Skip to content

Navigation Menu

Sign in
Appearance settings

Search code, repositories, users, issues, pull requests...

Provide feedback

We read every piece of feedback, and take your input very seriously.

Saved searches

Use saved searches to filter your results more quickly
Sign up
Appearance settings
This repository was archived by the owner on Mar 10, 2023. It is now read-only.

Commit 581490e

Browse files
committed
Added functionality for training neural network with multiple files.
1 parent e3be282 commit 581490e

File tree

1 file changed

+111
-72
lines changed

1 file changed

+111
-72
lines changed

‎train.py‎

Lines changed: 111 additions & 72 deletions
Original file line numberDiff line numberDiff line change
@@ -8,6 +8,7 @@
88
import re
99
import logging
1010
import sys
11+
import math
1112
import unicodedata
1213
import codecs
1314

@@ -324,40 +325,48 @@ def make_sequences_same_length(sequences, sequence_lengths, default_value=0.0):
324325

325326

326327
def main(argv):
327-
# Read text file.
328-
text_file_path = TRAIN_DIR + "211-122425-0059.txt"
329-
text = read_text_file(text_file_path)
330-
text = normalize_text(text)
331-
332-
# Read audio file.
333-
wav_file_path = TRAIN_DIR + "211-122425-0059.wav"
334-
audio_rate, audio_data = wav.read(wav_file_path)
335-
inputs = mfcc(audio_data, samplerate=audio_rate)
336-
337-
# Make text as as char array.
338-
labels = make_char_array(text, SPACE_TOKEN)
339-
labels = np.asarray([SPACE_INDEX if x == SPACE_TOKEN else ord(x) - FIRST_INDEX for x in labels])
340-
341-
# Labels sparse representation for feeding the placeholder.
342-
train_labels = sparse_tuples_from_sequences([labels])
343-
344-
# Train inputs.
345-
train_inputs = np.asarray(inputs[np.newaxis, :])
328+
# Read train data files.
329+
train_texts = read_text_files(TRAIN_DIR)
330+
train_labels = texts_encoder(train_texts,
331+
first_index=FIRST_INDEX,
332+
space_index=SPACE_INDEX,
333+
space_token=SPACE_TOKEN)
334+
train_inputs = read_audio_files(TRAIN_DIR)
346335
train_inputs = standardize_audios(train_inputs)
347-
train_sequence_length = [train_inputs.shape[1]]
348-
349-
# TODO(ugnelis): define different validation variables.
350-
validation_inputs = train_inputs
351-
validation_labels = train_labels
352-
validation_sequence_length = train_sequence_length
336+
train_sequence_lengths = get_sequence_lengths(train_inputs)
337+
train_inputs = make_sequences_same_length(train_inputs, train_sequence_lengths)
338+
339+
# Read validation data files.
340+
validation_texts = read_text_files(DEV_DIR)
341+
validation_labels = texts_encoder(validation_texts,
342+
first_index=FIRST_INDEX,
343+
space_index=SPACE_INDEX,
344+
space_token=SPACE_TOKEN)
345+
validation_labels = sparse_tuples_from_sequences(validation_labels)
346+
validation_inputs = read_audio_files(DEV_DIR)
347+
validation_inputs = standardize_audios(validation_inputs)
348+
validation_sequence_lengths = get_sequence_lengths(validation_inputs)
349+
validation_inputs = make_sequences_same_length(validation_inputs, validation_sequence_lengths)
350+
351+
# Read test data files.
352+
test_texts = read_text_files(TEST_DIR)
353+
test_labels = texts_encoder(test_texts,
354+
first_index=FIRST_INDEX,
355+
space_index=SPACE_INDEX,
356+
space_token=SPACE_TOKEN)
357+
test_labels = sparse_tuples_from_sequences(test_labels)
358+
test_inputs = read_audio_files(DEV_DIR)
359+
test_inputs = standardize_audios(test_inputs)
360+
test_sequence_lengths = get_sequence_lengths(test_inputs)
361+
test_inputs = make_sequences_same_length(test_inputs, test_sequence_lengths)
353362

354363
with tf.device('/gpu:0'):
355364
config = tf.ConfigProto(allow_soft_placement=True)
356365
config.gpu_options.allow_growth = True
357366

358367
graph = tf.Graph()
359368
with graph.as_default():
360-
369+
logging.debug("Starting new TensorFlow graph.")
361370
inputs_placeholder = tf.placeholder(tf.float32, [None, None, NUM_FEATURES])
362371

363372
# SparseTensor placeholder required by ctc_loss op.
@@ -382,13 +391,13 @@ def main(argv):
382391
# Reshaping to apply the same weights over the time steps.
383392
outputs = tf.reshape(outputs, [-1, NUM_HIDDEN])
384393

385-
weights = tf.Variable(tf.truncated_normal([NUM_HIDDEN,
394+
weigths = tf.Variable(tf.truncated_normal([NUM_HIDDEN,
386395
NUM_CLASSES],
387396
stddev=0.1))
388397
biases = tf.Variable(tf.constant(0., shape=[NUM_CLASSES]))
389398

390399
# Doing the affine projection.
391-
logits = tf.matmul(outputs, weights) + biases
400+
logits = tf.matmul(outputs, weigths) + biases
392401

393402
# Reshaping back to the original shape.
394403
logits = tf.reshape(logits, [batch_s, -1, NUM_CLASSES])
@@ -402,53 +411,83 @@ def main(argv):
402411
optimizer = tf.train.MomentumOptimizer(INITIAL_LEARNING_RATE, 0.9).minimize(cost)
403412

404413
# CTC decoder.
405-
decoded, log_prob = tf.nn.ctc_greedy_decoder(logits, sequence_length_placeholder)
414+
decoded, neg_sum_logits = tf.nn.ctc_greedy_decoder(logits, sequence_length_placeholder)
406415

