"""PyTorch compatible samplers.
These determine the order of iteration through a dataset.
Authors:
* Aku Rouhe 2020
* Samuele Cornell 2020
* Ralf Leibold 2020
"""
import torch
import logging
from operator import itemgetter
from torch.utils.data import (
RandomSampler,
WeightedRandomSampler,
DistributedSampler,
Sampler,
)
import numpy as np
from typing import List
from speechbrain.dataio.dataset import DynamicItemDataset
logger = logging.getLogger(__name__)
[docs]class ReproducibleRandomSampler(RandomSampler):
"""A modification of RandomSampler which always returns the same values.
Also look at `torch.utils.data.RandomSampler`. This has mostly
the same behaviour and arguments, except for adding 'seed' and 'epoch' and
not supporting 'generator'.
Note
----
Call `set_epoch` before every epoch. Otherwise, the sampler will produce the
same sequence of indices every epoch.
Arguments
---------
data_source : Dataset
The data source to sample indices for.
seed : int
The base seed to use for the random number generator. It is recommended
to use a value which has a good mix of 0 and 1 bits.
epoch : int
The epoch to start at.
Example
-------
>>> import torch
>>> from speechbrain.utils.checkpoints import Checkpointer
>>> from speechbrain.dataio.dataloader import SaveableDataLoader
>>> # An example "dataset"
>>> dataset = torch.arange(10).unsqueeze(1)
>>> # Create the random sampler:
>>> sampler = ReproducibleRandomSampler(dataset)
>>> dataloader = SaveableDataLoader(dataset, sampler = sampler,
... num_workers = 3)
>>> # Setup the checkpointer.
>>> # Note that the sampler doesn't need to be saved itself.
>>> tmpdir = getfixture('tmpdir')
>>> checkpointer = Checkpointer(tmpdir, {"dataloader": dataloader})
>>> # Iterate:
>>> subset = []
>>> for i, data_point in enumerate(dataloader):
... # Say you save a checkpoint on the fourth batch:
... if i == 3:
... _ = checkpointer.save_checkpoint(end_of_epoch = False)
... # So let's save the numbers you would get if you continue
... if i >= 4:
... subset.append(data_point.item())
>>> # What if instead you had to restart the experiment?
>>> new_sampler = ReproducibleRandomSampler(dataset)
>>> new_dataloader = SaveableDataLoader(dataset, sampler = new_sampler,
... num_workers = 3)
>>> new_checkpointer = Checkpointer(tmpdir, {"dataloader": new_dataloader})
>>> _ = new_checkpointer.recover_if_possible()
>>> # You'll get the same random order again:
>>> new_subset = [data_point.item() for data_point in new_dataloader]
>>> assert subset == new_subset
"""
def __init__(self, data_source, seed=563375142, epoch=0, **kwargs):
if "generator" in kwargs:
MSG = (
"Cannot give a separate generator when using "
+ "ReproducibleRandomSampler"
)
raise ValueError(MSG)
super().__init__(data_source, **kwargs)
self.seed = int(seed)
self.epoch = epoch
self.generator = torch.Generator()
[docs] def set_epoch(self, epoch):
"""
You can also just access self.epoch, but we maintain this interface
to mirror torch.utils.data.distributed.DistributedSampler
"""
self.epoch = epoch
def __iter__(self):
self.generator.manual_seed(self.seed + self.epoch)
return super().__iter__()
[docs]class ReproducibleWeightedRandomSampler(WeightedRandomSampler):
"""A reproducible modification of WeightedRandomSampler.
Also look at `torch.utils.data.WeightedRandomSampler`. This has the
the same behaviour and arguments, except for adding 'seed' and 'epoch' and
not supporting 'generator'.
Note
----
Call `set_epoch` before every epoch. Otherwise, the sampler will produce the
same sequence of indices every epoch.
Arguments
---------
weights : sequence of float
Weights for each index. Doesn't need to sum to one.
num_samples : int
Number of samples to draw
replacement : bool
To draw with replacement or not (within an epoch of num_samples).
seed : int
The base seed to use for the random number generator. It is recommended
to use a value which has a good mix of 0 and 1 bits.
epoch : int
The epoch to start at.
