global n_input
n_input = 40 # MFCC, maybe need add Delta
# The number of frames in the context
global n_context
n_context = 0
global feature_len
feature_len = 100 # all input is 1s wavfiel 1000ms/10ms = 100.
global feature_dim
feature_dim = n_input * (n_context * 2 + 1)
global alphabet
alphabet = Alphabet('./alphabet.txt')
print('alphabet.size() ', alphabet.size())
print(alphabet._label_to_str)
# The number of characters in the target language plus one
global n_character
n_character = alphabet.size() + 1 # +1 for CTC blank label
global max_labellen
max_labellen = 6
global n_hidden
n_hidden = 128
trfile = 'data/speechcmd_train.csv'
cvfile = 'data/speechcmd_dev.csv'
testfile = 'data/speechcmd_test.csv'
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(os.path.abspath(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=N_FEATURES,
numcontext=N_CONTEXT,
hdf5_cache_path=FLAGS.hdf5_test_set).sort_values(
by="features_len", ascending=False)
def create_windows(features):
num_strides = len(features) - (N_CONTEXT * 2)
# Create a view into the array with overlapping strides of size
# numcontext (past) + 1 (present) + numcontext (future)
window_size = 2 * N_CONTEXT + 1
features = np.lib.stride_tricks.as_strided(
features, (num_strides, window_size, N_FEATURES),
(features.strides[0], features.strides[0], features.strides[1]),
writeable=False)
return features
test_data['features'] = test_data['features'].apply(create_windows)
with tf.Session() as session:
inputs, outputs = create_inference_graph(
batch_size=FLAGS.test_batch_size, n_steps=N_STEPS)
seq_lengths_ph = tf.placeholder(tf.int32, [FLAGS.test_batch_size])
decode_logits_ph = tf.placeholder(
tf.float32, [None, FLAGS.test_batch_size,
alphabet.size() + 1])
labels_ph = tf.placeholder(tf.int32, [FLAGS.test_batch_size, None])
label_lengths_ph = tf.placeholder(tf.int32, [FLAGS.test_batch_size])
decoded, _ = decode_with_lm(decode_logits_ph,
seq_lengths_ph,
merge_repeated=False,
beam_width=FLAGS.beam_width)
sparse_labels = tf.cast(
ctc_label_dense_to_sparse(labels_ph, label_lengths_ph,
FLAGS.test_batch_size), tf.int32)
loss = tf.nn.ctc_loss(labels=sparse_labels,
inputs=decode_logits_ph,
sequence_length=seq_lengths_ph)
distance = tf.edit_distance(tf.cast(decoded[0], tf.int32),
sparse_labels)
# 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
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)
logitses = []
batch_count = len(test_data) // FLAGS.test_batch_size
bar = progressbar.ProgressBar(max_value=batch_count - 1,
widget=progressbar.AdaptiveETA)
for batch in bar(split_data(test_data, FLAGS.test_batch_size)):
session.run(outputs['initialize_state'])
batch_features = pad_to_dense(batch['features'].values)
batch_features_len = batch['features_len'].values
full_step_len = np.full_like(batch_features_len, N_STEPS)
logits = np.empty([0, FLAGS.test_batch_size, alphabet.size() + 1])
for i in range(0, batch_features.shape[1], N_STEPS):
chunk_features = batch_features[:, i:i + N_STEPS, :, :]
chunk_features_len = np.minimum(batch_features_len,
full_step_len)
# pad with zeros if the chunk does not have enough steps
steps_in_chunk = chunk_features.shape[1]
if steps_in_chunk < FLAGS.n_steps:
chunk_features = np.pad(
chunk_features,
((0, 0), (0, FLAGS.n_steps - steps_in_chunk), (0, 0),
(0, 0)),
mode='constant',
constant_values=0)
output = session.run(outputs['outputs'],
feed_dict={
inputs['input']:
chunk_features,
inputs['input_lengths']:
chunk_features_len,
})
logits = np.concatenate((logits, output))
# we have processed N_STEPS so subtract from remaining steps
batch_features_len -= N_STEPS
# clip to zero
batch_features_len = np.maximum(
batch_features_len, np.zeros_like(batch_features_len))
logitses.append(logits)
ground_truths = []
predictions = []
distances = []
losses = []
bar = progressbar.ProgressBar(max_value=batch_count - 1,
widget=progressbar.AdaptiveETA)
for logits, batch in bar(
zip(logitses, split_data(test_data, FLAGS.test_batch_size))):
seq_lengths = batch['features_len'].values
labels = pad_to_dense(batch['transcript'].values)
label_lengths = batch['transcript_len'].values
