def main(_):
initialize_globals()
input_dir = './data/CommonVoice/pertubed_sets'
output_dir = './results/randomly-initialized'
tfv1.reset_default_graph()
activations_common_voice_pertubed_sets(input_dir=input_dir,
output_dir=output_dir,
test_only=True,
randomly_initialized=True)
output_dir = './results'
# Obtain activations for all sets without pruning of common voice test set
tfv1.reset_default_graph()
activations_common_voice_pertubed_sets(input_dir=input_dir,
output_dir=output_dir)
# Obtain activations for non-training sets with pruning of common voice test set
tfv1.reset_default_graph()
activations_peractivations_common_voice_pertubed_setstubed_sets(
input_dir=input_dir,
output_dir=output_dir,
test_only=True,
prune_percentage=.1,
scores_file='./results/activations_combined.npy')
# Obtain activations for non-training sets with pruning of common voice test set
tfv1.reset_default_graph()
activations_peractivations_common_voice_pertubed_setstubed_sets(
input_dir=input_dir,
output_dir=output_dir,
test_only=True,
random=True,
prune_percentage=.1,
scores_file='./results/activations_combined.npy')
# Obtain activations for all sets without pruning of librispeech validation set
tfv1.reset_default_graph()
activations_libri_speech_test_set(input_dir=input_dir,
output_dir=output_dir)
# Obtain activations for non-training sets with pruning of librispeech validation set
tfv1.reset_default_graph()
activations_libri_speech_test_set(
input_dir=input_dir,
output_dir=output_dir,
test_only=True,
prune_percentage=.1,
scores_file='./results/activations_combined.npy')
# Obtain activations for non-training sets with random pruning of librispeech validation set
tfv1.reset_default_graph()
activations_libri_speech_test_set(
input_dir=input_dir,
output_dir=output_dir,
test_only=True,
random=True,
prune_percentage=.1,
scores_file='./results/activations_combined.npy')
def main(_):
initialize_globals()
if FLAGS.train or FLAGS.test:
if len(FLAGS.worker_hosts) == 0:
# Only one local task: this process (default case - no cluster)
with tf.Graph().as_default():
train()
if Config.is_chief:
export()
# Now do a final test epoch
if FLAGS.test:
with tf.Graph().as_default():
test()
log_debug('Done.')
else:
# Create and start a server for the local task.
server = tf.train.Server(Config.cluster,
job_name=FLAGS.job_name,
task_index=FLAGS.task_index)
if FLAGS.job_name == 'ps':
# We are a parameter server and therefore we just wait for all workers to finish
# by waiting for their stop tokens.
with tf.Session(server.target) as session:
for worker in FLAGS.worker_hosts:
log_debug('Waiting for stop token...')
token = session.run(
Config.done_dequeues[FLAGS.task_index])
if token < 0:
log_debug(
'Got a kill switch token from worker %i.' %
abs(token + 1))
break
log_debug('Got a stop token from worker %i.' % token)
log_debug('Session closed.')
if FLAGS.test:
test()
elif FLAGS.job_name == 'worker':
# We are a worker and therefore we have to do some work.
# Assigns ops to the local worker by default.
with tf.device(
tf.train.replica_device_setter(
worker_device=Config.worker_device,
cluster=Config.cluster)):
# Do the training
train(server)
log_debug('Server stopped.')
# Are we the main process?
if Config.is_chief:
# Doing solo/post-processing work just on the main process...
# Exporting the model
if FLAGS.export_dir:
export()
if len(FLAGS.one_shot_infer):
do_single_file_inference(FLAGS.one_shot_infer)
def main(_):
initialize_globals()
if FLAGS.train_files:
tfv1.reset_default_graph()
tfv1.set_random_seed(FLAGS.random_seed)
train()
if FLAGS.test_files:
tfv1.reset_default_graph()
test()
if FLAGS.export_dir and not FLAGS.export_zip:
tfv1.reset_default_graph()
export()
if FLAGS.export_zip:
tfv1.reset_default_graph()
FLAGS.export_tflite = True
if os.listdir(FLAGS.export_dir):
log_error('Directory {} is not empty, please fix this.'.format(FLAGS.export_dir))
sys.exit(1)
export()
package_zip()
if FLAGS.one_shot_infer:
tfv1.reset_default_graph()
do_single_file_inference(FLAGS.one_shot_infer)
def total_inference(input_folder, output_folder, checkpoint_dir, layer_wanted,
softmax_wanted, win_size_s, stride_size_s, fea_format,
csv_format):
initialize_globals()
for root, dirs, files in os.walk(input_folder):
nb = len(files)
it = 0
# Iterate over files
for filename in files:
print(filename)
if ((it + 1) % 100) == 0:
print(it, 'on', nb)
it += 1
if not filename.endswith('.wav'):
continue
else:
full_name = os.path.join(root.lstrip('./'), filename)
tf.reset_default_graph()
do_single_file_inference(checkpoint_dir=checkpoint_dir,
input_file_path='/' + full_name,
layer_wanted=layer_wanted,
win_size_s=win_size_s,
stride_size_s=stride_size_s,
save_folder=output_folder,
save_filename=filename[:-4],
softmax_wanted=softmax_wanted,
fea_format=fea_format,
csv_format=csv_format)
def main(_):
initialize_globals()
if not FLAGS.test_files:
log_error('You need to specify what files to use for evaluation via '
'the --test_files flag.')
