Delete main/src/sequence_async_reservoir.py with huggingface_hub
Browse files
main/src/sequence_async_reservoir.py
DELETED
|
@@ -1,278 +0,0 @@
|
|
| 1 |
-
from job_executor2 import JobExecutor2
|
| 2 |
-
import sequence
|
| 3 |
-
|
| 4 |
-
import collections
|
| 5 |
-
from queue import Empty
|
| 6 |
-
import copy
|
| 7 |
-
import time
|
| 8 |
-
|
| 9 |
-
|
| 10 |
-
class Performance:
|
| 11 |
-
def __init__(self):
|
| 12 |
-
self.nb_times = 0
|
| 13 |
-
self.total_time_sec = 0.0
|
| 14 |
-
|
| 15 |
-
def add(self, time_elapsed):
|
| 16 |
-
self.nb_times += 1
|
| 17 |
-
self.total_time_sec += time_elapsed
|
| 18 |
-
|
| 19 |
-
def get_average_time(self):
|
| 20 |
-
if self.nb_times == 0:
|
| 21 |
-
return 0
|
| 22 |
-
|
| 23 |
-
return self.total_time_sec / self.nb_times
|
| 24 |
-
|
| 25 |
-
|
| 26 |
-
class SequenceAsyncReservoir(sequence.Sequence):
|
| 27 |
-
"""
|
| 28 |
-
This sequence will asynchronously process data and keep a reserve of loaded samples
|
| 29 |
-
|
| 30 |
-
The idea is to have long loading processes work in the background while still using as efficiently as possible
|
| 31 |
-
the data that is currently loaded. The data is slowly being replaced by freshly loaded data over time.
|
| 32 |
-
|
| 33 |
-
Jobs are started and results retrieved at the beginning of each epoch
|
| 34 |
-
|
| 35 |
-
This sequence can be interrupted (e.g., after a certain number of batches have been returned). When the sequence
|
| 36 |
-
is restarted, the reservoir will not be emptied.
|
| 37 |
-
"""
|
| 38 |
-
def __init__(
|
| 39 |
-
self,
|
| 40 |
-
source_split,
|
| 41 |
-
max_reservoir_samples,
|
| 42 |
-
function_to_run,
|
| 43 |
-
*,
|
| 44 |
-
min_reservoir_samples=1,
|
| 45 |
-
nb_workers=1,
|
| 46 |
-
max_jobs_at_once=None,
|
| 47 |
-
reservoir_sampler=None,
|
| 48 |
-
collate_fn=sequence.remove_nested_list,
|
| 49 |
-
maximum_number_of_samples_per_epoch=None,
|
| 50 |
-
max_reservoir_replacement_size=None):
|
| 51 |
-
"""
|
| 52 |
-
Args:
|
| 53 |
-
source_split: the source split to iterate
|
| 54 |
-
max_reservoir_samples: the maximum number of samples of the reservoir
|
| 55 |
-
function_to_run: the function to run asynchronously
|
| 56 |
-
min_reservoir_samples: the minimum of samples of the reservoir needed before an output sequence
|
| 57 |
-
can be created
|
| 58 |
-
nb_workers: the number of workers that will process `function_to_run` to fill the reservoir. Must be >= 1
|
| 59 |
-
max_jobs_at_once: the maximum number of jobs that can be started and stored by epoch by the workers.
|
| 60 |
-
If 0, no limit. If None: set to the number of workers
|
| 61 |
-
reservoir_sampler: a sampler that will be used to sample the reservoir or None for sequential sampling
|
| 62 |
-
of the reservoir
|
| 63 |
-
collate_fn: a function to post-process the samples into a single batch, or None if not to be collated
|
| 64 |
-
maximum_number_of_samples_per_epoch: the maximum number of samples that will be generated per epoch.
|
| 65 |
-
If we reach this maximum, the sequence will be interrupted
|
| 66 |
-
max_reservoir_replacement_size: Specify the maximum number of samples replaced in the reservoir by epoch.
|
| 67 |
-
If `None`, we will use the whole result queue. This can be useful to control explicitly how the
|
| 68 |
-
reservoir is updated and depend less on the speed of hardware. Note that to have an effect,
|
| 69 |
-
`max_jobs_at_once` should be greater than `max_reservoir_replacement_size`.