decoded_, loss_, distance_, sparse_labels_ = session.run(
[decoded, loss, distance, sparse_labels],
feed_dict={
decode_logits_ph: logits,
seq_lengths_ph: seq_lengths,
labels_ph: labels,
label_lengths_ph: label_lengths
})
ground_truths.extend(
sparse_tensor_value_to_texts(sparse_labels_, alphabet))
predictions.extend(
sparse_tensor_value_to_texts(decoded_[0], alphabet))
distances.extend(distance_)
losses.extend(loss_)
wer, samples = calculate_report(ground_truths, predictions, distances,
losses)
mean_edit_distance = np.mean(distances)
mean_loss = np.mean(losses)
# Filter out all items with WER=0 and take only the first report_count items
report_samples = itertools.islice((s for s in samples if s.wer > 0),
FLAGS.report_count)
print('Test - WER: %f, loss: %f, mean edit distance: %f' %
(wer, mean_loss, mean_edit_distance))
print('-' * 80)
for sample in report_samples:
print('WER: %f, loss: %f, mean edit distance: %f' %
(sample.wer, sample.loss, sample.distance))
print(' - src: "%s"' % sample.src)
print(' - res: "%s"' % sample.res)
print('-' * 80)
if FLAGS.test_output_file:
json.dump(samples,
open(FLAGS.test_output_file, 'w'),
default=lambda x: float(x))
def run_inference():
"""Load frozen graph, run inference and display most likely predicted characters"""
parser = argparse.ArgumentParser(
description='Run Deepspeech inference to obtain char probabilities')
parser.add_argument('--input-file',
type=str,
help='Path to the wav file',
action="store",
dest="input_file_path")
parser.add_argument('--alphabet-file',
type=str,
help='Path to the alphabet.txt file',
action="store",
dest="alphabet_file_path")
parser.add_argument('--model-file',
type=str,
help='Path to the tf model file',
action="store",
dest="model_file_path")
parser.add_argument(
'--predicted-character-count',
type=int,
help='Number of most likely characters to be displayed',
action="store",
dest="predicted_character_count",
default=5)
args = parser.parse_args()
alphabet = Alphabet(os.path.abspath(args.alphabet_file_path))
if args.predicted_character_count >= alphabet.size():
args.predicted_character_count = alphabet.size() - 1
# Load frozen graph from file and parse it
with tf.io.gfile.GFile(args.model_file_path, "rb") as f:
graph_def = tf.compat.v1.GraphDef()
graph_def.ParseFromString(f.read())
# print(graph_def.node)
with tf.Graph().as_default() as graph:
tf.import_graph_def(graph_def, name="prefix")
# currently hardcoded values used during inference
with tf.compat.v1.Session(graph=graph) as session:
features, features_len = audiofile_to_features(
args.input_file_path)
previous_state_c = np.zeros([1, n_cell_dim])
previous_state_h = np.zeros([1, 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)
# we are interested only into logits, not CTC decoding
inputs = {
'input':
graph.get_tensor_by_name('prefix/input_node:0'),
'previous_state_c':
graph.get_tensor_by_name('prefix/previous_state_c:0'),
'previous_state_h':
graph.get_tensor_by_name('prefix/previous_state_h: 0'),
'input_lengths':
graph.get_tensor_by_name('prefix/input_lengths:0')
}
outputs = {
'outputs': graph.get_tensor_by_name('prefix/raw_logits:0'),
'new_state_c':
graph.get_tensor_by_name('prefix/new_state_c:0'),
'new_state_h':
graph.get_tensor_by_name('prefix/new_state_h: 0'),
}
logits = np.empty([0, 1, alphabet.size() + 1])
# the frozen model only accepts input split to 16 step chunks,
# if the inference was run from checkpoint instead (as in single inference in deepspeech script), this loop wouldn't be needed
for i in range(0, features_len[0], n_steps):
chunk = features[:, i:i + n_steps, :, :]
chunk_length = chunk.shape[1]
# pad with zeros if not enough steps (len(features) % FLAGS.n_steps != 0)
if chunk_length < n_steps:
chunk = np.pad(chunk, ((0, 0), (0, n_steps - chunk_length),
(0, 0), (0, 0)),
mode='constant',
constant_values=0)
# need to update the states with each loop iteration
logits_step, previous_state_c, previous_state_h = session.run(
[
outputs['outputs'], outputs['new_state_c'],
outputs['new_state_h']
],
feed_dict={
inputs['input']: chunk,
inputs['input_lengths']: [chunk_length],
inputs['previous_state_c']: previous_state_c,
inputs['previous_state_h']: previous_state_h,