exit(1)
global alphabet
alphabet = Alphabet(FLAGS.alphabet_config_path)
# sort examples by length, improves packing of batches and timesteps
test_data = preprocess(
FLAGS.test_files.split(','),
FLAGS.test_batch_size,
alphabet=alphabet,
numcep=Config.n_input,
numcontext=Config.n_context,
hdf5_cache_path=FLAGS.hdf5_test_set).sort_values(
by="features_len",
ascending=False)
from DeepSpeech import create_inference_graph
graph = create_inference_graph(batch_size=FLAGS.test_batch_size, n_steps=-1)
samples = evaluate(test_data, graph, alphabet)
if FLAGS.test_output_file:
# Save decoded tuples as JSON, converting NumPy floats to Python floats
json.dump(samples, open(FLAGS.test_output_file, 'w'), default=lambda x: float(x))
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 main(_):
initialize_globals()
if FLAGS.cmd != 'classify' and FLAGS.cmd != 'attack':
print('Unsupport command, please check your cmd parameter')
else:
adv_ds = AdvDeepSpeech(FLAGS.model_path, FLAGS.input)
if FLAGS.cmd == 'classify':
adv_ds.classify2(FLAGS.input)
else:
adv_ds.attack(FLAGS.iteration)
def main(_):
initialize_globals()
# Evaluate on:
# - files used for pruning (per set prune or total prune)
# - original test set for model
pertubed_sets = json.load(open('data/pertubed_input_sets_balanced.json'))
train_sets = json.load(open('./results/set_ids_used.json'))
common_voice_info = get_file_info('./data/common-voice-pertubed_sets.csv')
file_info = []
for set in pertubed_sets:
if str(set['set_id']) in train_sets: continue
for item in set['set_items']:
filename = item['path'][:-4]
file_info.append(common_voice_info[filename])
print('{} test files found...'.format(len(file_info)))
# Clean up characters in case they are in the transcript
not_allowed = [',', '.', '!', '?', '"', '-', ':', ';']
for info in file_info:
if any(c in info['transcript'] for c in not_allowed):
for c in not_allowed:
info['transcript'] = info['transcript'].replace(c, '')
# Prune on all pertubed sets combined and evaluate on test set
percents = [0, .05, .1, .2]
print('Evaluating pruning on common voice test set')
evaluation = {}
for percent in percents:
results = evaluate(scores_path='./results/activations_combined.npy',
prune_percent=percent,
evaluate_files=file_info)
evaluation['{}'.format(percent)] = results
print(results)
json.dump(evaluation,
open('./results/evaluations_all_pertubated_sets.json', 'w+'))
# Prune and evaluate on original test set
percents = [0, .1]
print('Evaluating pruning on original test set')
file_info = get_file_info('./data/librivox-test-clean.csv')
evaluation = {}
for percent in percents:
results = evaluate(scores_path='./results/activations_combined.npy',
prune_percent=percent,
evaluate_files=file_info.values())
evaluation['{}'.format(percent)] = results
print(results)
json.dump(evaluation,
open('./results/evaluations_original_test_set.json', 'w+'))
def main(_):
initialize_globals()
if FLAGS.train or FLAGS.test:
if len(FLAGS.worker_hosts) == 0:
# Only one local task: this process (default case - no cluster)
with tf.Graph().as_default():
tf.set_random_seed(FLAGS.random_seed)
train()
# Now do a final test epoch
if FLAGS.test:
with tf.Graph().as_default():
test()
log_debug('Done.')
else:
# Create and start a server for the local task.
server = tf.train.Server(Config.cluster, job_name=FLAGS.job_name, task_index=FLAGS.task_index)
if FLAGS.job_name == 'ps':
# We are a parameter server and therefore we just wait for all workers to finish
# by waiting for their stop tokens.
with tf.Session(server.target) as session:
for worker in FLAGS.worker_hosts:
log_debug('Waiting for stop token...')
token = session.run(Config.done_dequeues[FLAGS.task_index])
if token < 0:
log_debug('Got a kill switch token from worker %i.' % abs(token + 1))
break
log_debug('Got a stop token from worker %i.' % token)
log_debug('Session closed.')
if FLAGS.test:
test()
elif FLAGS.job_name == 'worker':
# We are a worker and therefore we have to do some work.
# Assigns ops to the local worker by default.
with tf.device(tf.train.replica_device_setter(
worker_device=Config.worker_device,
cluster=Config.cluster)):
# Do the training
train(server)
log_debug('Server stopped.')