|
| 70 |
-
"""
|
| 71 |
-
super().__init__(source_split)
|
| 72 |
-
self.max_reservoir_samples = max_reservoir_samples
|
| 73 |
-
self.min_reservoir_samples = min_reservoir_samples
|
| 74 |
-
self.function_to_run = function_to_run
|
| 75 |
-
self.nb_workers = nb_workers
|
| 76 |
-
self.collate_fn = collate_fn
|
| 77 |
-
self.max_jobs_at_once = max_jobs_at_once
|
| 78 |
-
if max_jobs_at_once is None:
|
| 79 |
-
self.max_jobs_at_once = nb_workers
|
| 80 |
-
self.reservoir_sampler = reservoir_sampler
|
| 81 |
-
assert min_reservoir_samples <= max_reservoir_samples
|
| 82 |
-
assert min_reservoir_samples >= 1
|
| 83 |
-
|
| 84 |
-
# this is where we store
|
| 85 |
-
self.reservoir = collections.deque(maxlen=max_reservoir_samples)
|
| 86 |
-
self.iter_source = None
|
| 87 |
-
self.iter_reservoir = None
|
| 88 |
-
|
| 89 |
-
self.perf_receiving = Performance()
|
| 90 |
-
self.perf_sending = Performance()
|
| 91 |
-
|
| 92 |
-
self.job_executer = None
|
| 93 |
-
assert nb_workers != 0, 'must have background workers'
|
| 94 |
-
|
| 95 |
-
if max_jobs_at_once is None:
|
| 96 |
-
# default: each worker can push at least one item
|
| 97 |
-
# before blocking
|
| 98 |
-
max_jobs_at_once = nb_workers
|
| 99 |
-
|
| 100 |
-
self.job_executer = JobExecutor2(
|
| 101 |
-
nb_workers=nb_workers,
|
| 102 |
-
function_to_run=self.function_to_run,
|
| 103 |
-
max_queue_size_pin_thread_per_worker=max_jobs_at_once,
|
| 104 |
-
max_queue_size_per_worker=max_jobs_at_once // nb_workers)
|
| 105 |
-
|
| 106 |
-
self.maximum_number_of_samples_per_epoch = maximum_number_of_samples_per_epoch
|
| 107 |
-
self.number_samples_generated = 0
|
| 108 |
-
self.max_reservoir_replacement_size = max_reservoir_replacement_size
|
| 109 |
-
|
| 110 |
-
def subsample(self, nb_samples):
|
| 111 |
-
subsampled_source = self.source_split.subsample(nb_samples)
|
| 112 |
-
return SequenceAsyncReservoir(subsampled_source, self.max_reservoir_samples, self.function_to_run,
|
| 113 |
-
min_reservoir_samples=self.min_reservoir_samples, nb_workers=1,
|
| 114 |
-
max_jobs_at_once=self.max_jobs_at_once,
|
| 115 |
-
reservoir_sampler=copy.deepcopy(self.reservoir_sampler),
|
| 116 |
-
collate_fn=self.collate_fn, maximum_number_of_samples_per_epoch=None,
|
| 117 |
-
max_reservoir_replacement_size=self.max_reservoir_replacement_size)
|
| 118 |
-
|
| 119 |
-
def reservoir_size(self):
|
| 120 |
-
"""
|
| 121 |
-
Returns:
|
| 122 |
-
The current number of samples in the reservoir
|
| 123 |
-
"""
|
| 124 |
-
return len(self.reservoir)
|
| 125 |
-
|
| 126 |
-
def subsample_uids(self, uids, uids_name, new_sampler=None):
|
| 127 |
-
if new_sampler is None:
|
| 128 |
-
sampler_to_use = copy.deepcopy(self.reservoir_sampler)
|
| 129 |
-
else:
|
| 130 |
-
sampler_to_use = copy.deepcopy(new_sampler)
|
| 131 |
-
|
| 132 |
-
subsampled_source = self.source_split.subsample_uids(uids, uids_name, new_sampler)
|
| 133 |
-
|
| 134 |
-
return SequenceAsyncReservoir(subsampled_source, self.max_reservoir_samples, self.function_to_run,
|
| 135 |
-
min_reservoir_samples=self.min_reservoir_samples, nb_workers=1,
|
| 136 |
-
max_jobs_at_once=self.max_jobs_at_once, reservoir_sampler=sampler_to_use,
|
| 137 |
-
collate_fn=self.collate_fn, maximum_number_of_samples_per_epoch=None,
|
| 138 |
-
max_reservoir_replacement_size=self.max_reservoir_replacement_size)
|
| 139 |
-
|
| 140 |
-
def initializer(self):
|
| 141 |
-
if self.iter_source is None:
|
| 142 |
-
self.iter_source = self.source_split.__iter__()
|
| 143 |
-
|
| 144 |
-
self._retrieve_results_and_fill_queue()
|
| 145 |
-
|
| 146 |
-
if len(self.reservoir) < self.min_reservoir_samples:
