def transcribe_file(audio_path, tlog_path):
from DeepSpeech import create_model # pylint: disable=cyclic-import,import-outside-toplevel
from util.checkpoints import load_or_init_graph
initialize_globals()
scorer = Scorer(FLAGS.lm_alpha, FLAGS.lm_beta, FLAGS.scorer_path,
Config.alphabet)
try:
num_processes = cpu_count()
except NotImplementedError:
num_processes = 1
with AudioFile(audio_path, as_path=True) as wav_path:
data_set = split_audio_file(
wav_path,
batch_size=FLAGS.batch_size,
aggressiveness=FLAGS.vad_aggressiveness,
outlier_duration_ms=FLAGS.outlier_duration_ms,
outlier_batch_size=FLAGS.outlier_batch_size)
iterator = tf.data.Iterator.from_structure(
data_set.output_types,
data_set.output_shapes,
output_classes=data_set.output_classes)
batch_time_start, batch_time_end, batch_x, batch_x_len = iterator.get_next(
)
no_dropout = [None] * 6
logits, _ = create_model(batch_x=batch_x,
seq_length=batch_x_len,
dropout=no_dropout)
transposed = tf.nn.softmax(tf.transpose(logits, [1, 0, 2]))
tf.train.get_or_create_global_step()
with tf.Session(config=Config.session_config) as session:
if FLAGS.load == 'auto':
method_order = ['best', 'last']
else:
method_order = [FLAGS.load]
load_or_init_graph(session, method_order)
session.run(iterator.make_initializer(data_set))
transcripts = []
while True:
try:
starts, ends, batch_logits, batch_lengths = \
session.run([batch_time_start, batch_time_end, transposed, batch_x_len])
except tf.errors.OutOfRangeError:
break
decoded = ctc_beam_search_decoder_batch(
batch_logits,
batch_lengths,
Config.alphabet,
FLAGS.beam_width,
num_processes=num_processes,
scorer=scorer)
decoded = list(d[0][1] for d in decoded)
transcripts.extend(zip(starts, ends, decoded))
transcripts.sort(key=lambda t: t[0])
transcripts = [{
'start': int(start),
'end': int(end),
'transcript': transcript
} for start, end, transcript in transcripts]
with open(tlog_path, 'w') as tlog_file:
json.dump(transcripts, tlog_file, default=float)
def init(self, debug=False):
'''
'''
if debug:
sample = self.load_audio(self.audio_path)
self.audio = tf.placeholder(tf.float32, [None, 1])
features, features_len = samples_to_mfccs(self.audio,
FLAGS.audio_sample_rate,
train_phase=False)
if debug:
print(
'MFCC features:',
features.eval(feed_dict={self.audio: sample},
session=self.sess))
# Add batch dimension
features = tf.expand_dims(features, 0)
features_len = tf.expand_dims(features_len, 0)
# Evaluate
features = create_overlapping_windows(features)
if debug:
print(
'After creating overlapping windows:',
features.eval(feed_dict={self.audio: sample},
session=self.sess))
print(
'Feature length:',
features_len.eval(feed_dict={self.audio: sample},
session=self.sess))
#print('features.shape:', features.shape, ', features_len:', features_len)
previous_state_c = tf.zeros([1, Config.n_cell_dim])
previous_state_h = tf.zeros([1, Config.n_cell_dim])
previous_state = tf.nn.rnn_cell.LSTMStateTuple(previous_state_c,
previous_state_h)
# One rate per layer
no_dropout = [None] * 6
rnn_impl = rnn_impl_lstmblockfusedcell
self.logits, _ = create_model(batch_x=features,
batch_size=1,
seq_length=features_len,
dropout=no_dropout,
previous_state=previous_state,
overlap=False,
rnn_impl=rnn_impl)
# Load checkpoint from file
self.load_checkpoint()
def evaluate(test_csvs, create_model, try_loading):
scorer = Scorer(FLAGS.lm_alpha, FLAGS.lm_beta,
FLAGS.lm_binary_path, FLAGS.lm_trie_path,
Config.alphabet)
test_csvs = FLAGS.test_files.split(',')
test_sets = [create_dataset([csv], batch_size=FLAGS.test_batch_size) for csv in test_csvs]
iterator = tf.data.Iterator.from_structure(test_sets[0].output_types,
