np.tile(embeddings[:, -1, :], [1, pad, 1])
],
axis=1)
embedding_last = trainer.predict(feature_array[-1])
pad = (feature_array[-1].shape[0] - embedding_last.shape[0]) / 2
embedding_last = np.concatenate([
np.tile(embedding_last[0, :], [pad, 1]), embedding_last,
np.tile(embedding_last[-1, :], [pad, 1])
],
axis=0)
embeddings = np.reshape(embeddings, [
embeddings.shape[0] * embeddings.shape[1], embeddings.shape[2]
])
embedding = np.concatenate([embeddings, embedding_last], axis=0)
else:
tf.logging.info("[INFO] Key %s length %d." %
(key, feature.shape[0]))
embedding = trainer.predict(feature)
# padding
pad = (feature.shape[0] - embedding.shape[0]) / 2
embedding = np.concatenate([
np.tile(embedding[0, :], [pad, 1]), embedding,
np.tile(embedding[-1, :], [pad, 1])
],
axis=0)
assert (embedding.shape[0] == feature.shape[0])
write_mat(fp_out, embedding, key=key)
fp_out.close()
trainer.close()
import sys
from dataset.kaldi_io import open_or_fd, read_mat_scp, write_mat
import numpy as np
if len(sys.argv) != 4:
print('Usage: %s phone_class post_in_scp post_out_ark' % sys.argv[0])
quit()
phone_class = []
with open(sys.argv[1], "r") as f:
for line in f.readlines():
phones = line.strip().split(" ")
phone_class.append([int(p) for p in phones])
num_classes = len(phone_class)
fp_out = open_or_fd(sys.argv[3], "wb")
for key, mat in read_mat_scp(sys.argv[2]):
post_new = np.zeros((mat.shape[0], num_classes))
for index, phones in enumerate(phone_class):
post_new[:, index] = np.sum(mat[:, phones], axis=1)
write_mat(fp_out, post_new, key=key)
fp_out.close()
#
# log_prob = np.reshape(log_prob, [log_prob.shape[0] * log_prob.shape[1], log_prob.shape[2]])
# log_prob = np.concatenate([log_prob, log_prob_last], axis=0)
# assert(log_prob.shape[0] == feature.shape[0] and log_prob.shape[1] == prior_vec.shape[0])
raise NotImplementedError("Do not let the utterance to be split.")
else:
tf.logging.info("[INFO] Key %s length %d." %
(key, feature.shape[0]))
log_prob = trainer.predict_phone(node, feature, [feature.shape[0]])
assert (log_prob.shape[0] == feature.shape[0]
and log_prob.shape[1] == prior_vec.shape[0])
# Convert to log-posteriors to log-likelihood
log_like = log_prob - log_prior_vec
write_mat(fp_out, log_like, key=key)
if is_output_logpost:
write_mat(fp_logpost,
np.array(log_prob, dtype=np.float32),
key=key)
num_done += 1
if is_output_logpost:
fp_logpost.close()
fp_out.close()
trainer.close()
tf.logging.info("Compute %d log-likelihood." % (num_done))
# start = i * args.chunk_size
# this_chunk_size = args.chunk_size if feature.shape[0] - start > args.chunk_size else feature.shape[
# 0] - start
# feature_length.append(this_chunk_size)
# feature_array.append(feature[start:start + this_chunk_size])
#
# # Except for the last feature, the length of other features should be the same (=chunk_size)
# log_prob = trainer.predict_phone(node,
# np.array(feature_array[:-1], dtype=np.float32),
# feature_length[:-1])
# log_prob_last = trainer.predict_phone(node, feature_array[-1], [feature_length[-1]])
#
# log_prob = np.reshape(log_prob, [log_prob.shape[0] * log_prob.shape[1], log_prob.shape[2]])
# log_prob = np.concatenate([log_prob, log_prob_last], axis=0)
# assert(log_prob.shape[0] == feature.shape[0] and log_prob.shape[1] == prior_vec.shape[0])
raise NotImplementedError("Do not let the utterance to be split.")
else:
tf.logging.info("[INFO] Key %s length %d." % (key, feature.shape[0]))
log_prob = trainer.predict_phone(node, feature, [feature.shape[0]])
assert(log_prob.shape[0] == feature.shape[0] and log_prob.shape[1] == prior_vec.shape[0])
# Convert to log-posteriors to log-likelihood
log_like = log_prob - log_prior_vec
write_mat(fp_out, log_like, key=key)
num_done += 1
fp_out.close()
trainer.close()
tf.logging.info("Compute %d log-likelihood." % (num_done))
config_json = os.path.join(args.model_dir, "nnet/config.json")
if not os.path.isfile(config_json):
sys.exit("Cannot find params.json in %s" % config_json)
params = Params(config_json)
# Attention weights
params.embedding_node = "attention_weights"
with open(os.path.join(nnet_dir, "feature_dim"), "r") as f:
dim = int(f.readline().strip())
trainer = Trainer(params, args.model_dir, dim, single_cpu=True)
trainer.build("predict")
if args.rspecifier.rsplit(".", 1)[1] == "scp":
# The rspecifier cannot be scp
sys.exit("The rspecifier must be ark or input pipe")
fp_out = open_or_fd(args.wspecifier, "wb")
for index, (key, feature) in enumerate(read_mat_ark(args.rspecifier)):
if feature.shape[0] < args.min_chunk_size:
tf.logging.info("[INFO] Key %s length too short, %d < %d, skip." %
(key, feature.shape[0], args.min_chunk_size))
continue
if feature.shape[0] > args.chunk_size:
# We only extract the first segment
feature = feature[:args.chunk_size]
attention_weights = trainer.predict(feature)
write_mat(fp_out, attention_weights, key=key)
fp_out.close()
trainer.close()