about the mfcc_batch_generator funtion, which generate a batch of sound feature data and labels, but in the trainning step, the data is not updated by the next() in the loop. So I add a new function mfcc_batch_generatorEx simaliar to mfcc_batch_generator in speech_data.py file:
def mfcc_batch_generatorEx(batch_size=10, source=Source.DIGIT_WAVES, target=Target.digits): maybe_download(source, DATA_DIR) if target == Target.speaker: speakers = get_speakers() batch_features = [] labels = [] files = os.listdir(path)
2、 in demo.py file, generate all sound features and labels by using bellow X, Y = speech_data.mfcc_batch_generatorEx(batch_size) 3、 in the training step, using code bellow: with tf.Session() as sess: model.fit(trainX, trainY, n_epoch=training_iters)#, validation_set=(testX, testY), show_metric=True,batch_size=batch_size) _y = model.predict(X) YY = [x.tolist() for x in Y] corrent_prediction = tf.equal(tf.arg_max(_y,1), tf.arg_max(YY,1)) accuracy = tf.reduce_mean(tf.cast(corrent_prediction, tf.float32)) print("\n\ncorrent_prediction = " , sess.run(accuracy) )
model.save("tflearn.lstm.model")