test_sets[0].output_shapes,
output_classes=test_sets[0].output_classes)
test_init_ops = [iterator.make_initializer(test_set) for test_set in test_sets]
(batch_x, batch_x_len), batch_y = iterator.get_next()
# One rate per layer
no_dropout = [None] * 6
logits, _ = create_model(batch_x=batch_x,
seq_length=batch_x_len,
dropout=no_dropout)
# Transpose to batch major and apply softmax for decoder
transposed = tf.nn.softmax(tf.transpose(logits, [1, 0, 2]))
loss = tf.nn.ctc_loss(labels=batch_y,
inputs=logits,
sequence_length=batch_x_len)
tf.train.get_or_create_global_step()
# Get number of accessible CPU cores for this process
try:
num_processes = cpu_count()
except NotImplementedError:
num_processes = 1
# Create a saver using variables from the above newly created graph
saver = tf.train.Saver()
with tf.Session(config=Config.session_config) as session:
# Restore variables from training checkpoint
loaded = try_loading(session, saver, 'best_dev_checkpoint', 'best validation')
if not loaded:
loaded = try_loading(session, saver, 'checkpoint', 'most recent')
if not loaded:
log_error('Checkpoint directory ({}) does not contain a valid checkpoint state.'.format(FLAGS.checkpoint_dir))
exit(1)
def run_test(init_op, dataset):
logitses = []
losses = []
seq_lengths = []
ground_truths = []
bar = create_progressbar(prefix='Computing acoustic model predictions | ',
widgets=['Steps: ', progressbar.Counter(), ' | ', progressbar.Timer()]).start()
log_progress('Computing acoustic model predictions...')
step_count = 0
# Initialize iterator to the appropriate dataset
session.run(init_op)
# First pass, compute losses and transposed logits for decoding
while True:
try:
logits, loss_, lengths, transcripts = session.run([transposed, loss, batch_x_len, batch_y])
except tf.errors.OutOfRangeError:
break
step_count += 1
bar.update(step_count)
logitses.append(logits)
losses.extend(loss_)
seq_lengths.append(lengths)
ground_truths.extend(sparse_tensor_value_to_texts(transcripts, Config.alphabet))
bar.finish()
predictions = []
bar = create_progressbar(max_value=step_count,
prefix='Decoding predictions | ').start()
log_progress('Decoding predictions...')
# Second pass, decode logits and compute WER and edit distance metrics
for logits, seq_length in bar(zip(logitses, seq_lengths)):
decoded = ctc_beam_search_decoder_batch(logits, seq_length, Config.alphabet, FLAGS.beam_width,
num_processes=num_processes, scorer=scorer)
predictions.extend(d[0][1] for d in decoded)
distances = [levenshtein(a, b) for a, b in zip(ground_truths, predictions)]
wer, cer, samples = calculate_report(ground_truths, predictions, distances, losses)
mean_loss = np.mean(losses)
# Take only the first report_count items
report_samples = itertools.islice(samples, FLAGS.report_count)
print('Test on %s - WER: %f, CER: %f, loss: %f' %
(dataset, wer, cer, mean_loss))
print('-' * 80)
for sample in report_samples:
print('WER: %f, CER: %f, loss: %f' %
(sample.wer, sample.distance, sample.loss))
print(' - src: "%s"' % sample.src)
print(' - res: "%s"' % sample.res)
print('-' * 80)
return samples
samples = []
for csv, init_op in zip(test_csvs, test_init_ops):
print('Testing model on {}'.format(csv))
samples.extend(run_test(init_op, dataset=csv))
return samples
def evaluate(test_csvs, create_model, try_loading):
if FLAGS.lm_binary_path:
scorer = Scorer(FLAGS.lm_alpha, FLAGS.lm_beta, FLAGS.lm_binary_path,
FLAGS.lm_trie_path, Config.alphabet)
else:
scorer = None
test_csvs = FLAGS.test_files.split(',')
test_sets = [
create_dataset([csv],
batch_size=FLAGS.test_batch_size,