|
| 147 |
-
# before we can iterate on sample, we must have enough samples in the reservoir!
|
| 148 |
-
self._wait_for_job_completion()
|
| 149 |
-
|
| 150 |
-
def fill_queue(self):
|
| 151 |
-
"""
|
| 152 |
-
Fill the input queue of jobs to be completed
|
| 153 |
-
"""
|
| 154 |
-
nb_queued = 0
|
| 155 |
-
try:
|
| 156 |
-
while not self.job_executer.is_full():
|
| 157 |
-
i = self.iter_source.next_item(blocking=False)
|
| 158 |
-
nb_queued += 1
|
| 159 |
-
|
| 160 |
-
time_blocked_start = time.perf_counter()
|
| 161 |
-
self.job_executer.put(i)
|
| 162 |
-
time_blocked_end = time.perf_counter()
|
| 163 |
-
self.perf_sending.add(time_blocked_end - time_blocked_start)
|
| 164 |
-
#print('RESERVOIR_QUEUE_FILL job_executer.put')
|
| 165 |
-
except StopIteration:
|
| 166 |
-
# we are done! Reset the input iterator
|
| 167 |
-
self.iter_source = self.source_split.__iter__()
|
| 168 |
-
|
| 169 |
-
def _retrieve_results_and_fill_queue(self):
|
| 170 |
-
"""
|
| 171 |
-
Retrieve results from the output queue
|
| 172 |
-
"""
|
| 173 |
-
# first make sure we can't be `starved`, so fill the input queue
|
| 174 |
-
self.fill_queue()
|
| 175 |
-
|
| 176 |
-
nb_samples_replaced = 0
|
| 177 |
-
|
| 178 |
-
# retrieve the results from the output queue and fill the reservoir
|
| 179 |
-
while not self.job_executer.pin_memory_queue.empty():
|
| 180 |
-
try:
|
| 181 |
-
time_blocked_start = time.perf_counter()
|
| 182 |
-
metadata, items = self.job_executer.pin_memory_queue.get()
|
| 183 |
-
if items is None:
