386 lines
15 KiB
Python
386 lines
15 KiB
Python
from __future__ import division
|
|
import time
|
|
|
|
import math
|
|
import random
|
|
from contextlib import contextmanager
|
|
from copy import deepcopy
|
|
|
|
import torch
|
|
if torch.distributed.is_available():
|
|
from torch.distributed import get_world_size, get_rank
|
|
|
|
from .batch import Batch
|
|
from .dataset import Dataset
|
|
|
|
|
|
class RandomShuffler(object):
|
|
"""Use random functions while keeping track of the random state to make it
|
|
reproducible and deterministic."""
|
|
|
|
def __init__(self, random_state=None):
|
|
self._random_state = random_state
|
|
self.extra = 0
|
|
if self._random_state is None:
|
|
self._random_state = random.getstate()
|
|
|
|
@contextmanager
|
|
def use_internal_state(self):
|
|
"""Use a specific RNG state."""
|
|
old_state = random.getstate()
|
|
random.setstate(self._random_state)
|
|
yield
|
|
self._random_state = random.getstate()
|
|
random.setstate(old_state)
|
|
|
|
@property
|
|
def random_state(self):
|
|
return deepcopy(self._random_state)
|
|
|
|
@random_state.setter
|
|
def random_state(self, s):
|
|
self._random_state = s
|
|
|
|
def __call__(self, data, subsample=None):
|
|
"""Shuffle and return a new list."""
|
|
with self.use_internal_state():
|
|
return random.sample(data, len(data))
|
|
|
|
def set_epoch(self, epoch):
|
|
self.epoch = epoch
|
|
|
|
|
|
class DistributedShuffler:
|
|
|
|
def __init__(self, num_replicas=None, rank=None):
|
|
if num_replicas is None:
|
|
num_replicas = get_world_size()
|
|
if rank is None:
|
|
rank = get_rank()
|
|
self.num_replicas = num_replicas
|
|
self.rank = rank
|
|
self.epoch = 0
|
|
self.extra = 0
|
|
|
|
def __call__(self, data, subsample=True):
|
|
|
|
# deterministically shuffle based on epoch
|
|
g = torch.Generator()
|
|
g.manual_seed(self.epoch)
|
|
|
|
indices = list(torch.randperm(len(data), generator=g))
|
|
if not subsample:
|
|
return [data[i] for i in indices]
|
|
return [data[i] for i in self.subsample(indices)]
|
|
|
|
def subsample(self, indices):
|
|
# add extra samples to make it evenly divisible
|
|
num_samples = int(math.ceil(len(indices) * 1.0 / self.num_replicas))
|
|
total_size = num_samples * self.num_replicas
|
|
extras = indices[:(total_size - len(indices))]
|
|
self.extra = len(extras)
|
|
indices += extras
|
|
assert len(indices) == total_size
|
|
|
|
# subsample
|
|
offset = num_samples * self.rank
|
|
indices = indices[offset:offset + num_samples]
|
|
assert len(indices) == num_samples
|
|
|
|
return indices
|
|
|
|
def set_epoch(self, epoch):
|
|
self.epoch = epoch
|
|
|
|
|
|
class Iterator(object):
|
|
"""Defines an iterator that loads batches of data from a Dataset.
|
|
|
|
Attributes:
|
|
dataset: The Dataset object to load Examples from.
|
|
batch_size: Batch size.
|
|
batch_size_fn: Function of three arguments (new example to add, current
|
|
count of examples in the batch, and current effective batch size)
|
|
that returns the new effective batch size resulting from adding
|
|
that example to a batch. This is useful for dynamic batching, where
|
|
this function would add to the current effective batch size the
|
|
number of tokens in the new example.
|
|
sort_key: A key to use for sorting examples in order to batch together
|
|
examples with similar lengths and minimize padding. The sort_key
|
|
provided to the Iterator constructor overrides the sort_key
|
|
attribute of the Dataset, or defers to it if None.
|
|
train: Whether the iterator represents a train set.