407416
label_error_rate = tf.reduce_mean(tf.edit_distance(tf.cast(decoded[0], tf.int32),
408417
labels_placeholder))
409-
with tf.Session(graph=graph) as session:
410-
# Initialize the weights and biases.
411-
tf.global_variables_initializer().run()
412-
413-
for current_epoch in range(NUM_EPOCHS):
414-
train_cost = train_label_error_rate = 0
415-
start_time = time.time()
416-
417-
for batch in range(NUM_BATCHES_PER_EPOCH):
418-
feed = {inputs_placeholder: train_inputs,
419-
labels_placeholder: train_labels,
420-
sequence_length_placeholder: train_sequence_length}
421-
422-
batch_cost, _ = session.run([cost, optimizer], feed)
423-
train_cost += batch_cost * BATCH_SIZE
424-
train_label_error_rate += session.run(label_error_rate, feed_dict=feed) * BATCH_SIZE
425-
426-
train_cost /= NUM_EXAMPLES
427-
train_label_error_rate /= NUM_EXAMPLES
428-
429-
val_feed = {inputs_placeholder: validation_inputs,
430-
labels_placeholder: validation_labels,
431-
sequence_length_placeholder: validation_sequence_length}
432-
433-
validation_cost, validation_label_error_rate = session.run([cost, label_error_rate], feed_dict=val_feed)
434-
435-
# Output intermediate step information.
436-
logging.info("Epoch %d/%d (time: %.3f s)",
437-
current_epoch + 1,
438-
NUM_EPOCHS,
439-
time.time() - start_time)
440-
logging.info("Train cost: %.3f, train label error rate: %.3f",
441-
train_cost,
442-
train_label_error_rate)
443-
logging.info("Validation cost: %.3f, validation label error rate: %.3f",
444-
validation_cost,
445-
validation_label_error_rate)
446-
447-
# Decoding.
448-
decoded_outputs = session.run(decoded[0], feed_dict=feed)
449-
decoded_text = sequence_decoder(decoded_outputs[1])
450-
451-
logging.info("Original:\n%s", text)
418+
419+
with tf.Session(graph=graph) as session:
420+
logging.debug("Starting TensorFlow session.")
421+
# Initialize the weights and biases.
422+
tf.global_variables_initializer().run()
423+
424+
train_num = train_inputs.shape[0]
425+
validation_num = validation_inputs.shape[0]
426+
427+
# Check if there is any example.
428+
if train_num <= 0:
429+
logging.error("There are no training examples.")
430+
return
431+
432+
num_batches_per_epoch = math.ceil(train_num / BATCH_SIZE)
433+
434+
for current_epoch in range(NUM_EPOCHS):
435+
train_cost = 0
436+
train_label_error_rate = 0
437+
start_time = time.time()
438+
439+
for batch in range(num_batches_per_epoch):
440+
# Format batches.
441+
if int(train_num / ((batch + 1) * BATCH_SIZE)) >= 1:
442+
indexes = [i % train_num for i in range(batch * BATCH_SIZE, (batch + 1) * BATCH_SIZE)]
443+
else:
444+
indexes = [i % train_num for i in range(batch * BATCH_SIZE, train_num)]
445+
446+
batch_train_inputs = train_inputs[indexes]
447+
batch_train_sequence_lengths = train_sequence_lengths[indexes]
448+
batch_train_targets = sparse_tuples_from_sequences(train_labels[indexes])
449+
450+
feed = {inputs_placeholder: batch_train_inputs,
451+
labels_placeholder: batch_train_targets,
452+
sequence_length_placeholder: batch_train_sequence_lengths}
453+
454+
batch_cost, _ = session.run([cost, optimizer], feed)
455+
train_cost += batch_cost * BATCH_SIZE
456+
train_label_error_rate += session.run(label_error_rate, feed_dict=feed) * BATCH_SIZE
457+
458+
train_cost /= train_num
459+
train_label_error_rate /= train_num
460+
461+
validation_feed = {inputs_placeholder: validation_inputs,
462+
labels_placeholder: validation_labels,
463+
sequence_length_placeholder: validation_sequence_lengths}
464+
465+
validation_cost, validation_label_error_rate = session.run([cost, label_error_rate],
466+
feed_dict=validation_feed)
467+
468+
validation_cost /= validation_num
469+
validation_label_error_rate /= validation_num
470+
471+
# Output intermediate step information.
472+
print("Epoch %d/%d (time: %.3f s)" %
473+
(current_epoch + 1, NUM_EPOCHS, time.time() - start_time))
474+
print("Train cost: %.3f, train label error rate: %.3f" %
475+
(train_cost, train_label_error_rate))
476+
print("Validation cost: %.3f, validation label error rate: %.3f" %
477+
(validation_cost, validation_label_error_rate))
478+
479+
test_feed = {inputs_placeholder: test_inputs,
480+
sequence_length_placeholder: test_sequence_lengths}
481+
# Decoding.
482+
decoded_outputs = session.run(decoded[0], feed_dict=test_feed)
483+
dense_decoded = tf.sparse_tensor_to_dense(decoded_outputs, default_value=-1).eval(session=session)
484+
485+
for i, sequence in enumerate(dense_decoded):
486+
sequence = [s for s in sequence if s != -1]
487+
decoded_text = sequence_decoder(sequence)
488+
489+
print('Sequence %d' % i)
490+
logging.info("Original:\n%s", test_texts[i])
452491
logging.info("Decoded:\n%s", decoded_text)
453492

454493

0 commit comments

Comments
(0)

AltStyle によって変換されたページ (->オリジナル) /