# Are we the main process?
if Config.is_chief:
# Doing solo/post-processing work just on the main process...
# Exporting the model
if FLAGS.export_dir:
export()
if len(FLAGS.one_shot_infer):
do_single_file_inference(FLAGS.one_shot_infer)
def main(_):
initialize_globals()
if not FLAGS.test_files:
log_error('You need to specify what files to use for evaluation via '
'the --test_files flag.')
exit(1)
from DeepSpeech import create_model, try_loading # pylint: disable=cyclic-import
samples = evaluate(FLAGS.test_files.split(','), create_model, try_loading)
if FLAGS.test_output_file:
# Save decoded tuples as JSON, converting NumPy floats to Python floats
json.dump(samples, open(FLAGS.test_output_file, 'w'), default=float)
def main(_):
initialize_globals()
if not FLAGS.test_files:
log_error('You need to specify what files to use for evaluation via '
'the --test_files flag.')
exit(1)
from DeepSpeech import create_model, try_loading # pylint: disable=cyclic-import
samples = evaluate(FLAGS.test_files.split(','), create_model, try_loading)
if FLAGS.test_output_file:
# Save decoded tuples as JSON, converting NumPy floats to Python floats
json.dump(samples, open(FLAGS.test_output_file, 'w'), default=float)
def main(_):
initialize_globals()
evaluation_csv = './data/librivox-test-clean.csv'
results_file = './results/evaluation_output.txt'
scores_file = './results/final_imp_scores.npy'
for prune_settings in [(0, False), (.05, False), (.05, True), (.1, False),
(.1, True), (.15, False), (.15, True)]:
evaluate_with_pruning(evaluation_csv,
create_model,
try_loading,
prune_settings[0],
random=prune_settings[1],
scores_file=scores_file,
result_file=results_file)
def main(_):
initialize_globals()
if not FLAGS.test_files:
log_error('You need to specify what files to use for evaluation via '
'the --test_files flag.')
sys.exit(1)
is_character_based = character_based()
study = optuna.create_study()
study.set_user_attr("is_character_based", is_character_based)
study.optimize(objective, n_jobs=1, n_trials=FLAGS.n_trials)
print('Best params: lm_alpha={} and lm_beta={} with WER={}'.format(
study.best_params['lm_alpha'], study.best_params['lm_beta'],
study.best_value))
def main(_):
initialize_globals()
if FLAGS.train:
with tf.Graph().as_default():
tf.set_random_seed(FLAGS.random_seed)
train()
if FLAGS.test:
with tf.Graph().as_default():
test()
if FLAGS.export_dir:
export()
if len(FLAGS.one_shot_infer):
do_single_file_inference(FLAGS.one_shot_infer)
def main(_):
initialize_globals()
if FLAGS.train_files:
tfv1.reset_default_graph()
tfv1.set_random_seed(FLAGS.random_seed)
train()
if FLAGS.test_files:
tfv1.reset_default_graph()
test()
if FLAGS.export_dir:
tfv1.reset_default_graph()
export()
if FLAGS.one_shot_infer:
tfv1.reset_default_graph()
do_single_file_inference(FLAGS.one_shot_infer)
def main(_):
initialize_globals()
if FLAGS.train or FLAGS.test:
# Only one local task: this process (default case - no cluster)
with tf.Graph().as_default():
tf.set_random_seed(FLAGS.random_seed)
train()
# Now do a final test epoch
if FLAGS.test:
with tf.Graph().as_default():
test()
log_debug('Done.')