Example
-------
>>> a = ReproducibleWeightedRandomSampler([0.1, 0.9, 0.4, 0.7, 3.0, 0.6], 5, replacement=True)
>>> b = ReproducibleWeightedRandomSampler([0.1, 0.9, 0.4, 0.7, 3.0, 0.6], 5, replacement=True)
>>> list(a)
[3, 1, 4, 4, 4]
>>> list(b)
[3, 1, 4, 4, 4]
>>> a.set_epoch(1)
>>> list(a)
[4, 5, 4, 4, 3]
>>> b.set_epoch(1)
>>> list(b)
[4, 5, 4, 4, 3]
"""
def __init__(
self,
weights,
num_samples,
replacement,
seed=129491412,
epoch=0,
**kwargs,
):
if "generator" in kwargs:
MSG = (
"Cannot give a separate generator when using "
+ "ReproducibleRandomSampler"
)
raise ValueError(MSG)
super().__init__(weights, num_samples, replacement, **kwargs)
self.seed = int(seed)
self.epoch = epoch
self.generator = torch.Generator()
[docs] def set_epoch(self, epoch):
"""
You can also just access self.epoch, but we maintain this interface
to mirror torch.utils.data.distributed.DistributedSampler
"""
self.epoch = epoch
def __iter__(self):
self.generator.manual_seed(self.seed + self.epoch)
return super().__iter__()
[docs]class ConcatDatasetBatchSampler(Sampler):
"""This sampler is built to work with a standard Pytorch ConcatDataset.
It is used to retrieve elements from the different concatenated datasets placing them in the same batch
with proportion specified by batch_sizes, e.g 8, 16 means each batch will
be of 24 elements with the first 8 belonging to the first dataset in ConcatDataset
object and the last 16 to the second.
More than two datasets are supported, in that case you need to provide 3 batch
sizes.
Note
----
Batched are drawn from the datasets till the one with smallest length is exhausted.
Thus number of examples in your training epoch is dictated by the dataset
whose length is the smallest.
Arguments
---------
samplers : int
The base seed to use for the random number generator. It is recommended
to use a value which has a good mix of 0 and 1 bits.
batch_sizes: list
Batch sizes.
epoch : int
The epoch to start at.
Example
-------
>>> import torch
>>> from speechbrain.dataio.sampler import ConcatDatasetBatchSampler, ReproducibleRandomSampler
>>> from speechbrain.dataio.sampler import ReproducibleRandomSampler
>>> from speechbrain.dataio.dataloader import SaveableDataLoader
>>> # example "datasets"
>>> dataset1 = torch.arange(0, 10).unsqueeze(1)
>>> dataset2 = torch.arange(20, 40).unsqueeze(1)
>>> tot_dataset = torch.utils.data.ConcatDataset([dataset1, dataset2])
>>> sampler1 = ReproducibleRandomSampler(dataset1)
>>> sampler2 = ReproducibleRandomSampler(dataset2)
>>> tot_sampler = ConcatDatasetBatchSampler([sampler1, sampler2], [2, 4])
>>> dataloader = SaveableDataLoader(tot_dataset, batch_sampler = tot_sampler,
... num_workers = 3)
>>> for data_point in dataloader:
... assert len(data_point) == 6
... for i in range(2):
... assert data_point[i] in [x for x in range(0, 10)]
... for i in range(2, 4):
... assert data_point[i] in [x for x in range(10, 40)]
"""
def __init__(self, samplers, batch_sizes: (tuple, list), epoch=0) -> None:
if not isinstance(samplers, (list, tuple)):
raise ValueError(
"samplers should be a list or tuple of Pytorch Samplers, "
"but got samplers={}".format(batch_sizes)
)
if not isinstance(batch_sizes, (list, tuple)):
raise ValueError(
"batch_sizes should be a list or tuple of integers, "
"but got batch_sizes={}".format(batch_sizes)
)
if not len(batch_sizes) == len(samplers):
raise ValueError(
"batch_sizes and samplers should be have same length"
)
self.batch_sizes = batch_sizes
self.samplers = samplers
self.offsets = [0] + np.cumsum(
[len(x) for x in self.samplers]
).tolist()[:-1]
self.epoch = epoch
self.set_epoch(self.epoch)
def _iter_one_dataset(self, c_batch_size, c_sampler, c_offset):
batch = []
for idx in c_sampler:
batch.append(c_offset + idx)
if len(batch) == c_batch_size:
yield batch
[docs] def set_epoch(self, epoch):
"""You can also just access self.epoch, but we maintain this interface
to mirror ``torch.utils.data.distributed.DistributedSampler``.
