def __init__(
self,
batch_size: int,
shape_file: str,
sort_in_batch: str = "descending",
sort_batch: str = "ascending",
drop_last: bool = False,
):
assert check_argument_types()
assert batch_size > 0
self.batch_size = batch_size
self.shape_file = shape_file
self.sort_in_batch = sort_in_batch
self.sort_batch = sort_batch
self.drop_last = drop_last
# utt2shape: (Length, ...)
# uttA 100,...
# uttB 201,...
utt2shape = load_num_sequence_text(shape_file, loader_type="csv_int")
if sort_in_batch == "descending":
# Sort samples in descending order (required by RNN)
keys = sorted(utt2shape, key=lambda k: -utt2shape[k][0])
elif sort_in_batch == "ascending":
# Sort samples in ascending order
keys = sorted(utt2shape, key=lambda k: utt2shape[k][0])
else:
raise ValueError(
f"sort_in_batch must be either one of "
f"ascending, descending, or None: {sort_in_batch}"
)
if len(keys) == 0:
raise RuntimeError(f"0 lines found: {shape_file}")
# Apply max(, 1) to avoid 0-batches
N = max(len(keys) // batch_size, 1)
if not self.drop_last:
# Split keys evenly as possible as. Note that If N != 1,
# the these batches always have size of batch_size at minimum.
self.batch_list = [
keys[i * len(keys) // N : (i + 1) * len(keys) // N] for i in range(N)
]
else:
self.batch_list = [
tuple(keys[i * batch_size : (i + 1) * batch_size]) for i in range(N)
]
if len(self.batch_list) == 0:
logging.warning(f"{shape_file} is empty")
if sort_in_batch != sort_batch:
if sort_batch not in ("ascending", "descending"):
raise ValueError(
f"sort_batch must be ascending or descending: {sort_batch}"
)
self.batch_list.reverse()
if len(self.batch_list) == 0:
raise RuntimeError("0 batches")
def test_load_num_sequence_text(loader_type: str, tmp_path: Path):
p = tmp_path / "dummy.txt"
if "csv" in loader_type:
delimiter = ","
else:
delimiter = " "
with p.open("w") as f:
f.write("abc " + delimiter.join(["0", "1", "2"]) + "\n")
f.write("def " + delimiter.join(["3", "4", "5"]) + "\n")
desired = {"abc": np.array([0, 1, 2]), "def": np.array([3, 4, 5])}
if loader_type == "dummy":
with pytest.raises(ValueError):
load_num_sequence_text(p, loader_type=loader_type)
return
else:
target = load_num_sequence_text(p, loader_type=loader_type)
for k in desired:
np.testing.assert_array_equal(target[k], desired[k])
def __init__(
self,
shape_file: Union[Path, str],
dtype: Union[str, np.dtype] = "float32",
loader_type: str = "csv_int",
):
assert check_argument_types()
shape_file = Path(shape_file)
self.utt2shape = load_num_sequence_text(shape_file, loader_type)
self.dtype = np.dtype(dtype)
def __init__(
self,
shape_file: Union[Path, str],
low: int,
high: int = None,
dtype: Union[str, np.dtype] = "int64",
loader_type: str = "csv_int",
):
assert check_argument_types()
shape_file = Path(shape_file)
self.utt2shape = load_num_sequence_text(shape_file, loader_type)
self.dtype = np.dtype(dtype)
self.low = low
self.high = high
def test_load_num_sequence_text_invalid(tmp_path: Path):
p = tmp_path / "dummy.txt"
with p.open("w") as f:
f.write("abc 12.3.3.,4.44\n")
with pytest.raises(ValueError):
load_num_sequence_text(p)
with p.open("w") as f:
f.write("abc\n")
with pytest.raises(RuntimeError):
load_num_sequence_text(p)
with p.open("w") as f:
f.write("abc 1 2\n")
f.write("abc 2 4\n")
with pytest.raises(RuntimeError):
load_num_sequence_text(p)
def __init__(
self,
batch_bins: int,
shape_files: Union[Tuple[str, ...], List[str]],
min_batch_size: int = 1,
sort_in_batch: str = "descending",
sort_batch: str = "ascending",
drop_last: bool = False,
padding: bool = True,
):
assert check_argument_types()
assert batch_bins > 0
if sort_batch != "ascending" and sort_batch != "descending":
raise ValueError(
f"sort_batch must be ascending or descending: {sort_batch}")
if sort_in_batch != "descending" and sort_in_batch != "ascending":
raise ValueError(
f"sort_in_batch must be ascending or descending: {sort_in_batch}"
)
