Spaces:
Running
Running
File size: 4,886 Bytes
3dd84f8 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 |
import torch
# reference: https://github.com/jaywalnut310/vits/blob/main/data_utils.py
class DistributedBucketSampler(torch.utils.data.distributed.DistributedSampler):
"""
Maintain similar input lengths in a batch.
Length groups are specified by boundaries.
Ex) boundaries = [b1, b2, b3] -> any batch is included either {x | b1 < length(x) <=b2} or {x | b2 < length(x) <= b3}.
It removes samples which are not included in the boundaries.
Ex) boundaries = [b1, b2, b3] -> any x s.t. length(x) <= b1 or length(x) > b3 are discarded.
"""
def __init__(
self,
dataset,
batch_size,
boundaries,
num_replicas=None,
rank=None,
shuffle=True,
):
super().__init__(dataset, num_replicas=num_replicas, rank=rank, shuffle=shuffle)
self.lengths = dataset.lengths
self.batch_size = batch_size
self.boundaries = boundaries
self.buckets, self.num_samples_per_bucket = self._create_buckets()
self.total_size = sum(self.num_samples_per_bucket)
self.num_samples = self.total_size // self.num_replicas
def _create_buckets(self):
buckets = [[] for _ in range(len(self.boundaries) - 1)]
for i in range(len(self.lengths)):
length = self.lengths[i]
idx_bucket = self._bisect(length)
if idx_bucket != -1:
buckets[idx_bucket].append(i)
# from https://github.com/Plachtaa/VITS-fast-fine-tuning/blob/main/data_utils.py
# avoid "integer division or modulo by zero" error for very small dataset
try:
for i in range(len(buckets) - 1, 0, -1):
if len(buckets[i]) == 0:
buckets.pop(i)
self.boundaries.pop(i + 1)
assert all(len(bucket) > 0 for bucket in buckets)
# When one bucket is not traversed
except Exception as e:
print('Bucket warning ', e)
for i in range(len(buckets) - 1, -1, -1):
if len(buckets[i]) == 0:
buckets.pop(i)
self.boundaries.pop(i + 1)
num_samples_per_bucket = []
for i in range(len(buckets)):
len_bucket = len(buckets[i])
total_batch_size = self.num_replicas * self.batch_size
rem = (
total_batch_size - (len_bucket % total_batch_size)
) % total_batch_size
num_samples_per_bucket.append(len_bucket + rem)
return buckets, num_samples_per_bucket
def __iter__(self):
# deterministically shuffle based on epoch
g = torch.Generator()
g.manual_seed(self.epoch)
indices = []
if self.shuffle:
for bucket in self.buckets:
indices.append(torch.randperm(len(bucket), generator=g).tolist())
else:
for bucket in self.buckets:
indices.append(list(range(len(bucket))))
batches = []
for i in range(len(self.buckets)):
bucket = self.buckets[i]
len_bucket = len(bucket)
ids_bucket = indices[i]
num_samples_bucket = self.num_samples_per_bucket[i]
# add extra samples to make it evenly divisible
rem = num_samples_bucket - len_bucket
ids_bucket = (
ids_bucket
+ ids_bucket * (rem // len_bucket)
+ ids_bucket[: (rem % len_bucket)]
)
# subsample
ids_bucket = ids_bucket[self.rank :: self.num_replicas]
# batching
for j in range(len(ids_bucket) // self.batch_size):
batch = [
bucket[idx]
for idx in ids_bucket[
j * self.batch_size : (j + 1) * self.batch_size
]
]
batches.append(batch)
if self.shuffle:
batch_ids = torch.randperm(len(batches), generator=g).tolist()
batches = [batches[i] for i in batch_ids]
self.batches = batches
assert len(self.batches) * self.batch_size == self.num_samples
return iter(self.batches)
def _bisect(self, x, lo=0, hi=None):
if hi is None:
hi = len(self.boundaries) - 1
if hi > lo:
mid = (hi + lo) // 2
if self.boundaries[mid] < x and x <= self.boundaries[mid + 1]:
return mid
elif x <= self.boundaries[mid]:
return self._bisect(x, lo, mid)
else:
return self._bisect(x, mid + 1, hi)
else:
return -1
def __len__(self):
return self.num_samples // self.batch_size |