train_phase=False) for csv in test_csvs
]
iterator = tfv1.data.Iterator.from_structure(
tfv1.data.get_output_types(test_sets[0]),
tfv1.data.get_output_shapes(test_sets[0]),
output_classes=tfv1.data.get_output_classes(test_sets[0]))
test_init_ops = [
iterator.make_initializer(test_set) for test_set in test_sets
]
batch_wav_filename, (batch_x, batch_x_len), batch_y = iterator.get_next()
# One rate per layer
no_dropout = [None] * 6
logits, _ = create_model(batch_x=batch_x,
batch_size=FLAGS.test_batch_size,
seq_length=batch_x_len,
dropout=no_dropout)
# Transpose to batch major and apply softmax for decoder
transposed = tf.nn.softmax(tf.transpose(a=logits, perm=[1, 0, 2]))
loss = tfv1.nn.ctc_loss(labels=batch_y,
inputs=logits,
sequence_length=batch_x_len)
tfv1.train.get_or_create_global_step()
# Get number of accessible CPU cores for this process
try:
num_processes = cpu_count()
except NotImplementedError:
num_processes = 1
# Create a saver using variables from the above newly created graph
saver = tfv1.train.Saver()
with tfv1.Session(config=Config.session_config) as session:
# Restore variables from training checkpoint
loaded = try_loading(session, saver, 'best_dev_checkpoint',
'best validation')
if not loaded:
loaded = try_loading(session, saver, 'checkpoint', 'most recent')
if not loaded:
log_error(
'Checkpoint directory ({}) does not contain a valid checkpoint state.'
.format(FLAGS.checkpoint_dir))
exit(1)
def run_test(init_op, dataset):
wav_filenames = []
losses = []
predictions = []
ground_truths = []
bar = create_progressbar(prefix='Test epoch | ',
widgets=[
'Steps: ',
progressbar.Counter(), ' | ',
progressbar.Timer()
]).start()
log_progress('Test epoch...')
step_count = 0
# Initialize iterator to the appropriate dataset
session.run(init_op)
# First pass, compute losses and transposed logits for decoding
while True:
try:
batch_wav_filenames, batch_logits, batch_loss, batch_lengths, batch_transcripts = \
session.run([batch_wav_filename, transposed, loss, batch_x_len, batch_y])
except tf.errors.OutOfRangeError:
break
decoded = ctc_beam_search_decoder_batch(
batch_logits,
batch_lengths,
Config.alphabet,
FLAGS.beam_width,
num_processes=num_processes,
scorer=scorer,
cutoff_prob=FLAGS.cutoff_prob,
cutoff_top_n=FLAGS.cutoff_top_n)
predictions.extend(d[0][1] for d in decoded)
ground_truths.extend(
sparse_tensor_value_to_texts(batch_transcripts,
Config.alphabet))
wav_filenames.extend(
wav_filename.decode('UTF-8')
for wav_filename in batch_wav_filenames)
losses.extend(batch_loss)
step_count += 1
bar.update(step_count)
bar.finish()
wer, cer, samples = calculate_report(wav_filenames, ground_truths,
predictions, losses)
mean_loss = np.mean(losses)
# Take only the first report_count items
report_samples = itertools.islice(samples, FLAGS.report_count)
print('Test on %s - WER: %f, CER: %f, loss: %f' %
(dataset, wer, cer, mean_loss))
print('-' * 80)
for sample in report_samples:
print('WER: %f, CER: %f, loss: %f' %
(sample.wer, sample.cer, sample.loss))
print(' - wav: file://%s' % sample.wav_filename)
print(' - src: "%s"' % sample.src)
print(' - res: "%s"' % sample.res)
print('-' * 80)
return samples
samples = []
for csv, init_op in zip(test_csvs, test_init_ops):
print('Testing model on {}'.format(csv))
samples.extend(run_test(init_op, dataset=csv))
return samples
def evaluate_with_pruning(test_csvs,
prune_percentage,
random,
scores_file,
result_file,
verbose=True,
skip_lstm=False):
'''Code originaly comes from the DeepSpeech repository (./DeepSpeech/evaluate.py).
The code is adapted for evaluation on pruned versions of the DeepSpeech model.
