def do_single_file_inference(input_file_path):
with tfv1.Session(config=Config.session_config) as session:
inputs, outputs, _ = create_inference_graph(batch_size=1, n_steps=-1)
# Create a saver using variables from the above newly created graph
saver = tfv1.train.Saver()
# Restore variables from training checkpoint
# TODO: This restores the most recent checkpoint, but if we use validation to counteract
# over-fitting, we may want to restore an earlier checkpoint.
checkpoint = tf.train.get_checkpoint_state(FLAGS.checkpoint_dir)
if not checkpoint:
log_error(
'Checkpoint directory ({}) does not contain a valid checkpoint state.'
.format(FLAGS.checkpoint_dir))
exit(1)
checkpoint_path = checkpoint.model_checkpoint_path
saver.restore(session, checkpoint_path)
features, features_len = audiofile_to_features(input_file_path)
previous_state_c = np.zeros([1, Config.n_cell_dim])
previous_state_h = np.zeros([1, Config.n_cell_dim])
# Add batch dimension
features = tf.expand_dims(features, 0)
features_len = tf.expand_dims(features_len, 0)
# Evaluate
features = create_overlapping_windows(features).eval(session=session)
features_len = features_len.eval(session=session)
logits = outputs['outputs'].eval(feed_dict={
inputs['input']:
features,
inputs['input_lengths']:
features_len,
inputs['previous_state_c']:
previous_state_c,
inputs['previous_state_h']:
previous_state_h,
},
session=session)
logits = np.squeeze(logits)
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
decoded = ctc_beam_search_decoder(logits,
Config.alphabet,
FLAGS.beam_width,
scorer=scorer,
cutoff_prob=FLAGS.cutoff_prob,
cutoff_top_n=FLAGS.cutoff_top_n)
# Print highest probability result
print(decoded[0][1])
def do_single_file_inference(input_file_path):
with tf.Session(config=Config.session_config) as session:
inputs, outputs, _ = create_inference_graph(batch_size=1, n_steps=-1)
# Create a saver using variables from the above newly created graph
mapping = {
v.op.name: v
for v in tf.global_variables()
if not v.op.name.startswith('previous_state_')
}
saver = tf.train.Saver(mapping)
# Restore variables from training checkpoint
# TODO: This restores the most recent checkpoint, but if we use validation to counteract
# over-fitting, we may want to restore an earlier checkpoint.
checkpoint = tf.train.get_checkpoint_state(FLAGS.checkpoint_dir)
if not checkpoint:
log_error(
'Checkpoint directory ({}) does not contain a valid checkpoint state.'
.format(FLAGS.checkpoint_dir))
exit(1)
checkpoint_path = checkpoint.model_checkpoint_path
saver.restore(session, checkpoint_path)
session.run(outputs['initialize_state'])
features, features_len = audiofile_to_features(input_file_path)
# Add batch dimension
features = tf.expand_dims(features, 0)
features_len = tf.expand_dims(features_len, 0)
# Evaluate
features = create_overlapping_windows(features).eval(session=session)
features_len = features_len.eval(session=session)
logits = outputs['outputs'].eval(feed_dict={
inputs['input']: features,
inputs['input_lengths']: features_len,
},
session=session)
logits = np.squeeze(logits)
scorer = Scorer(FLAGS.lm_alpha, FLAGS.lm_beta, FLAGS.lm_binary_path,
FLAGS.lm_trie_path, Config.alphabet)
decoded = ctc_beam_search_decoder(logits,
Config.alphabet,
FLAGS.beam_width,
scorer=scorer)
# Print highest probability result
print(decoded[0][1])
def do_single_file_inference(input_file_path):
with tfv1.Session(config=Config.session_config) as session:
inputs, outputs, _ = create_inference_graph(batch_size=1, n_steps=-1)
# Restore variables from training checkpoint
if FLAGS.load == 'auto':
method_order = ['best', 'last']
else:
method_order = [FLAGS.load]
load_or_init_graph(session, method_order)
features, features_len = audiofile_to_features(input_file_path)
previous_state_c = np.zeros([1, Config.n_cell_dim])
previous_state_h = np.zeros([1, Config.n_cell_dim])
# Add batch dimension
features = tf.expand_dims(features, 0)
features_len = tf.expand_dims(features_len, 0)
# Evaluate
features = create_overlapping_windows(features).eval(session=session)
features_len = features_len.eval(session=session)
logits = outputs['outputs'].eval(feed_dict={
inputs['input']:
features,
inputs['input_lengths']:
features_len,
inputs['previous_state_c']:
previous_state_c,
inputs['previous_state_h']:
previous_state_h,
},
session=session)
logits = np.squeeze(logits)
if FLAGS.scorer_path:
scorer = Scorer(FLAGS.lm_alpha, FLAGS.lm_beta, FLAGS.scorer_path,
Config.alphabet)
