def run(args):
num_done = 0
with common_lib.smart_open(
args.pasted_targets) as targets_reader, common_lib.smart_open(
args.out_targets, "w") as targets_writer:
for key, mat in common_lib.read_mat_ark(targets_reader):
mat = np.matrix(mat)
if mat.shape[1] % args.dim != 0:
raise RuntimeError(
"For utterance {utt} in {f}, num-columns {nc} "
"is not a multiple of dim {dim}"
"".format(
utt=key,
f=args.pasted_targets.name,
nc=mat.shape[1],
dim=args.dim,
))
num_sources = mat.shape[1] // args.dim
out_mat = np.matrix(np.zeros([mat.shape[0], args.dim]))
if args.remove_mismatch_frames:
for n in range(mat.shape[0]):
if should_remove_frame(mat[n, :].getA()[0], args.dim):
out_mat[n, :] = np.zeros([1, args.dim])
else:
for i in range(num_sources):
out_mat[n, :] += mat[n, (i * args.dim):(
(i + 1) *
args.dim)] * (1.0 if args.weights is None else
args.weights[i])
else:
# Just interpolate the targets
for i in range(num_sources):
out_mat += mat[:, (i * args.dim):((i + 1) * args.dim)] * (
1.0 if args.weights is None else args.weights[i])
common_lib.write_matrix_ascii(targets_writer,
out_mat.tolist(),
key=key)
num_done += 1
logger.info("Merged {num_done} target matrices"
"".format(num_done=num_done))
if num_done == 0:
raise RuntimeError
def run(args):
num_done = 0
with common_lib.smart_open(args.pasted_targets) as targets_reader, \
common_lib.smart_open(args.out_targets, 'w') as targets_writer:
for key, mat in common_lib.read_mat_ark(targets_reader):
mat = np.matrix(mat)
if mat.shape[1] % args.dim != 0:
raise RuntimeError(
"For utterance {utt} in {f}, num-columns {nc} "
"is not a multiple of dim {dim}"
"".format(utt=key, f=args.pasted_targets.name,
nc=mat.shape[1], dim=args.dim))
num_sources = mat.shape[1] // args.dim
out_mat = np.matrix(np.zeros([mat.shape[0], args.dim]))
if args.remove_mismatch_frames:
for n in range(mat.shape[0]):
if should_remove_frame(mat[n, :].getA()[0], args.dim):
out_mat[n, :] = np.zeros([1, args.dim])
else:
for i in range(num_sources):
out_mat[n, :] += (
mat[n, (i * args.dim) : ((i+1) * args.dim)]
* (1.0 if args.weights is None
else args.weights[i]))
else:
# Just interpolate the targets
for i in range(num_sources):
out_mat += (
mat[:, (i * args.dim) : ((i+1) * args.dim)]
* (1.0 if args.weights is None else args.weights[i]))
common_lib.write_matrix_ascii(targets_writer, out_mat.tolist(),
key=key)
num_done += 1
logger.info("Merged {num_done} target matrices"
"".format(num_done=num_done))
if num_done == 0:
raise RuntimeError
def run(args):
num_utts = 0
for key, mat in common_lib.read_mat_ark(args.targets_in_ark):
mat = np.matrix(mat)
if args.subsampling_factor > 0:
num_indexes = (old_div(
(mat.shape[0] + args.subsampling_factor - 1),
args.subsampling_factor))
out_mat = np.zeros([num_indexes, mat.shape[1]])
i = 0
for k in range(int(old_div(args.subsampling_factor, 2.0)),
mat.shape[0], args.subsampling_factor):
st = int(k - old_div(float(args.subsampling_factor), 2.0))
end = int(k + old_div(float(args.subsampling_factor), 2.0))
if st < 0:
st = 0
if end > mat.shape[0]:
end = mat.shape[0]
try:
out_mat[i, :] = old_div(np.sum(mat[st:end, :], axis=0),
float(end - st))
except IndexError:
logger.error("mat.shape = {0}, st = {1}, end = {2}"
"".format(mat.shape, st, end))
raise
assert i == old_div(k, args.subsampling_factor)
i += 1
common_lib.write_matrix_ascii(args.targets_out_ark, out_mat, key=key)
num_utts += 1
args.targets_in_ark.close()
args.targets_out_ark.close()
logger.info("Sub-sampled {num_utts} target matrices"
"".format(num_utts=num_utts))
def run(args):
num_utts = 0
for key, mat in common_lib.read_mat_ark(args.targets_in_ark):
mat = np.matrix(mat)
if args.subsampling_factor > 0:
num_indexes = ((mat.shape[0] + args.subsampling_factor - 1)
/ args.subsampling_factor)
out_mat = np.zeros([num_indexes, mat.shape[1]])
i = 0
for k in range(int(args.subsampling_factor / 2.0),
mat.shape[0], args.subsampling_factor):
st = int(k - float(args.subsampling_factor) / 2.0)
end = int(k + float(args.subsampling_factor) / 2.0)
if st < 0:
st = 0
if end > mat.shape[0]:
end = mat.shape[0]
try:
out_mat[i, :] = np.sum(mat[st:end, :], axis=0) / float(end - st)
except IndexError:
logger.error("mat.shape = {0}, st = {1}, end = {2}"
"".format(mat.shape, st, end))
raise
assert i == k / args.subsampling_factor
i += 1
common_lib.write_matrix_ascii(args.targets_out_ark, out_mat, key=key)
num_utts += 1
args.targets_in_ark.close()
args.targets_out_ark.close()
logger.info("Sub-sampled {num_utts} target matrices"
"".format(num_utts=num_utts))