'''
tfv1.reset_default_graph()
if FLAGS.lm_binary_path:
scorer = Scorer(FLAGS.lm_alpha, FLAGS.lm_beta, FLAGS.lm_binary_path,
FLAGS.lm_trie_path, Config.alphabet)
else:
scorer = None
test_csvs = test_csvs.split(',')
test_sets = [
create_dataset([csv],
batch_size=FLAGS.test_batch_size,
train_phase=False) for csv in test_csvs
]
iterator = tfv1.data.Iterator.from_structure(
tfv1.data.get_output_types(test_sets[0]),
tfv1.data.get_output_shapes(test_sets[0]),
output_classes=tfv1.data.get_output_classes(test_sets[0]))
test_init_ops = [
iterator.make_initializer(test_set) for test_set in test_sets
]
batch_wav_filename, (batch_x, batch_x_len), batch_y = iterator.get_next()
# One rate per layer
no_dropout = [None] * 6
logits, _ = create_model(batch_x=batch_x,
batch_size=FLAGS.test_batch_size,
seq_length=batch_x_len,
dropout=no_dropout)
# Transpose to batch major and apply softmax for decoder
transposed = tf.nn.softmax(tf.transpose(a=logits, perm=[1, 0, 2]))
loss = tfv1.nn.ctc_loss(labels=batch_y,
inputs=logits,
sequence_length=batch_x_len)
tfv1.train.get_or_create_global_step()
# Get number of accessible CPU cores for this process
try:
num_processes = cpu_count()
except NotImplementedError:
num_processes = 1
# Create a saver using variables from the above newly created graph
saver = tfv1.train.Saver()
with tfv1.Session(config=Config.session_config) as session:
# Create a saver using variables from the above newly created graph
saver = tfv1.train.Saver()
# Restore variables from training checkpoint
loaded = False
if not loaded and FLAGS.load in ['auto', 'last']:
loaded = try_loading(session,
saver,
'checkpoint',
'most recent',
load_step=False)
if not loaded and FLAGS.load in ['auto', 'best']:
loaded = try_loading(session,
saver,
'best_dev_checkpoint',
'best validation',
load_step=False)
if not loaded:
print('Could not load checkpoint from {}'.format(
FLAGS.checkpoint_dir))
sys.exit(1)
###### PRUNING PART ######
if verbose:
if not prune_percentage: print('No pruning done.')
else:
if verbose: print('-' * 80)
if verbose: print('pruning with {}%...'.format(prune_percentage))
scores_per_layer = np.load(scores_file)
layer_masks = prune_matrices(scores_per_layer,
prune_percentage=prune_percentage,
random=random,
verbose=verbose,
skip_lstm=skip_lstm)
n_layers_to_prune = len(layer_masks)
i = 0
for index, v in enumerate(tf.trainable_variables()):
lstm_layer_name = 'cudnn_lstm/rnn/multi_rnn_cell/cell_0/cudnn_compatible_lstm_cell/kernel:0'
if 'weights' not in v.name and v.name != lstm_layer_name:
continue
if (i >= n_layers_to_prune):
break # if i < total_ops, it is not yet the last layer
# make mask into the shape of the weights
if v.name == lstm_layer_name:
if skip_lstm: continue
# Shape of LSTM weights: [(2*neurons), (4*neurons)]
cell_template = np.ones((2, 4))
mask = np.repeat(layer_masks[i], v.shape[0] // 2, axis=0)
mask = mask.reshape(
[layer_masks[i].shape[0], v.shape[0] // 2])
mask = np.swapaxes(mask, 0, 1)
mask = np.kron(mask, cell_template)
else:
idx = layer_masks[i] == 1
mask = np.repeat(layer_masks[i], v.shape[0], axis=0)
mask = mask.reshape([layer_masks[i].shape[0], v.shape[0]])
mask = np.swapaxes(mask, 0, 1)
# apply mask to weights
session.run(v.assign(tf.multiply(v, mask)))
i += 1
###### END PRUNING PART ######
def run_test(init_op, dataset):
wav_filenames = []
losses = []
predictions = []
ground_truths = []
bar = create_progressbar(prefix='Test epoch | ',
widgets=[
'Steps: ',
progressbar.Counter(), ' | ',
progressbar.Timer()
]).start()
log_progress('Test epoch...')