"""
if hasattr(self.samplers[0], "epoch"):
for s in self.samplers:
s.set_epoch(epoch)
def __iter__(self):
iterators = [iter(i) for i in self.samplers]
tot_batch = []
for b_num in range(len(self)):
for samp_idx in range(len(self.samplers)):
c_batch = []
while len(c_batch) < self.batch_sizes[samp_idx]:
c_batch.append(
self.offsets[samp_idx] + next(iterators[samp_idx])
)
tot_batch.extend(c_batch)
yield tot_batch
tot_batch = []
def __len__(self):
min_len = float("inf")
for idx, sampler in enumerate(self.samplers):
c_len = len(sampler) // self.batch_sizes[idx]
min_len = min(c_len, min_len)
return min_len
[docs]class DynamicBatchSampler(Sampler):
"""This BatchSampler batches examples together by grouping them by their length.
Every example in the batch have approximatively the same length and
thus padding is minimized.
This enables faster training on datasets
where length of examples can vary significantly (e.g Librispeech).
Inspired by: https://www.tensorflow.org/api_docs/python/tf/data/experimental/bucket_by_sequence_length
Dynamic batching is performed by specifying a max_batch_length which is the
upper limit for the sum of the length of examples in a batch:
e.g., if ex1 has length 4, ex2 length 5 andn if max_batch_length is set to 6
ex1 and ex2 will be placed, alone, in two distinct batches.
Length for each example can be obtained in two manners.
If the input dataset is a DynamicItemDataset it can be obtained by specifying a
length_func. Default assumes a "duration" entry is in the annotation.
Length for each example can also be passed to this class upon instantiation
by specifying a list containing the length for each example and passing it to
lengths_list.
Examples are grouped together by defining a set of possible discrete intervals
(buckets) multiple of a left_bucket_length.
A bucket_length_multiplier is used to specify the number of possible buckets.
E.g., if max_batch_length = 32 and left_bucket_length = 10, bucket_length_multiplier = 2
there will be 3 buckets: [0, 10), [10, 20), [20, 40).
A common choice would be setting left_bucket_length to approximatively the length
of your shortest example in the dataset.
Decreasing bucket_length_multiplier creates more buckets in the whole interval
of [left_bucket_length, max_batch_size]: e.g. if max_batch_length = 32 and left_bucket_length = 10,
bucket_length_multiplier = 1.5 the number of buckets increases to 8.
With right boundaries: [10 12 14 17 21 25 30 36].
Thus examples with length less than 10 are all grouped together but more buckets
are created for longer examples.
Note that the bucket boundary grows exponentially using the multiplier.
The buckets can also be specified by passing a list to the bucket_boundaries
argument instead of specifying a left_bucket_length and a bucket_length_multiplier.
Example
-------
>>> import torch
>>> import speechbrain as sb
>>> from speechbrain.dataio.sampler import DynamicBatchSampler
>>> from speechbrain.dataio.dataset import DynamicItemDataset
>>> from speechbrain.dataio.dataloader import SaveableDataLoader
>>> from speechbrain.dataio.batch import PaddedBatch
>>> import numpy as np
>>> item_lengths = sorted([np.random.randint(10, 100) for x in range(20)])
>>> dataset = {"ex_{}".format(x) : {"wav" :torch.randn(x)} for x in item_lengths}
>>> dataset = DynamicItemDataset(dataset)
>>> dataset.set_output_keys(["wav"])
>>> length_func = lambda x : len(x) # trivial in this example
>>> bsampler = DynamicBatchSampler(dataset, 20, 10, 1.1, length_func, shuffle=False)
>>> dataloader = SaveableDataLoader(dataset, batch_sampler=bsampler, collate_fn=PaddedBatch)
>>> for i, b in enumerate(dataloader):
... data, length = b["wav"]
>>> assert data.shape[-1] == max(item_lengths)
Arguments
---------
dataset : torch.utils.data.Dataset
Pytorch Dataset from which elements will be sampled.
max_batch_length : int
Upper limit for the sum of the length of examples in a batch.