|
|
repeat: Whether to repeat the iterator for multiple epochs.
|
|
shuffle: Whether to shuffle examples between epochs.
|
|
sort: Whether to sort examples according to self.sort_key.
|
|
Note that repeat, shuffle, and sort default to train, train, and
|
|
(not train).
|
|
sort_within_batch: Whether to sort (in descending order according to
|
|
self.sort_key) within each batch. If None, defaults to self.sort.
|
|
If self.sort is True and this is False, the batch is left in the
|
|
original (ascending) sorted order.
|
|
device: Device to create batches on. Use -1 for CPU and None for the
|
|
currently active GPU device.
|
|
"""
|
|
|
|
def __init__(self, dataset, batch_size, sort_key=None, device=None,
|
|
batch_size_fn=None, train=True,
|
|
repeat=None, shuffle=None, sort=None, reverse=False,
|
|
sort_within_batch=None, distributed=False, num_replicas=None, rank=None):
|
|
self.batch_size, self.train, self.dataset = batch_size, train, dataset
|
|
self.batch_size_fn = batch_size_fn
|
|
self.iterations = 0
|
|
self.epoch = 0
|
|
self.reverse = reverse
|
|
self.repeat = train if repeat is None else repeat
|
|
self.shuffle = train if shuffle is None else shuffle
|
|
self.sort = not train if sort is None else sort
|
|
if sort_within_batch is None:
|
|
self.sort_within_batch = self.sort
|
|
else:
|
|
self.sort_within_batch = sort_within_batch
|
|
if sort_key is None:
|
|
self.sort_key = dataset.sort_key
|
|
else:
|
|
self.sort_key = sort_key
|
|
self.device = device
|
|
|
|
self.distributed = distributed
|
|
if distributed:
|
|
self.random_shuffler = DistributedShuffler(num_replicas=num_replicas, rank=rank)
|
|
else:
|
|
self.random_shuffler = RandomShuffler()
|
|
|
|
# For state loading/saving only
|
|
self._iterations_this_epoch = 0
|
|
self._random_state_this_epoch = None
|
|
self._restored_from_state = False
|
|
|
|
@classmethod
|
|
def splits(cls, datasets, batch_sizes=None, **kwargs):
|
|
"""Create Iterator objects for multiple splits of a dataset.
|
|
|
|
Arguments:
|
|
datasets: Tuple of Dataset objects corresponding to the splits. The
|
|
first such object should be the train set.
|
|
batch_sizes: Tuple of batch sizes to use for the different splits,
|
|
or None to use the same batch_size for all splits.
|
|
Remaining keyword arguments: Passed to the constructor of the
|
|
iterator class being used.
|
|
"""
|
|
if batch_sizes is None:
|
|
batch_sizes = [kwargs.pop('batch_size')] * len(datasets)
|
|
ret = []
|
|
for i in range(len(datasets)):
|
|
train = i == 0
|
|
ret.append(cls(
|
|
datasets[i], batch_size=batch_sizes[i], train=train, **kwargs))
|
|
return tuple(ret)
|
|
|
|
def data(self):
|
|
"""Return the examples in the dataset in order, sorted, or shuffled."""
|
|
if self.sort:
|
|
xs = sorted(self.dataset, key=self.sort_key, reverse=self.reverse)
|
|
if self.distributed:
|
|
xs = [xs[i] for i in self.random_shuffler.subsample(list(range(len(xs))))]
|
|
elif self.shuffle:
|
|
xs = self.random_shuffler(list(self.dataset))
|
|
else:
|
|
xs = self.dataset
|
|
self.extra = self.random_shuffler.extra
|
|
return xs
|
|
|
|
def init_epoch(self):
|
|
"""Set up the batch generator for a new epoch."""