})
logits = np.concatenate((logits, logits_step))
logits = np.squeeze(logits)
row_output = []
for j in range(args.predicted_character_count):
row_output.append([])
# now sort logits and turn them into characters + probabilities
for i in range(0, len(logits)):
softmax_output = softmax(logits[i])
indexes_sorted = softmax_output.argsort(
)[args.predicted_character_count * -1:][::-1]
most_likely_chars = ''
chars_probability = ''
for j in range(args.predicted_character_count):
char_index = indexes_sorted[j]
if char_index < alphabet.size():
text = alphabet._string_from_label(char_index)
most_likely_chars += text + ' '
row_output[j].append(text)
chars_probability += ' (' + str(
softmax_output[char_index]) + ')'
else:
most_likely_chars += '- '
row_output[j].append('-')
chars_probability += ' (' + str(
softmax_output[char_index]) + ')'
print(most_likely_chars, " ", chars_probability)
with open(args.input_file_path + "_acoustic.txt", "w") as out:
for j in range(len(row_output)):
out.write(', '.join(row_output[j]) + "\n")
print(row_output[j])
def run_inference():
"""Load frozen graph, run inference and display most likely predicted characters"""
parser = argparse.ArgumentParser(description='Run Deepspeech inference to obtain char probabilities')
parser.add_argument('--input-file', type=str,
help='Path to the wav file', action="store", dest="input_file_path")
parser.add_argument('--alphabet-file', type=str,
help='Path to the alphabet.txt file', action="store", dest="alphabet_file_path")
parser.add_argument('--model-file', type=str,
help='Path to the tf model file', action="store", dest="model_file_path")
parser.add_argument('--predicted-character-count', type=int,
help='Number of most likely characters to be displayed', action="store",
dest="predicted_character_count", default=5)
args = parser.parse_args()
alphabet = Alphabet(os.path.abspath(args.alphabet_file_path))
if args.predicted_character_count >= alphabet.size():
args.predicted_character_count = alphabet.size() - 1
# Load frozen graph from file and parse it
with tf.gfile.GFile(args.model_file_path, "rb") as f:
graph_def = tf.GraphDef()
graph_def.ParseFromString(f.read())
with tf.Graph().as_default() as graph:
tf.import_graph_def(graph_def, name="prefix")
# currently hardcoded values used during inference
n_input = 26
n_context = 9
n_steps = 16
with tf.Session(graph=graph) as session:
session.run('prefix/initialize_state')
features = util.audio.audiofile_to_input_vector(args.input_file_path, n_input, n_context)
num_strides = len(features) - (n_context * 2)
window_size = 2 * n_context + 1
features = np.lib.stride_tricks.as_strided(
features,
(num_strides, window_size, n_input),
(features.strides[0], features.strides[0], features.strides[1]),
writeable=False)
# we are interested only into logits, not CTC decoding
inputs = {'input': graph.get_tensor_by_name('prefix/input_node:0'),
'input_lengths': graph.get_tensor_by_name('prefix/input_lengths:0')}
outputs = {'outputs': graph.get_tensor_by_name('prefix/logits:0')}
logits = np.empty([0, 1, alphabet.size() + 1])
for i in range(0, len(features), n_steps):
chunk = features[i:i + n_steps]
# pad with zeros if not enough steps (len(features) % FLAGS.n_steps != 0)
if len(chunk) < n_steps:
chunk = np.pad(chunk,
(
(0, n_steps - len(chunk)),
(0, 0),
(0, 0)
),
mode='constant',
constant_values=0)
output = session.run(outputs['outputs'], feed_dict={
inputs['input']: [chunk],
inputs['input_lengths']: [len(chunk)],
})
logits = np.concatenate((logits, output))
for i in range(0, len(logits)):
softmax_output = softmax(logits[i][0])
indexes_sorted = softmax_output.argsort()[args.predicted_character_count * -1:][::-1]
most_likely_chars = ''
chars_probability = ''
for j in range(args.predicted_character_count):
char_index = indexes_sorted[j]
if char_index < alphabet.size():
text = alphabet.string_from_label(char_index)
most_likely_chars += text+' '
chars_probability += ' (' + str(softmax_output[char_index]) + ')'
else:
most_likely_chars += '- '
chars_probability += ' (' + str(softmax_output[char_index]) + ')'
print(most_likely_chars, " ", chars_probability)