Should be chosen based on your GPU memory.
left_bucket_length : int
Minimum length of a bucket. Specifies resolution of buckets and thus this sampler
stochasticity. A common choice is to set this to length of your
shortest example.
bucket_length_multiplier : float
Multiplier for bucket length, specifies number of buckets from left_bucket_length to
max_batch_length.
length_func : callable
Function used to get length of each example from the dataset.
This argument can be used only when the dataset is a Speechbrain DynamicItemDataset object.
Can be anything: e.g. lambda x: x["duration"]*16000 returns number of samples
if duration key in the annotation is in seconds and the file has 16kHz sampling freq.
shuffle : bool
Whether or not shuffle examples between each epoch.
bucket_boundaries : list
Overrides bucket_length_multiplier and left_bucket_length by specifying manually
the buckets right boundaries.
lengths_list: list
Overrides length_func by passing a list containing the length of each example
in the dataset. This argument must be set when the dataset is a plain
Pytorch Dataset object and not a DynamicItemDataset object as length_func
cannot be used on Pytorch Datasets.
epoch : int
The epoch to start at.
drop_last : bool
If ``True``, the sampler will drop the last examples which
have not been grouped.
"""
def __init__(
self,
dataset,
max_batch_length: int,
left_bucket_length: int,
bucket_length_multiplier: float = 1.1,
length_func=lambda x: x["duration"],
shuffle: bool = True,
bucket_boundaries: List[int] = [],
lengths_list: List[int] = None,
seed: int = 42,
epoch: int = 0,
drop_last: bool = False,
):
self._dataset = dataset
self._ex_lengths = {}
ex_ids = self._dataset.data_ids
if lengths_list is not None:
# take length of examples from this argument and bypass length_key
for indx in range(len(lengths_list)):
self._ex_lengths[str(indx)] = lengths_list[indx]
else:
# use length func
if not isinstance(dataset, DynamicItemDataset):
raise NotImplementedError(
"Dataset should be a Speechbrain DynamicItemDataset when using length function"
)
for indx in range(len(self._dataset)):
self._ex_lengths[str(indx)] = length_func(
self._dataset.data[ex_ids[indx]]
)
if bucket_boundaries is not None:
if not all([x >= 1 for x in bucket_boundaries]):
raise ValueError(
"All elements in bucket boundaries should be >= 1."
)
if not len(set(bucket_boundaries)) == len(bucket_boundaries):
raise ValueError(
"Bucket_boundaries should not contain duplicates."