|
|
if not self.distributed:
|
|
if self._restored_from_state:
|
|
self.random_shuffler.random_state = self._random_state_this_epoch
|
|
else:
|
|
self._random_state_this_epoch = self.random_shuffler.random_state
|
|
|
|
self.create_batches()
|
|
|
|
if not self.distributed:
|
|
if self._restored_from_state:
|
|
self._restored_from_state = False
|
|
else:
|
|
self._iterations_this_epoch = 0
|
|
else:
|
|
self._iterations_this_epoch = 0
|
|
|
|
|
|
if not self.repeat:
|
|
self.iterations = 0
|
|
self.epoch += 1
|
|
if self.distributed:
|
|
self.random_shuffler.set_epoch(self.epoch)
|
|
|
|
def create_batches(self):
|
|
self.batches = batch(self.data(), self.batch_size, self.batch_size_fn)
|
|
|
|
def __len__(self):
|
|
if self.batch_size_fn is not None:
|
|
raise NotImplementedError
|
|
return math.ceil(len(self.dataset) / self.batch_size)
|
|
|
|
def __iter__(self):
|
|
while True:
|
|
self.init_epoch()
|
|
for idx, minibatch in enumerate(self.batches):
|
|
# fast-forward if loaded from state
|
|
if self._iterations_this_epoch > idx:
|
|
continue
|
|
self.iterations += 1
|
|
self._iterations_this_epoch += 1
|
|
if self.sort_within_batch:
|
|
# NOTE: `rnn.pack_padded_sequence` requires that a minibatch
|
|
# be sorted by decreasing order, which requires reversing
|
|
# relative to typical sort keys
|
|
if self.sort:
|
|
minibatch.reverse()
|
|
else:
|
|
minibatch.sort(key=self.sort_key, reverse=self.reverse)
|
|
b = Batch(minibatch, self.dataset, self.device,
|
|
self.train)
|
|
yield b
|
|
if not self.repeat:
|
|
return
|
|
|
|
def state_dict(self):
|
|
d = {"iterations": self.iterations}
|
|
if not self.distributed:
|
|
d.update({
|
|
"iterations_this_epoch": self._iterations_this_epoch,
|
|
"random_state_this_epoch": self._random_state_this_epoch
|
|
})
|
|
|
|
def load_state_dict(self, state_dict):
|
|
self.iterations = state_dict["iterations"]
|
|
self._iterations_this_epoch = state_dict["iterations_this_epoch"]
|
|
|
|
if not self.distributed:
|
|
self._random_state_this_epoch = state_dict["random_state_this_epoch"]
|
|
self._restored_from_state = True
|
|
|
|
|
|
class BPTTIterator(Iterator):
|
|
"""Defines an iterator for language modeling tasks that use BPTT.
|
|
|
|
Provides contiguous streams of examples together with targets that are
|
|
one timestep further forward, for language modeling training with
|
|
backpropagation through time (BPTT). Expects a Dataset with a single
|
|
example and a single field called 'text' and produces Batches with text and
|
|
target attributes.
|
|
|
|
Attributes:
|
|
dataset: The Dataset object to load Examples from.
|
|
batch_size: Batch size.
|
|
bptt_len: Length of sequences for backpropagation through time.
|
|
sort_key: A key to use for sorting examples in order to batch together
|
|
examples with similar lengths and minimize padding. The sort_key
|
|
provided to the Iterator constructor overrides the sort_key
|
|
attribute of the Dataset, or defers to it if None.
|
|
train: Whether the iterator represents a train set.
|
|
repeat: Whether to repeat the iterator for multiple epochs.
|
|
shuffle: Whether to shuffle examples between epochs.
|
|
sort: Whether to sort examples according to self.sort_key.
|
|
Note that repeat, shuffle, and sort default to train, train, and
|
|
(not train).
|
|
device: Device to create batches on. Use -1 for CPU and None for the
|
|
currently active GPU device.