step_count = 0
# Initialize iterator to the appropriate dataset
session.run(init_op)
# First pass, compute losses and transposed logits for decoding
while True:
try:
batch_wav_filenames, batch_logits, batch_loss, batch_lengths, batch_transcripts = \
session.run([batch_wav_filename, transposed, loss, batch_x_len, batch_y])
except tf.errors.OutOfRangeError:
break
decoded = ctc_beam_search_decoder_batch(
batch_logits,
batch_lengths,
Config.alphabet,
FLAGS.beam_width,
num_processes=num_processes,
scorer=scorer,
cutoff_prob=FLAGS.cutoff_prob,
cutoff_top_n=FLAGS.cutoff_top_n)
predictions.extend(d[0][1] for d in decoded)
ground_truths.extend(
sparse_tensor_value_to_texts(batch_transcripts,
Config.alphabet))
wav_filenames.extend(
wav_filename.decode('UTF-8')
for wav_filename in batch_wav_filenames)
losses.extend(batch_loss)
step_count += 1
bar.update(step_count)
bar.finish()
wer, cer, samples = calculate_report(wav_filenames, ground_truths,
predictions, losses)
mean_loss = np.mean(losses)
# Take only the first report_count items
report_samples = itertools.islice(samples, FLAGS.report_count)
if verbose:
print('Test on %s - WER: %f, CER: %f, loss: %f' %
(dataset, wer, cer, mean_loss))
if verbose: print('-' * 80)
if result_file:
pruning_type = 'score-based' if not random else 'random'
result_string = '''Results for evaluating model with pruning percentage of {}% and {} pruning:
Test on {} - WER: {}, CER: {}, loss: {}
'''.format(prune_percentage * 100, pruning_type, dataset, wer,
cer, mean_loss)
write_to_file(result_file, result_string, 'a+')
return wer, cer, mean_loss
results = []
for csv, init_op in zip(test_csvs, test_init_ops):
if verbose: print('Testing model on {}'.format(csv))
results.extend(run_test(init_op, dataset=csv))
return results
def evaluate(test_csvs, create_model, try_loading):
scorer = Scorer(FLAGS.lm_alpha, FLAGS.lm_beta, FLAGS.lm_binary_path,
FLAGS.lm_trie_path, Config.alphabet)
test_set = create_dataset(test_csvs,
batch_size=FLAGS.test_batch_size,
cache_path=FLAGS.test_cached_features_path)
it = test_set.make_one_shot_iterator()
(batch_x, batch_x_len), batch_y = it.get_next()
# One rate per layer
no_dropout = [None] * 6
logits, _ = create_model(batch_x=batch_x,
seq_length=batch_x_len,
dropout=no_dropout)
# Transpose to batch major and apply softmax for decoder
transposed = tf.nn.softmax(tf.transpose(logits, [1, 0, 2]))
loss = tf.nn.ctc_loss(labels=batch_y,
inputs=logits,
sequence_length=batch_x_len)
global_step = tf.train.get_or_create_global_step()
with tf.Session(config=Config.session_config) as session:
# Create a saver using variables from the above newly created graph
saver = tf.train.Saver()
# Restore variables from training checkpoint
loaded = try_loading(session, saver, 'best_dev_checkpoint',
'best validation')
if not loaded:
loaded = try_loading(session, saver, 'checkpoint', 'most recent')
if not loaded:
log_error(
'Checkpoint directory ({}) does not contain a valid checkpoint state.'
.format(FLAGS.checkpoint_dir))
exit(1)
logitses = []
losses = []
seq_lengths = []
ground_truths = []
print('Computing acoustic model predictions...')
bar = progressbar.ProgressBar(widgets=[
'Steps: ',
progressbar.Counter(), ' | ',
progressbar.Timer()
])
step_count = 0
# First pass, compute losses and transposed logits for decoding
while True:
try:
logits, loss_, lengths, transcripts = session.run(
[transposed, loss, batch_x_len, batch_y])
except tf.errors.OutOfRangeError:
break
step_count += 1
bar.update(step_count)
logitses.append(logits)
losses.extend(loss_)
seq_lengths.append(lengths)
ground_truths.extend(
sparse_tensor_value_to_texts(transcripts, Config.alphabet))
bar.finish()
predictions = []
# Get number of accessible CPU cores for this process
try:
num_processes = cpu_count()
except:
num_processes = 1
print('Decoding predictions...')
