TFDS와 결정론

TensorFlow.org에서 보기 Google Colab에서 실행 GitHub에서 보기 노트북 다운로드

이 문서에서는 다음을 설명합니다.

  • TFDS는 결정론을 보장합니다.
  • TFDS는 어떤 순서로 예제를 읽습니까?
  • 다양한 주의 사항 및 문제

설정

데이터세트

TFDS가 데이터를 읽는 방법을 이해하려면 몇 가지 컨텍스트가 필요합니다.

세대 동안, TFDS 표준화에 원본 데이터 쓰기 .tfrecord 파일을. 큰 데이터 세트의 경우, 여러 .tfrecord 파일이 생성되며, 각각의 여러 예제를 포함. 우리는 서로 전화 .tfrecord 파편을 제기.

이 가이드에서는 1024개의 샤드가 있는 imagenet을 사용합니다.

import re
import tensorflow_datasets as tfds

imagenet = tfds.builder('imagenet2012')

num_shards = imagenet.info.splits['train'].num_shards
num_examples = imagenet.info.splits['train'].num_examples
print(f'imagenet has {num_shards} shards ({num_examples} examples)')
imagenet has 1024 shards (1281167 examples)

데이터세트 예시 ID 찾기

결정론에 대해서만 알고 싶다면 다음 섹션으로 건너뛸 수 있습니다.

각 데이터 세트는 예를 고유로 식별 id (예를 들어 'imagenet2012-train.tfrecord-01023-of-01024__32' ). 이 복구 할 수있는 id 전달하여 read_config.add_tfds_id = True 추가 할 것 'tfds_id' 로부터의 DICT에 키를 tf.data.Dataset .

이 튜토리얼에서는 데이터 세트의 예제 ID를 인쇄할 작은 유틸리티를 정의합니다(사람이 더 읽기 쉽게 정수로 변환).

읽을 때의 결정론

이 섹션의 deterministim 보증을 설명 tfds.load .

shuffle_files=False (기본)

기본 TFDS는 결정 론적으로 예를 산출하여 ( shuffle_files=False )

# Same as: imagenet.as_dataset(split='train').take(20)
print_ex_ids(imagenet, split='train', take=20)
print_ex_ids(imagenet, split='train', take=20)
[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 1251, 1252, 1253, 1254]
[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 1251, 1252, 1253, 1254]

성능을 위해, TFDS 사용하여 동시에 여러 파편을 읽을 tf.data.Dataset.interleave을 . 우리 TFDS 16 예를 읽은 후 샤드 (2)에 전환하는 것이 본 실시 예에서 볼 ( ..., 14, 15, 1251, 1252, ... ). 인터리브에 대한 자세한 내용은 아래를 참조하세요.

마찬가지로 subsplit API도 결정적입니다.

print_ex_ids(imagenet, split='train[67%:84%]', take=20)
print_ex_ids(imagenet, split='train[67%:84%]', take=20)
[858382, 858383, 858384, 858385, 858386, 858387, 858388, 858389, 858390, 858391, 858392, 858393, 858394, 858395, 858396, 858397, 859533, 859534, 859535, 859536]
[858382, 858383, 858384, 858385, 858386, 858387, 858388, 858389, 858390, 858391, 858392, 858393, 858394, 858395, 858396, 858397, 859533, 859534, 859535, 859536]

있는 거 교육 이상의 시대에 당신이 경우 (임의성이 제한되어 있으므로 모든 시대는 같은 순서로 파편을 읽으로, 위의 설정은하지 않는 것이 좋습니다 ds = ds.shuffle(buffer) 버퍼 사이즈).

shuffle_files=True

함께 shuffle_files=True 읽기가 더 이상 결정되지 않도록, 파편, 각 시대에 대한 섞습니다.

print_ex_ids(imagenet, split='train', shuffle_files=True, take=20)
print_ex_ids(imagenet, split='train', shuffle_files=True, take=20)
[568017, 329050, 329051, 329052, 329053, 329054, 329056, 329055, 568019, 568020, 568021, 568022, 568023, 568018, 568025, 568024, 568026, 568028, 568030, 568031]
[43790, 43791, 43792, 43793, 43796, 43794, 43797, 43798, 43795, 43799, 43800, 43801, 43802, 43803, 43804, 43805, 43806, 43807, 43809, 43810]

결정적 파일 셔플링을 얻으려면 아래 레시피를 참조하십시오.