|
|
"""
|
|
|
|
def __init__(self, dataset, batch_size, bptt_len, **kwargs):
|
|
self.bptt_len = bptt_len
|
|
super(BPTTIterator, self).__init__(dataset, batch_size, **kwargs)
|
|
|
|
def __len__(self):
|
|
return math.ceil((len(self.dataset[0].text) / self.batch_size - 1) /
|
|
self.bptt_len)
|
|
|
|
def __iter__(self):
|
|
text = self.dataset[0].text
|
|
TEXT = self.dataset.fields['text']
|
|
TEXT.eos_token = None
|
|
text = text + ([TEXT.pad_token] * int(math.ceil(len(text) / self.batch_size) *
|
|
self.batch_size - len(text)))
|
|
data = TEXT.numericalize(
|
|
[text], device=self.device, train=self.train)
|
|
data = data.view(self.batch_size, -1).t().contiguous()
|
|
dataset = Dataset(examples=self.dataset.examples, fields=[
|
|
('text', TEXT), ('target', TEXT)])
|
|
while True:
|
|
for i in range(0, len(self) * self.bptt_len, self.bptt_len):
|
|
seq_len = min(self.bptt_len, len(data) - i - 1)
|
|
yield Batch.fromvars(
|
|
dataset, self.batch_size, train=self.train,
|
|
text=data[i:i + seq_len],
|
|
target=data[i + 1:i + 1 + seq_len])
|
|
if not self.repeat:
|
|
raise StopIteration
|
|
|
|
|
|
class BucketIterator(Iterator):
|
|
"""Defines an iterator that batches examples of similar lengths together.
|
|
|
|
Minimizes amount of padding needed while producing freshly shuffled
|
|
batches for each new epoch. See pool for the bucketing procedure used.
|
|
"""
|
|
|
|
def create_batches(self):
|
|
if self.sort:
|
|
self.batches = batch(self.data(), self.batch_size,
|
|
self.batch_size_fn, repeat=self.repeat)
|
|
else:
|
|
self.batches = pool(self.data(), self.batch_size,
|
|
self.sort_key, self.batch_size_fn,
|
|
random_shuffler=self.random_shuffler, repeat=self.repeat,
|
|
reverse=self.reverse, shuffle=self.shuffle)
|
|
|
|
|
|
def batch(data, batch_size, batch_size_fn=None, repeat=False):
|
|
"""Yield elements from data in chunks of batch_size."""
|
|
if batch_size_fn is None:
|
|
def batch_size_fn(new, count, sofar):
|
|
return count
|
|
minibatch = []
|
|
size_so_far = 0
|
|
for ex in data:
|
|
minibatch.append(ex)
|
|
size_so_far = batch_size_fn(ex, len(minibatch), size_so_far)
|
|
if size_so_far == batch_size:
|
|
yield minibatch
|
|
minibatch, size_so_far = [], 0
|
|
elif size_so_far > batch_size:
|
|
if len(minibatch) == 1: # if we only have one really big example
|
|
yield minibatch
|
|
minibatch, size_so_far = [], 0
|
|
else:
|
|
yield minibatch[:-1]
|
|
minibatch, size_so_far = minibatch[-1:], batch_size_fn(ex, 1, 0)
|
|
if size_so_far > batch_size: # if we add a really big example that needs to be on its own to a batch
|
|
yield minibatch
|
|
minibatch, size_so_far = [], 0
|
|
if minibatch:
|
|
yield minibatch
|
|
|
|
|
|
def pool(data, batch_size, key, batch_size_fn=lambda new, count, sofar: count,
|
|
random_shuffler=None, reverse=False, shuffle=False, repeat=False, leftovers=None):
|
|
"""Sort within buckets, then batch, then shuffle batches.
|
|
|
|
Partitions data into chunks of size 100*batch_size, sorts examples within
|
|
each chunk using sort_key, then batch these examples and shuffle the
|
|
batches.
|
|
"""
|
|
if random_shuffler is None:
|
|
random_shuffler = random.shuffle
|
|
for p in batch(data, batch_size * 100, batch_size_fn):
|
|
p_batch = batch(sorted(p, key=key, reverse=reverse), batch_size, batch_size_fn, repeat=repeat)
|
|
if shuffle:
|
|
for b in random_shuffler(list(p_batch), subsample=False):
|
|
yield b
|
|
else:
|
|
for b in list(p_batch):
|
|
yield b
|