)
self._bucket_boundaries = np.array(
self._get_data_boundaries(
max_batch_length=max_batch_length,
bucket_boundaries=bucket_boundaries,
left_bucket_length=left_bucket_length,
bucket_length_multiplier=bucket_length_multiplier,
)
)
self._max_batch_length = max_batch_length
self._shuffle = shuffle
self._seed = seed
self._drop_last = drop_last
# Calculate bucket lengths
self._bucket_lens = [
max(1, int(max_batch_length / self._bucket_boundaries[i]))
for i in range(len(self._bucket_boundaries))
] + [1]
self._epoch = epoch
self._generate_batches()
def _get_data_boundaries(
self,
max_batch_length: int,
bucket_boundaries: List[int],
left_bucket_length: int,
bucket_length_multiplier: float,
) -> List[int]:
if not bucket_boundaries:
if left_bucket_length <= 0:
raise ValueError(
"left_bucket_length must be >0 if no bucket_boundaries set"
)
if bucket_length_multiplier < 1.0:
raise ValueError(
"bucket_length_multiplier must be >1.0 if no bucket_boundaries set"
)
bucket_boundaries = {left_bucket_length}
bucket_boundary = float(left_bucket_length)
while True:
bucket_boundary *= bucket_length_multiplier
if bucket_boundary >= max_batch_length:
break
bucket_boundaries.add(bucket_boundary)
return list(sorted(bucket_boundaries))
def _generate_batches(self):
logger.info("DynamicBatchSampler: Generating dynamic batches")
if self._shuffle:
# deterministically shuffle based on epoch and seed
g = torch.Generator()
g.manual_seed(self._seed + self._epoch)
sampler = torch.randperm(len(self._dataset), generator=g).tolist() # type: ignore
else:
sampler = range(len(self._dataset)) # type: ignore
self._batches = []
bucket_batches = [[] for i in self._bucket_lens]
bucket_stats = [0 for i in self._bucket_lens]
for idx in sampler:
item_len = self._ex_lengths[str(idx)]
bucket_id = np.searchsorted(self._bucket_boundaries, item_len)
bucket_batches[bucket_id].append(idx)
bucket_stats[bucket_id] += 1
if len(bucket_batches[bucket_id]) >= self._bucket_lens[bucket_id]:
self._batches.append(bucket_batches[bucket_id])
bucket_batches[bucket_id] = []
# Dump remaining batches - we might even want to shuffle those
if not self._drop_last:
for batch in bucket_batches:
if batch:
self._batches.append(batch)
if self._epoch == 0: # only log at first epoch
logger.info(
"DynamicBatchSampler: Created {} batches, {} buckets used.".format(
len(self._batches), len(self._bucket_boundaries)
)
)
boundaries = [0] + self._bucket_boundaries.tolist()
for i in range(len(self._bucket_boundaries)):
logger.info(
"DynamicBatchSampler: Bucket {} with boundary {}-{} and batch_size {} has {} examples.".format(
i,
np.around(boundaries[i], 2),
np.around(boundaries[i + 1], 2),
self._bucket_lens[i],
bucket_stats[i],
)
)
def __iter__(self):
for batch in self._batches:
yield batch
if self._shuffle: # re-generate batches only if shuffling
self._generate_batches()
[docs] def set_epoch(self, epoch):
"""
You can also just access self.epoch, but we maintain this interface
to mirror torch.utils.data.distributed.DistributedSampler
"""
self._epoch = epoch
self._generate_batches()
def __len__(self):
return len(self._batches)
# Heavily inspired by Catalyst, which is under Apache 2.0 licence.
# https://github.com/catalyst-team/catalyst/blob/51428d7756e62b9b8ee5379f38e9fd576eeb36e5/catalyst/data/sampler.py#L522
[docs]class DistributedSamplerWrapper(DistributedSampler):
"""This wrapper allows using any sampler with Distributed Data Parallel (DDP) correctly.
Passing blindly the sampler to each DDP process will cause to have access
within each process to all the data in the dataset instead of only a subset
of it which is unique to each process. This wrapper prevents this and
allows to use only a subset of the original data for each process.
NOTE
----
This is is automatically applied to any sampler in the Brain class when DDP
training is used.
"""
def __init__(self, sampler, *args, **kwargs):
# DistributedSampler only calls len() on dataset
# so a sampler is fine to pass there, as well.
super().__init__(dataset=sampler, *args, **kwargs)
self.sampler = sampler
def __iter__(self):
# It is easiest to use a random access interface to the wrapped
# sampler's indices, so we just fetch all indices from the wrapped
# sampler
sampler_indices = list(self.sampler.__iter__())
indices_of_indices = super().__iter__()
# Itemgetter fetches the wrapped sampler indices from the positions
# pointed to by DistributedSampler
return iter(itemgetter(*indices_of_indices)(sampler_indices))
[docs] def set_epoch(self, epoch):
"""Pass set_epoch() through to DistributedSampler and the wrapper one"""
super().set_epoch(epoch)
if hasattr(self.sampler, "set_epoch"):
self.sampler.set_epoch(epoch)