결정론 주의 사항: 인터리브 인수

변경 read_config.interleave_cycle_length , read_config.interleave_block_length 예제의 순서를 변경합니다.

TFDS은에 의존 tf.data.Dataset.interleave 성능을 향상시키고 메모리 사용량을 줄이고, 한 번에 몇 파편을로드 할 수 있습니다.

예제 순서는 인터리브 인수의 고정 값에 대해서만 동일하도록 보장됩니다. 참조 인터리브 문서 를 이해하기 cycle_lengthblock_length 대응도합니다.

  • cycle_length=16 , block_length=16 (기본적으로 상기와 동일) :
print_ex_ids(imagenet, split='train', take=20)
[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 1251, 1252, 1253, 1254]
  • cycle_length=3 , block_length=2 :
read_config = tfds.ReadConfig(
    interleave_cycle_length=3,
    interleave_block_length=2,
)
print_ex_ids(imagenet, split='train', read_config=read_config, take=20)
[0, 1, 1251, 1252, 2502, 2503, 2, 3, 1253, 1254, 2504, 2505, 4, 5, 1255, 1256, 2506, 2507, 6, 7]

두 번째 예에서, 우리는 데이터 세트 2 (읽기 볼 block_length=2 파편에) 예를, 그 다음 파편으로 전환. 매 2 * 3 ( cycle_length=3 )의 예는, 제 샤드 (로 되돌아 간다 shard0-ex0, shard0-ex1, shard1-ex0, shard1-ex1, shard2-ex0, shard2-ex1, shard0-ex2, shard0-ex3, shard1-ex2, shard1-ex3, shard2-ex2,... ).

서브스플릿 및 예시 주문

각 예는 보유 ID 0, 1, ..., num_examples-1 . subsplit의 API는 (예를 들어 실시 예 슬라이스 선택 train[:x] 선택 0, 1, ..., x-1 ).

그러나 subsplit 내에서 예제는 id 순서가 오름차순으로 읽히지 않습니다(shard 및 interleave로 인해).

구체적으로는, ds.take(x)split='train[:x]' 동등하지!

이것은 예제가 다른 샤드에서 오는 위의 인터리브 예제에서 쉽게 볼 수 있습니다.

print_ex_ids(imagenet, split='train', take=25)  # tfds.load(..., split='train').take(25)
print_ex_ids(imagenet, split='train[:25]', take=-1)  # tfds.load(..., split='train[:25]')
[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 1251, 1252, 1253, 1254, 1255, 1256, 1257, 1258, 1259]
[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24]

16 (block_length의) 예 후 .take(25) 동안 다음 샤드 스위치 train[:25] 제 샤드 예에서 계속 읽기.

조리법

결정적 파일 셔플링 가져오기

결정적 셔플링에는 2가지 방법이 있습니다.

  1. 설정 shuffle_seed . 참고: 이를 위해서는 각 에포크에서 시드를 변경해야 합니다. 그렇지 않으면 에포크 간에 동일한 순서로 샤드를 읽습니다.
read_config = tfds.ReadConfig(
    shuffle_seed=32,
)

# Deterministic order, different from the default shuffle_files=False above
print_ex_ids(imagenet, split='train', shuffle_files=True, read_config=read_config, take=22)
print_ex_ids(imagenet, split='train', shuffle_files=True, read_config=read_config, take=22)
[176411, 176412, 176413, 176414, 176415, 176416, 176417, 176418, 176419, 176420, 176421, 176422, 176423, 176424, 176425, 176426, 710647, 710648, 710649, 710650, 710651, 710652]
[176411, 176412, 176413, 176414, 176415, 176416, 176417, 176418, 176419, 176420, 176421, 176422, 176423, 176424, 176425, 176426, 710647, 710648, 710649, 710650, 710651, 710652]
  1. 사용 experimental_interleave_sort_fn :이 파편 읽고 어떤 순서로, 오히려에 의존하는 것보다하는 동안 완벽하게 제어 할 수 있습니다 ds.shuffle 순서를.
def _reverse_order(file_instructions):
  return list(reversed(file_instructions))

read_config = tfds.ReadConfig(
    experimental_interleave_sort_fn=_reverse_order,
)