bar = progressbar.ProgressBar(max_value=step_count,
widget=progressbar.AdaptiveETA)
# Second pass, decode logits and compute WER and edit distance metrics
for logits, seq_length in bar(zip(logitses, seq_lengths)):
decoded = ctc_beam_search_decoder_batch(logits,
seq_length,
Config.alphabet,
FLAGS.beam_width,
num_processes=num_processes,
scorer=scorer)
predictions.extend(d[0][1] for d in decoded)
distances = [levenshtein(a, b) for a, b in zip(ground_truths, predictions)]
wer, cer, samples = calculate_report(ground_truths, predictions, distances,
losses)
mean_loss = np.mean(losses)
# Take only the first report_count items
report_samples = itertools.islice(samples, FLAGS.report_count)
print('Test - WER: %f, CER: %f, loss: %f' % (wer, cer, mean_loss))
print('-' * 80)
for sample in report_samples:
print('WER: %f, CER: %f, loss: %f' %
(sample.wer, sample.distance, sample.loss))
print(' - src: "%s"' % sample.src)
print(' - res: "%s"' % sample.res)
print('-' * 80)
return samples
def evaluate(test_csvs, create_model):
if FLAGS.scorer_path:
scorer = Scorer(FLAGS.lm_alpha, FLAGS.lm_beta, FLAGS.scorer_path,
Config.alphabet)
else:
scorer = None
test_csvs = FLAGS.test_files.split(',')
test_sets = [
create_dataset([csv],
batch_size=FLAGS.test_batch_size,
train_phase=False) for csv in test_csvs
]
iterator = tfv1.data.Iterator.from_structure(
tfv1.data.get_output_types(test_sets[0]),
tfv1.data.get_output_shapes(test_sets[0]),
output_classes=tfv1.data.get_output_classes(test_sets[0]))
test_init_ops = [
iterator.make_initializer(test_set) for test_set in test_sets
]
batch_wav_filename, (batch_x, batch_x_len), batch_y = iterator.get_next()
# One rate per layer
no_dropout = [None] * 6
logits, _ = create_model(batch_x=batch_x,
batch_size=FLAGS.test_batch_size,
seq_length=batch_x_len,
dropout=no_dropout)
# Transpose to batch major and apply softmax for decoder
transposed = tf.nn.softmax(tf.transpose(a=logits, perm=[1, 0, 2]))
loss = tfv1.nn.ctc_loss(labels=batch_y,
inputs=logits,
sequence_length=batch_x_len)
tfv1.train.get_or_create_global_step()
# Get number of accessible CPU cores for this process
try:
num_processes = cpu_count()
except NotImplementedError:
num_processes = 1
with tfv1.Session(config=Config.session_config) as session:
if FLAGS.load == 'auto':
method_order = ['best', 'last']
else:
method_order = [FLAGS.load]
load_or_init_graph(session, method_order)
def run_test(init_op, dataset):
wav_filenames = []
losses = []
predictions = []
ground_truths = []
bar = create_progressbar(prefix='Test epoch | ',
widgets=[
'Steps: ',
progressbar.Counter(), ' | ',
progressbar.Timer()
]).start()
log_progress('Test epoch...')
step_count = 0
# Initialize iterator to the appropriate dataset
session.run(init_op)
# First pass, compute losses and transposed logits for decoding
while True:
try:
batch_wav_filenames, batch_logits, batch_loss, batch_lengths, batch_transcripts = \
session.run([batch_wav_filename, transposed, loss, batch_x_len, batch_y])
except tf.errors.OutOfRangeError:
break
decoded = ctc_beam_search_decoder_batch(
batch_logits,
batch_lengths,
Config.alphabet,
FLAGS.beam_width,
num_processes=num_processes,
scorer=scorer,
cutoff_prob=FLAGS.cutoff_prob,
cutoff_top_n=FLAGS.cutoff_top_n)
predictions.extend(d[0][1] for d in decoded)
ground_truths.extend(
sparse_tensor_value_to_texts(batch_transcripts,
Config.alphabet))
wav_filenames.extend(
wav_filename.decode('UTF-8')
for wav_filename in batch_wav_filenames)
losses.extend(batch_loss)
step_count += 1
bar.update(step_count)
bar.finish()
# Print test summary
test_samples = calculate_and_print_report(wav_filenames,
ground_truths,
predictions, losses,
dataset)
return test_samples
samples = []
for csv, init_op in zip(test_csvs, test_init_ops):
print('Testing model on {}'.format(csv))
samples.extend(run_test(init_op, dataset=csv))
return samples
