# Copyright 2022 Cerebras Systems.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import logging
import torch
import cerebras_pytorch as cstorch
import cerebras_pytorch.distributed as dist
from cerebras_pytorch.distributed import get_worker_state
from modelzoo.common.pytorch.input_utils import get_streaming_batch_size
from modelzoo.vision.pytorch.input.utils import create_worker_cache
from .readers import H5Reader, Mixture
from .samplers import CBSampler
[docs]class RestartableDataLoader(torch.utils.data.DataLoader):
"""
The state we care about for allowing deterministic restart of instances
of `HDF5Dataset` is the total number of samples streamed globally,
which gets consumed by the sampler. Accordingly each worker saves the number
of samples that it has streamed in `state_dict()`. We aggregate these
together via summation to save the global number of samples streamed across
all workers, which is the same thing that is used to set the state of the
sampler on state dict load.
"""
[docs] def __init__(self, *args, **kwargs):
# keep track of how many samples were streamed in the previous portion
# of the run so that we can track cumulative samples streamed in the
# state_dict
self.previous_samples_streamed = 0
super().__init__(*args, **kwargs)
[docs] def state_dict(self):
"""
Save number of samples streamed for current worker
"""
worker_state = get_worker_state()
return {
"samples_streamed": worker_state.samples_streamed,
"previous_samples_streamed": self.previous_samples_streamed,
}
[docs] def load_state_dict(self, state_dict):
"""
Set sampler state with the total number of samples streamed globally
"""
self.validate_state_dict(state_dict)
self.previous_samples_streamed = state_dict["samples_streamed"]
if state_dict["seed"] != self.dataset.seed:
raise ValueError(
f"shuffle seed {self.dataset.seed} doesn't match the seed used "
f"for the previous portion of the run {state_dict['seed']}"
)
self.dataset.sampler.set_state(state_dict["samples_streamed"])
[docs] def aggregate_state_dict(self, worker_states):
"""
Sum samples streamed across all workers to get the number of samples
streamed globally
"""
return {
"samples_streamed": sum(
sd["samples_streamed"] for sd in worker_states
)
+ worker_states[0]["previous_samples_streamed"],
"seed": self.dataset.seed,
}
[docs] def deaggregate_state_dict(self, aggregated_state_dict):
"""
No deaggregation needed since the sampler needs the global number of
samples streamed
"""
return aggregated_state_dict
@staticmethod
def validate_state_dict(sd):
if len(sd) != 2 or "samples_streamed" not in sd or "seed" not in sd:
raise RuntimeError(
"The keys in state_dict must be 'samples_streamed' and 'seed', "
f"found {sd.keys()}. This means that the dataloader state in "
"the checkpoint you are loading from is not compatible with "
"the dataloader currently in use. Consider re-running without "
"loading the dataloader state."
)
[docs]class HDF5Dataset(torch.utils.data.Dataset):
"""
Dynamically read samples from disk for using mapping paradigms.
It supports two different data formats on disk. The first is data stored
in an H5 file in the shape `(num_tokens,)`, i.e. a series of documents
tokenized and concatenated together. We call this format the 'corpus' format
The second format is H5 data of shape `(num_sequences, ...)`, i.e. data has
already been tokenized and split into sequences. We call this format the
'sample' format.
The corpus format supports flexible choice of MSL backed by a single copy of
the data on disk. Both formats support deterministic restart, and a data
order that is independent of the configuration of the cluster you are
running on. I.e. you can pause a run, increase or decrease the number of
systems you are running on, and restart the run with no change in data
order.
When used in combination with shuffling, this implementation relies on
random access reads to disk to dynamically split samples into sequences
and shuffle. Users with unusually slow storage should look out for data
loading bottlenecks and might consider using `use_worker_cache=True` if
disk access is indeed a bottleneck.
Args:
params (dict): a dictionary containing the following fields:
- "data_dir" (str or list[str]): the path to the HDF5 files.
Exactly one of "data_dir" or "mixture" must be specified.
- "batch_size" (int): batch size
- "shuffle" (bool): whether or not to shuffle the dataset. Defaults
to `False`
- "shuffle_seed" (int): seed used for deterministic shuffling.