else:
scorer = None
decoded = ctc_beam_search_decoder(logits,
Config.alphabet,
FLAGS.beam_width,
scorer=scorer,
cutoff_prob=FLAGS.cutoff_prob,
cutoff_top_n=FLAGS.cutoff_top_n)
# Print highest probability result
print(decoded[0][1])
def do_single_file_inference(input_file_path):
with tfv1.Session(config=Config.session_config) as session:
inputs, outputs, _ = create_inference_graph(batch_size=1, n_steps=-1)
# 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)
features, features_len = audiofile_to_features(input_file_path)
previous_state_c = np.zeros([1, Config.n_cell_dim])
previous_state_h = np.zeros([1, Config.n_cell_dim])
# Add batch dimension
features = tf.expand_dims(features, 0)
features_len = tf.expand_dims(features_len, 0)
# Evaluate
features = create_overlapping_windows(features).eval(session=session)
features_len = features_len.eval(session=session)
logits = outputs['outputs'].eval(feed_dict={
inputs['input']: features,
inputs['input_lengths']: features_len,
inputs['previous_state_c']: previous_state_c,
inputs['previous_state_h']: previous_state_h,
}, session=session)
logits = np.squeeze(logits)
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
decoded = ctc_beam_search_decoder(logits, Config.alphabet, FLAGS.beam_width,
scorer=scorer, cutoff_prob=FLAGS.cutoff_prob,
cutoff_top_n=FLAGS.cutoff_top_n)
# Print highest probability result
print(decoded[0][1])
def activations_common_voice_pertubed_sets(input_dir,
output_dir,
test_only=False,
prune_percentage=0,
scores_file=None,
random=False,
verbose=True,
randomly_initialized=False):
'''Obtains activations for wavs in input_dir and saves them to output_dir'''
inputs, outputs, layers = create_inference_graph(batch_size=1, n_steps=-1)
intermediate_layer_names = [
'layer_1', 'layer_2', 'layer_3', 'rnn_output', 'layer_4', 'layer_5'
]
intermediate_layers = [
l for n, l in layers.items() if n in intermediate_layer_names
]
pertubed_sets = json.load(open('data/pertubed_input_sets_balanced.json'))
skip_sets = []
if test_only: skip_sets = json.load(open('./results/set_ids_used.json'))
if not prune_percentage: base_path = '{}/activations'.format(output_dir)
else:
base_path = '{}/activations/pruned-{}'.format(output_dir,
prune_percentage * 100)
if random: base_path += '-random'
with tfv1.Session(config=Config.session_config) as session:
# Create a saver using variables from the above newly created graph
if not randomly_initialized:
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)
else:
initializer = tfv1.global_variables_initializer()
session.run(initializer)
###### 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=False)
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 ######
# Default states for LSTM cell
previous_state_c = np.zeros([1, Config.n_cell_dim])
previous_state_h = np.zeros([1, Config.n_cell_dim])
sets_to_process = [
set for set in pertubed_sets if str(set['set_id']) not in skip_sets
]
print('{} sets found'.format(len(sets_to_process)))
for set in sets_to_process:
print('Processing set {}, {} items...'.format(
set['set_id'], set['set_length']))
# Only process files that are not yet available in results directory
create_dir_if_not_exists('{}/{}'.format(
base_path, set['set_id'])) # Check if directory exists
files_done = [
f[:-4]
for f in os.listdir('{}/{}'.format(base_path, set['set_id']))
if f.endswith('.npy')
]
for item in set['set_items']:
file_name = item['path'][:-4]
print(file_name)
if file_name in files_done:
print('Skipped.')
continue
print('current file: {}'.format(file_name))
input_file_path = '{}/{}.wav'.format(input_dir, file_name)
# Prepare features
features, features_len = audiofile_to_features(input_file_path)
features = tf.expand_dims(features, 0)
features_len = tf.expand_dims(features_len, 0)
features = create_overlapping_windows(features).eval(
session=session)
features_len = features_len.eval(session=session)
feed_dict = {
inputs['input']: features,
inputs['input_lengths']: features_len,
inputs['previous_state_c']: previous_state_c,
inputs['previous_state_h']: previous_state_h,
}
intermediate_activations = session.run(intermediate_layers,
feed_dict=feed_dict)
# Save activations of actual input
save_to_path_activations = '{}/{}/{}.npy'.format(
base_path, set['set_id'], file_name)
write_numpy_to_file(save_to_path_activations,
np.array(intermediate_activations))
print('Activations for {} are saved to: {}'.format(
file_name, save_to_path_activations))
return True
