File size: 6,975 Bytes
fb11af9 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 | # Copyright 2025 Bytedance Ltd. and/or its affiliates
#
# 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 copy
import sys
import traceback
from collections import deque
from typing import TYPE_CHECKING, Any, Callable, Dict, Generator, Iterator, Optional
from ..utils import logging
logger = logging.get_logger(__name__)
if TYPE_CHECKING:
from .batching_strategy import BaseBatchingStrategy
class DynamicBatchSizeDataLoader:
"""Dynamic batch DataLoader.
Args:
dataloader: torch DataLoader
batching_strategy: dynamic batch strategy
collate_fn: DataLoader collate_fn, collate data after get data from batching_strategy
num_micro_batch: num_micro_batch, if num_micro_batch == 1, return micro_batch for gradient accumulation
length: length of dataloader, if length == -1, length = sys.maxsize, default len(dataloader)
drop_last: if True, drop last batch if batch size < num_micro_batch
"""
def __init__(
self,
dataloader: Any,
batching_strategy: "BaseBatchingStrategy",
collate_fn: Optional[Callable] = None,
num_micro_batch: int = 1,
length: int = 0,
drop_last: bool = True,
) -> None:
self.batching_strategy = batching_strategy
self.num_micro_batch = num_micro_batch
self.dataloader_item_buffer = deque()
self.item_buffer = deque()
self.step = 0
self._collate_fn = collate_fn
self._dataloader = dataloader
self._drop_last = drop_last
self._data_iter: Iterator
self._resume = False
self._batch_data_iter: Generator
if length > 0:
self._length = length
elif length == -1:
self._length = sys.maxsize
else:
self._length = len(self._dataloader)
def __len__(self):
if self._length:
return self._length
else:
raise RuntimeError("length must set at init. before call len()")
def __iter__(self) -> Iterator:
if not self._resume:
self.step = 0
self._data_iter = iter(self._dataloader)
self._batch_data_iter = self.batch_data_generator()
self._resume = False
return self
def __next__(self):
return next(self._batch_data_iter)
def batch_data_generator(self):
batch = []
while True:
if self._length and self.step >= self._length:
return
if self.batching_strategy.is_full_filled():
micro_batch = self.batching_strategy.get_micro_batch(self.step)
if self._collate_fn:
micro_batch = self._collate_fn(micro_batch)
batch.append(micro_batch)
if len(batch) == self.num_micro_batch:
yield batch
self.step += 1
batch = []
try:
processing_item = next(self._data_iter)
except Exception as e:
if isinstance(e, StopIteration):
if self.step < self._length:
# call iter until reach length
self._data_iter = iter(self._dataloader)
processing_item = next(self._data_iter)
elif not self._drop_last and not self.batching_strategy.empty():
while not self.batching_strategy.empty():
micro_batch = self.batching_strategy.get_micro_batch(self.step)
if self._collate_fn:
micro_batch = self._collate_fn(micro_batch)
batch.append(micro_batch)
if len(batch) == self.num_micro_batch:
yield batch
self.step += 1
batch = []
while len(batch) < self.num_micro_batch:
padding_batch = copy.deepcopy(micro_batch)
padding_batch["padding_flag"] = True
batch.append(padding_batch)
yield batch
self.step += 1
return
else:
return
else:
logger.error(f"DynamicBatchDataset iter data exception: {e} \n{traceback.format_exc()}")
raise
# put processing_item to buffer
if isinstance(processing_item, dict):
processing_item = [processing_item]
for item in processing_item:
self.batching_strategy.put_item(item)
def state_dict(self):
# save state
state = self.__dict__.copy()
# remove internal fields
for k in list(state.keys()):
if k.startswith("_"):
del state[k]
# save dataloader state
if hasattr(self._dataloader, "state_dict"):
state["dataloader_state"] = self._dataloader.state_dict()
elif hasattr(self._dataloader, "__getstate__"):
state["dataloader_state"] = self._dataloader.__getstate__()
if hasattr(self.batching_strategy, "state_dict"):
state["batching_strategy_state"] = self.batching_strategy.state_dict() # type: ignore
del state["batching_strategy"]
return copy.deepcopy(state)
def load_state_dict(self, state: Dict[str, Any]):
if state["num_micro_batch"] != self.num_micro_batch:
logger.warning(
f"num_micro_batch changed: [ {state['num_micro_batch']} -> {self.num_micro_batch} ], will clear prefetch buffer"
)
del state["num_micro_batch"]
self.__dict__.update(state)
self._resume = True
if hasattr(self._dataloader, "load_state_dict"):
self._dataloader.load_state_dict(state["dataloader_state"])
elif hasattr(self._dataloader, "__getstate__"):
self._dataloader.__setstate__(state["dataloader_state"])
if "batching_strategy_state" in state:
self.batching_strategy.load_state_dict( # type: ignore
state["batching_strategy_state"]
)
del state["batching_strategy_state"]
self._data_iter = iter(self._dataloader)
self._batch_data_iter = self.batch_data_generator()
|