Defaults to 0.
- "use_worker_cache" (bool): whether or not to copy data to storage
that is directly attached to each individual worker node.
Useful when your network storage is unusually slow, but
otherwise discouraged.
- "max_sequence_length" (int): the sequence length of samples
produced by the dataloader. When using the 'corpus' data format,
the same preprocessed data will work with any max sequence
length, so this may be set at runtime. When using the 'sample'
format this must be set to `None`.
- "data_subset" (str): an optional specification to only consider a
subset of the full dataset, useful for sequence length
scheduling and multi-epoch testing. Expected to be a comma
separated list of ranges, e.g. '0.0-0.5' or '0.1-0.3,0.7-1.0'.
Specifying '0.0-0.5' creates a dataset from the first half of
the data on disk and disregards the second half.
- "mixture" list[dict]: an optional specification of multiple
datasets to mix over to create one single weighted combination.
Each element must be a dictionary containing keys `data_dir`
and `weight`. `data_dir` serves the same purpose as mentioned
above. `weight` defines the probability with which this dataset
should be sampled from. Weights are normalized to sum to 1.
Optionally, the dictionary may also contain a `data_subset`
field which functions the same as the `data_subset` argument
above.
- "drop_last" (bool): similar to the PyTorch drop_last setting
except that samples that when set to `True`, samples that would
have been dropped at the end of one epoch are yielded at the
start of the next epoch so that there is no data loss. This is
necessary for a data ordering that is independent of the
distributed setup being used.
- "num_samples" (int): the number of samples to shuffle over (if
shuffling is enabled). In multi-epoch training, it is common to
set this to the total number of samples that you plan to train
on so that epochs are not sequential but instead shuffled
together for potentially improved convergence.
- "sort_files" (bool): whether or not the reader should sort the input
files. This is included for backwards compatibility and should
almost always be set to `True`.
"""
[docs] def __init__(self, params):
self.use_worker_cache = params.get("use_worker_cache", False)
self.msl = params.get("max_sequence_length", None)
shuffle = params.get("shuffle", False)
self._seed = params.get("shuffle_seed", 0)
data_dir = params.get("data_dir", None)
mixture_params = params.get("mixture", None)
batch_size = get_streaming_batch_size(params["batch_size"])
micro_batch_size = params.get("micro_batch_size")
drop_last = params.get("drop_last", True)
num_samples = params.get("num_samples", None)
self.sort_files = params.get("sort_files", True)
if data_dir and mixture_params:
raise ValueError(
"you can't specify `data_dir` and `mixture` at the same time"
)
if data_dir is not None:
self.reader = self._set_up_reader(
data_dir, params.get("data_subset", None)
)
else:
self.reader = Mixture(
[
self._set_up_reader(
x["data_dir"], x.get("data_subset", None)
)
for x in mixture_params
],
[x["weight"] for x in mixture_params],
interleave=not shuffle,
seed=self._seed,
)
self.sampler = CBSampler(
self,
shuffle=shuffle,
seed=self._seed,
shard=True,
batch_size=batch_size,
drop_last=drop_last,
num_samples=num_samples,
)
self.map_fn = None
if self.by_sample and shuffle:
logging.warning(
"You have chosen to use the sample data format with shuffling. "
"If you are doing a single-epoch run, it is usually beneficial "
"to shuffle at preprocessing time instead of runtime. On some "
"storage setups, shuffling at runtime can cause performance "
"degredation."
)
@property
def by_sample(self):
return self.reader.by_sample
@property
def seed(self):
return self._seed
def map(self, fn):
if self.map_fn is not None:
raise ValueError(
f"You may only apply one map function to a H5MapDataset"
)
self.map_fn = fn
def _set_up_reader(self, data_dir, subset):
if not isinstance(data_dir, list):
data_dir = [data_dir]
if self.use_worker_cache and cstorch.use_cs() and dist.is_streamer():
data_dir = [create_worker_cache(d) for d in data_dir]
reader = H5Reader(data_dir, self.msl, True, subset, self.sort_files)
return reader
def __getitem__(self, i):
x = self.reader[i]
if self.map_fn is not None:
return self.map_fn(x)
return x
def __len__(self):
return len(self.reader)