self.batch_bins = batch_bins
self.shape_files = shape_files
self.sort_in_batch = sort_in_batch
self.sort_batch = sort_batch
self.drop_last = drop_last
# utt2shape: (Length, ...)
# uttA 100,...
# uttB 201,...
utt2shapes = [
load_num_sequence_text(s, loader_type="csv_int")
for s in shape_files
]
first_utt2shape = utt2shapes[0]
for s, d in zip(shape_files, utt2shapes):
if set(d) != set(first_utt2shape):
raise RuntimeError(
f"keys are mismatched between {s} != {shape_files[0]}")
# Sort samples in ascending order
# (shape order should be like (Length, Dim))
keys = sorted(first_utt2shape, key=lambda k: first_utt2shape[k][0])
if len(keys) == 0:
raise RuntimeError(f"0 lines found: {shape_files[0]}")
# Decide batch-sizes
start = 0
batch_sizes = []
bs = 1
while True:
# shape: (Length, dim1, dim2, ...)
if padding:
max_lengths = [
max(d[keys[i]][0] for i in range(start, start + bs))
for d in utt2shapes
]
# bins = bs x max_length
bins = sum(bs * lg for lg in max_lengths)
else:
# bins = sum of lengths
bins = sum(d[keys[i]][0] for i in range(start, start + bs)
for d in utt2shapes)
if bins > batch_bins and bs >= min_batch_size:
batch_sizes.append(bs)
start += bs
bs = 1
else:
bs += 1
if start >= len(keys):
break
if start + bs > len(keys):
if not self.drop_last or len(batch_sizes) == 0:
batch_sizes.append(len(keys) - start)
break
if len(batch_sizes) == 0:
# Maybe we can't reach here
raise RuntimeError("0 batches")
# If the last batch-size is smaller than minimum batch_size,
# the samples are redistributed to the other mini-batches
if len(batch_sizes) > 1 and batch_sizes[-1] < min_batch_size:
for i in range(batch_sizes.pop(-1)):
batch_sizes[-(i % len(batch_sizes)) - 2] += 1
if not self.drop_last:
# Bug check
assert sum(batch_sizes) == len(
keys), f"{sum(batch_sizes)} != {len(keys)}"
# Set mini-batch
self.batch_list = []
start = 0
for bs in batch_sizes:
assert len(keys) >= start + bs, "Bug"
minibatch_keys = keys[start:start + bs]
start += bs
if sort_in_batch == "descending":
minibatch_keys.reverse()
elif sort_in_batch == "ascending":
# Key are already sorted in ascending
pass
else:
raise ValueError(
f"sort_in_batch must be ascending or descending: {sort_in_batch}"
)
self.batch_list.append(tuple(minibatch_keys))
if sort_batch == "ascending":
pass
elif sort_batch == "descending":
self.batch_list.reverse()
else:
raise ValueError(
f"sort_batch must be ascending or descending: {sort_batch}")
def __init__(
self,
batch_size: int,
shape_files: Union[Tuple[str, ...], List[str]],
fold_lengths: Sequence[int],
min_batch_size: int = 1,
sort_in_batch: str = "descending",
sort_batch: str = "ascending",
drop_last: bool = False,
):
assert check_argument_types()
assert batch_size > 0
if sort_batch != "ascending" and sort_batch != "descending":
raise ValueError(
f"sort_batch must be ascending or descending: {sort_batch}")
if sort_in_batch != "descending" and sort_in_batch != "ascending":
raise ValueError(
f"sort_in_batch must be ascending or descending: {sort_in_batch}"
)
self.batch_size = batch_size
self.shape_files = shape_files
self.sort_in_batch = sort_in_batch
self.sort_batch = sort_batch
self.drop_last = drop_last
# utt2shape: (Length, ...)
# uttA 100,...
# uttB 201,...
utt2shapes = [
load_num_sequence_text(s, loader_type="csv_int")
for s in shape_files
]
first_utt2shape = utt2shapes[0]
for s, d in zip(shape_files, utt2shapes):
if set(d) != set(first_utt2shape):
raise RuntimeError(
f"keys are mismatched between {s} != {shape_files[0]}")
# Sort samples in ascending order
# (shape order should be like (Length, Dim))
keys = sorted(first_utt2shape, key=lambda k: first_utt2shape[k][0])
if len(keys) == 0:
raise RuntimeError(f"0 lines found: {shape_files[0]}")
# Decide batch-sizes
start = 0
batch_sizes = []
while True:
k = keys[start]
factor = max(
int(d[k][0] / m) for d, m in zip(utt2shapes, fold_lengths))
bs = max(min_batch_size, int(batch_size / (1 + factor)))
if self.drop_last and start + bs > len(keys):
# This if-block avoids 0-batches
if len(self.batch_list) > 0:
break
bs = min(len(keys) - start, bs)
batch_sizes.append(bs)
start += bs
if start >= len(keys):
break
if len(batch_sizes) == 0:
# Maybe we can't reach here
raise RuntimeError("0 batches")
# If the last batch-size is smaller than minimum batch_size,
# the samples are redistributed to the other mini-batches
if len(batch_sizes) > 1 and batch_sizes[-1] < min_batch_size:
for i in range(batch_sizes.pop(-1)):
batch_sizes[-(i % len(batch_sizes)) - 2] += 1
if not self.drop_last:
# Bug check
assert sum(batch_sizes) == len(
keys), f"{sum(batch_sizes)} != {len(keys)}"
# Set mini-batch
self.batch_list = []
start = 0
for bs in batch_sizes:
assert len(keys) >= start + bs, "Bug"
minibatch_keys = keys[start:start + bs]
start += bs
if sort_in_batch == "descending":
minibatch_keys.reverse()
elif sort_in_batch == "ascending":
# Key are already sorted in ascending
pass
else:
raise ValueError(
f"sort_in_batch must be ascending or descending: {sort_in_batch}"
)
self.batch_list.append(tuple(minibatch_keys))
if sort_batch == "ascending":
pass
elif sort_batch == "descending":
self.batch_list.reverse()
else:
raise ValueError(
f"sort_batch must be ascending or descending: {sort_batch}")