# Last shard (01023-of-01024) is read first
print_ex_ids(imagenet, split='train', read_config=read_config, take=5)
[1279916, 1279917, 1279918, 1279919, 1279920]

결정적 선점형 파이프라인 가져오기

이것은 더 복잡합니다. 쉽고 만족스러운 솔루션은 없습니다.

  1. 않고 ds.shuffle 결정적으로 셔플 이론적는 (의 함수로서 각 실시 예에 샤드에서 판독 된 판독를 추론 된 예를 계산하는 것이 가능해야 cycle_length , block_length 및 샤드 순서). 그 후 skip , take 각 샤드 용 관통 삽입 될 수 experimental_interleave_sort_fn .

  2. ds.shuffle 는 전체 교육 파이프 라인을 재생하지 않고 가능성이 불가능합니다. 그것은 구원의 필요 ds.shuffle 예 읽은 추론 버퍼 상태. 예를 들면 (예를 들어, 비 연속 일 수 shard5_ex2 , shard5_ex4 읽을 수 있지만하지 shard5_ex3 ).

  3. ds.shuffle , 한 가지 방법은 / example_ids이 (추론 읽기 모든 shards_ids 저장하는 것입니다 tfds_id 그에서 파일 명령을 추론).

가장 간단한 경우 1. 에 가지고있다 .skip(x).take(y) 검색 train[x:x+y] 매치. 다음이 필요합니다.

  • 집합 cycle_length=1 (파편 순차적으로 판독되도록)
  • 설정 shuffle_files=False
  • 사용하지 마십시오 ds.shuffle

훈련이 1 epoch에 불과한 거대한 데이터 세트에서만 사용해야 합니다. 예제는 기본 셔플 순서로 읽습니다.

read_config = tfds.ReadConfig(
    interleave_cycle_length=1,  # Read shards sequentially
)

print_ex_ids(imagenet, split='train', read_config=read_config, skip=40, take=22)
# If the job get pre-empted, using the subsplit API will skip at most `len(shard0)`
print_ex_ids(imagenet, split='train[40:]', read_config=read_config, take=22)
[40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61]
[40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61]

주어진 하위 분할에 대해 읽은 샤드/예제 찾기

으로 tfds.core.DatasetInfo , 당신은 읽기 지침에 직접 액세스 할 수 있습니다.

imagenet.info.splits['train[44%:45%]'].file_instructions
[FileInstruction(filename='imagenet2012-train.tfrecord-00450-of-01024', skip=700, take=-1, num_examples=551),
 FileInstruction(filename='imagenet2012-train.tfrecord-00451-of-01024', skip=0, take=-1, num_examples=1251),
 FileInstruction(filename='imagenet2012-train.tfrecord-00452-of-01024', skip=0, take=-1, num_examples=1251),
 FileInstruction(filename='imagenet2012-train.tfrecord-00453-of-01024', skip=0, take=-1, num_examples=1251),
 FileInstruction(filename='imagenet2012-train.tfrecord-00454-of-01024', skip=0, take=-1, num_examples=1252),
 FileInstruction(filename='imagenet2012-train.tfrecord-00455-of-01024', skip=0, take=-1, num_examples=1251),
 FileInstruction(filename='imagenet2012-train.tfrecord-00456-of-01024', skip=0, take=-1, num_examples=1251),
 FileInstruction(filename='imagenet2012-train.tfrecord-00457-of-01024', skip=0, take=-1, num_examples=1251),
 FileInstruction(filename='imagenet2012-train.tfrecord-00458-of-01024', skip=0, take=-1, num_examples=1251),
 FileInstruction(filename='imagenet2012-train.tfrecord-00459-of-01024', skip=0, take=-1, num_examples=1251),
 FileInstruction(filename='imagenet2012-train.tfrecord-00460-of-01024', skip=0, take=1001, num_examples=1001)]