|
| 184 |
-
# the job failed!
|
| 185 |
-
continue
|
| 186 |
-
time_blocked_end = time.perf_counter()
|
| 187 |
-
self.perf_receiving.add(time_blocked_end - time_blocked_start)
|
| 188 |
-
self.reservoir.append(items)
|
| 189 |
-
nb_samples_replaced += 1
|
| 190 |
-
|
| 191 |
-
if self.max_reservoir_replacement_size is not None and nb_samples_replaced >= self.max_reservoir_replacement_size:
|
| 192 |
-
# we have reached the maximum replacement, so stop here
|
| 193 |
-
break
|
| 194 |
-
except Empty:
|
| 195 |
-
break
|
| 196 |
-
|
| 197 |
-
def _wait_for_job_completion(self):
|
| 198 |
-
"""
|
| 199 |
-
Block the processing until we have enough result in the reservoir
|
| 200 |
-
"""
|
| 201 |
-
while len(self.reservoir) < self.min_reservoir_samples:
|
| 202 |
-
self._retrieve_results_and_fill_queue()
|
| 203 |
-
|
| 204 |
-
def __iter__(self):
|
| 205 |
-
self.initializer()
|
| 206 |
-
iterator = SequenceAsyncReservoirIterator(self, copy.deepcopy(self.reservoir_sampler))
|
| 207 |
-
|
| 208 |
-
# keep track of average performance of receiving and sending objects to the workers
|
| 209 |
-
self.perf_receiving = Performance()
|
| 210 |
-
self.perf_sending = Performance()
|
| 211 |
-
return iterator
|
| 212 |
-
|
| 213 |
-
def close(self):
|
| 214 |
-
"""
|
| 215 |
-
Finish and join the existing pool processes
|
| 216 |
-
"""
|
| 217 |
-
if self.job_executer is not None:
|
| 218 |
-
self.job_executer.close()
|
| 219 |
-
|
| 220 |
-
if self.source_split is not None:
|
| 221 |
-
self.source_split.close()
|
| 222 |
-
|
| 223 |
-
|
| 224 |
-
class SequenceAsyncReservoirIterator(sequence.SequenceIterator):
|
| 225 |
-
"""
|
| 226 |
-
Iterate through the SequenceAsyncReservoir sequence
|
| 227 |
-
"""
|
| 228 |
-
def __init__(self, base_sequence, reservoir_sampler):
|
| 229 |
-
super().__init__()
|
| 230 |
-
self.base_sequence = base_sequence
|
| 231 |
-
self.reservoir_sampler = reservoir_sampler
|
| 232 |
-
|
| 233 |
-
self._reset_iter_reservoir()
|
| 234 |
-
self.number_samples_generated = 0
|
| 235 |
-
|
| 236 |
-
def _reset_iter_reservoir(self):
|
| 237 |
-
# make sure the reservoir is not changed during iteration, so take a copy
|
| 238 |
-
# when the iterator is created
|
| 239 |
-
self.reservoir_copy = list(self.base_sequence.reservoir)
|
| 240 |
-
|
| 241 |
-
# print(f'LAST_receiving=', self.base_sequence.perf_receiving.get_average_time())
|
| 242 |
-
# Note: in case the queue receive stalls the sequence, it is worth trying a different
|
| 243 |
-
# sharing strategy: torch.multiprocessing.set_sharing_strategy('file_system')
|
| 244 |
-
|
| 245 |
-
# create the reservoir sampler if necessary
|
| 246 |
-
if self.reservoir_sampler is not None:
|
| 247 |
-
# If no sampler, we can make it faster
|
| 248 |
-
self.reservoir_sampler.initializer(self.reservoir_copy)
|
| 249 |
-
self.iter_reservoir = iter(self.reservoir_sampler)
|
| 250 |
-
else:
|
| 251 |
-
self.iter_reservoir = self.reservoir_copy.__iter__()
|
| 252 |
-
|
| 253 |
-
def __next__(self):
|
| 254 |
-
if self.base_sequence.maximum_number_of_samples_per_epoch is not None and \
|
| 255 |
-
self.number_samples_generated >= self.base_sequence.maximum_number_of_samples_per_epoch:
|
| 256 |
-
# we have reached the maximum number of samples, stop the sequence
|
| 257 |
-
raise StopIteration()
|
| 258 |
-
|
| 259 |
-
if self.reservoir_sampler is not None:
|
| 260 |
-
# items are actually a list of indices!
|
| 261 |
-
indices = self.iter_reservoir.__next__()
|
| 262 |
-
if isinstance(indices, collections.Iterable):
|
| 263 |
-
items = [self.reservoir_copy[index] for index in indices]
|
| 264 |
-
self.number_samples_generated += len(items)
|
| 265 |
-
else:
|
| 266 |
-
self.number_samples_generated += 1
|
| 267 |
-
items = [self.reservoir_copy[indices]]
|
| 268 |
-
else:
|
| 269 |
-
self.number_samples_generated += 1
|
| 270 |
-
items = [self.iter_reservoir.__next__()]
|
| 271 |
-
|
| 272 |
-
if self.base_sequence.collate_fn is not None:
|
| 273 |
-
return self.base_sequence.collate_fn(items)
|
| 274 |
-
|
| 275 |
-
return items
|
| 276 |
-
|
| 277 |
-
def close(self):
|
| 278 |
-
self.base_sequence.close()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|