# Are we the main process?
if Config.is_chief:
# Doing solo/post-processing work just on the main process...
# Exporting the model
if FLAGS.export_dir:
export()
def main(_):
initialize_globals()
if not FLAGS.test_files:
log_error('You need to specify what files to use for evaluation via '
'the --test_files flag.')
exit(1)
#if FLAGS.embeddings_output_dir:
# prefix = FLAGS.embeddings_output_dir
# print('Prefix :', prefix)
# #print('LAYER4 :', LAYER4)
# EMBEDDINGS = prefix + 'embeddings/'
# LAYER4 = EMBEDDINGS + 'layer4/'
# LAYER5 = EMBEDDINGS + 'layer5/'
# LAYER6 = EMBEDDINGS + 'layer6/'
# c.TEXT = EMBEDDINGS + 'text/'
# print('LAYER4 :', LAYER4)
# sort examples by length, improves packing of batches and timesteps
test_data = preprocess(FLAGS.test_files.split(','),
FLAGS.test_batch_size,
alphabet=Config.alphabet,
numcep=Config.n_input,
numcontext=Config.n_context,
hdf5_cache_path=FLAGS.hdf5_test_set).sort_values(
by="features_len", ascending=False)
#print('test_data', test_data)
#print(test_data.fname[1])
#return 1
#print(test_data[0].fname)
print('Batch Size: ', FLAGS.test_batch_size)
from DeepSpeech import create_inference_graph
graph = create_inference_graph(batch_size=FLAGS.test_batch_size,
n_steps=-1)
samples = evaluate(test_data, graph)
if FLAGS.test_output_file:
# Save decoded tuples as JSON, converting NumPy floats to Python floats
json.dump(samples,
open(FLAGS.test_output_file, 'w'),
default=lambda x: float(x))
def load_graph(frozen_graph_filename=GRAPH_PB_PATH):
# We load the protobuf file from the disk and parse it to retrieve the
# unserialized graph_def
with tf.gfile.GFile(frozen_graph_filename, "rb") as f:
graph_def = tf.GraphDef()
graph_def.ParseFromString(f.read())
return graph_def
if __name__ == '__main__':
# Initialize Globals.
create_flags()
initialize_globals()
# Load pretrained model.
graph_def = load_graph()
with tf.Graph().as_default() as graph:
# The name var will prefix every op/nodes in your graph
# Since we load everything in a new graph, this is not needed
tf.import_graph_def(graph_def, name="prefix")
# Open tf.Session.
with tf.Session(graph=graph) as sess:
# Extract graph node names.
tf.import_graph_def(graph_def, name='')
graph_nodes = [n for n in graph_def.node]
names = []
def main(_):
initialize_globals()
if FLAGS.train or FLAGS.test:
if len(FLAGS.worker_hosts) == 0:
# Only one local task: this process (default case - no cluster)
#with tf.Graph().as_default():
#train()
#if Config.is_chief:
# export()
# Now do a final test epoch
if FLAGS.test:
print("$$$$$$$$$ Testing on entire test dataset $$$$$$$$$$")
ckpt_files = [
f for f in sorted(os.listdir(FLAGS.checkpoint_dir))
if os.path.isfile(os.path.join(FLAGS.checkpoint_dir, f))
and '.meta' in f
]
for ckpt_file in ckpt_files:
print("************* Testing on ckpt file: " + ckpt_file +
" ***************")
with tf.Graph().as_default():
test(ckpt_file.replace(".meta", ""), FLAGS.test_files)
log_debug('Done.')
for test_file in FLAGS.test_files.split(","):
print("$$$$$$$$$ Testing on " + test_file +
" dataset $$$$$$$$$$")
ckpt_files = [
f for f in sorted(os.listdir(FLAGS.checkpoint_dir))
if os.path.isfile(os.path.join(FLAGS.checkpoint_dir,
f)) and '.meta' in f
]
for ckpt_file in ckpt_files:
print("************* Testing on ckpt file: " +
ckpt_file + " ***************")
with tf.Graph().as_default():
test(ckpt_file.replace(".meta", ""), test_file)
log_debug('Done.')
else:
# Create and start a server for the local task.
server = tf.train.Server(Config.cluster,
job_name=FLAGS.job_name,
task_index=FLAGS.task_index)
if FLAGS.job_name == 'ps':
# We are a parameter server and therefore we just wait for all workers to finish
# by waiting for their stop tokens.
with tf.Session(server.target) as session:
for worker in FLAGS.worker_hosts:
log_debug('Waiting for stop token...')
token = session.run(
Config.done_dequeues[FLAGS.task_index])
if token < 0:
log_debug(
'Got a kill switch token from worker %i.' %
abs(token + 1))
break
log_debug('Got a stop token from worker %i.' % token)
log_debug('Session closed.')
if FLAGS.test:
test()
elif FLAGS.job_name == 'worker':
# We are a worker and therefore we have to do some work.
# Assigns ops to the local worker by default.
with tf.device(
tf.train.replica_device_setter(
worker_device=Config.worker_device,
cluster=Config.cluster)):
# Do the training
train(server)
log_debug('Server stopped.')
# Are we the main process?
#if Config.is_chief:
# Doing solo/post-processing work just on the main process...
# Exporting the model
#if FLAGS.export_dir:
#export()
if len(FLAGS.one_shot_infer):
do_single_file_inference(FLAGS.one_shot_infer)