SUMIN/HuggingFace_transformer
1
0
Fork 0
Code Issues Pull Requests Actions 7 Packages Projects Releases Wiki Activity

Adding batch_size support for (almost) all pipelines (#13724)

Browse Source 
* Tentative enabling of `batch_size` for pipelines.

* Add systematic test for pipeline batching.

* Enabling batch_size on almost all pipelines

- Not `zero-shot` (it's already passing stuff as batched so trickier)
- Not `QA` (preprocess uses squad features, we need to switch to real
tensors at this boundary.

* Adding `min_length_for_response` for conversational.

* Making CTC, speech mappings avaiable regardless of framework.

* Attempt at fixing automatic tests (ffmpeg not enabled for fast tests)

* Removing ffmpeg dependency in tests.

* Small fixes.

* Slight cleanup.

* Adding docs

and adressing comments.

* Quality.

* Update docs/source/main_classes/pipelines.rst

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/pipelines/question_answering.py

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/pipelines/zero_shot_classification.py

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Improving docs.

* Update docs/source/main_classes/pipelines.rst

Co-authored-by: Philipp Schmid <32632186+philschmid@users.noreply.github.com>

* N -> oberved_batch_size

softmax trick.

* Follow `padding_side`.

* Supporting image pipeline batching (and padding).

* Rename `unbatch` -> `loader_batch`.

* unbatch_size forgot.

* Custom padding for offset mappings.

* Attempt to remove librosa.

* Adding require_audio.

* torchaudio.

* Back to using datasets librosa.

* Adding help to set a pad_token on the tokenizer.

* Update src/transformers/pipelines/base.py

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/pipelines/base.py

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/pipelines/base.py

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Quality.

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
Co-authored-by: Philipp Schmid <32632186+philschmid@users.noreply.github.com>
This commit is contained in:
Nicolas Patry
2021-10-29 11:34:18 +02:00
committed by GitHub
parent 4469010c1b
commit be236361f1
27 changed files with 629 additions and 64 deletions
Download Patch File Download Diff File Expand all files Collapse all files

118
tests/test_pipelines_common.py
View File

@@ -12,8 +12,10 @@
# See the License for the specific language governing permissions and
# limitations under the License.
import copy
import importlib
import logging
import random
import string
import unittest
from abc import abstractmethod
@@ -21,6 +23,7 @@ from functools import lru_cache
from unittest import skipIf
from transformers import FEATURE_EXTRACTOR_MAPPING, TOKENIZER_MAPPING, AutoFeatureExtractor, AutoTokenizer, pipeline
from transformers.pipelines.base import _pad
from transformers.testing_utils import is_pipeline_test, require_torch
@@ -73,6 +76,12 @@ def get_tiny_config_from_class(configuration_class):
@lru_cache(maxsize=100)
def get_tiny_tokenizer_from_checkpoint(checkpoint):
tokenizer = AutoTokenizer.from_pretrained(checkpoint)
if tokenizer.vocab_size < 300:
# Wav2Vec2ForCTC for instance
# ByT5Tokenizer
# all are already small enough and have no Fast version that can
# be retrained
return tokenizer
logger.info("Training new from iterator ...")
vocabulary = string.ascii_letters + string.digits + " "
tokenizer = tokenizer.train_new_from_iterator(vocabulary, vocab_size=len(vocabulary), show_progress=False)
@@ -87,6 +96,12 @@ def get_tiny_feature_extractor_from_checkpoint(checkpoint, tiny_config):
feature_extractor = None
if hasattr(tiny_config, "image_size") and feature_extractor:
feature_extractor = feature_extractor.__class__(size=tiny_config.image_size, crop_size=tiny_config.image_size)
# Speech2TextModel specific.
if hasattr(tiny_config, "input_feat_per_channel") and feature_extractor:
feature_extractor = feature_extractor.__class__(
feature_size=tiny_config.input_feat_per_channel, num_mel_bins=tiny_config.input_feat_per_channel
)
return feature_extractor
@@ -136,7 +151,26 @@ class PipelineTestCaseMeta(type):
else:
tokenizer = None
feature_extractor = get_tiny_feature_extractor_from_checkpoint(checkpoint, tiny_config)
self.run_pipeline_test(model, tokenizer, feature_extractor)
pipeline, examples = self.get_test_pipeline(model, tokenizer, feature_extractor)
if pipeline is None:
# The test can disable itself, but it should be very marginal
# Concerns: Wav2Vec2ForCTC without tokenizer test (FastTokenizer don't exist)
return
self.run_pipeline_test(pipeline, examples)
def run_batch_test(pipeline, examples):
# Need to copy because `Conversation` are stateful
if pipeline.tokenizer is not None and pipeline.tokenizer.pad_token_id is None:
return # No batching for this and it's OK
# 10 examples with batch size 4 means there needs to be a unfinished batch
# which is important for the unbatcher
dataset = [copy.deepcopy(random.choice(examples)) for i in range(10)]
for item in pipeline(dataset, batch_size=4):
pass
run_batch_test(pipeline, examples)
return test
@@ -211,3 +245,85 @@ class CommonPipelineTest(unittest.TestCase):
dataset = MyDataset()
for output in text_classifier(dataset):
self.assertEqual(output, {"label": ANY(str), "score": ANY(float)})
@is_pipeline_test
class PipelinePadTest(unittest.TestCase):
@require_torch
def test_pipeline_padding(self):
import torch
items = [
{
"label": "label1",
"input_ids": torch.LongTensor([[1, 23, 24, 2]]),
"attention_mask": torch.LongTensor([[0, 1, 1, 0]]),
},
{
"label": "label2",
"input_ids": torch.LongTensor([[1, 23, 24, 43, 44, 2]]),
"attention_mask": torch.LongTensor([[0, 1, 1, 1, 1, 0]]),
},
]
self.assertEqual(_pad(items, "label", 0, "right"), ["label1", "label2"])
self.assertTrue(
torch.allclose(
_pad(items, "input_ids", 10, "right"),
torch.LongTensor([[1, 23, 24, 2, 10, 10], [1, 23, 24, 43, 44, 2]]),
)
)
self.assertTrue(
torch.allclose(
_pad(items, "input_ids", 10, "left"),
torch.LongTensor([[10, 10, 1, 23, 24, 2], [1, 23, 24, 43, 44, 2]]),
)
)
self.assertTrue(
torch.allclose(
_pad(items, "attention_mask", 0, "right"), torch.LongTensor([[0, 1, 1, 0, 0, 0], [0, 1, 1, 1, 1, 0]])
)
)
@require_torch
def test_pipeline_image_padding(self):
import torch
items = [
{
"label": "label1",
"pixel_values": torch.zeros((1, 3, 10, 10)),
},
{
"label": "label2",
"pixel_values": torch.zeros((1, 3, 10, 10)),
},
]
self.assertEqual(_pad(items, "label", 0, "right"), ["label1", "label2"])
self.assertTrue(
torch.allclose(
_pad(items, "pixel_values", 10, "right"),
torch.zeros((2, 3, 10, 10)),
)
)
@require_torch
def test_pipeline_offset_mapping(self):
import torch
items = [
{
"offset_mappings": torch.zeros([1, 11, 2], dtype=torch.long),
},
{
"offset_mappings": torch.zeros([1, 4, 2], dtype=torch.long),
},
]
self.assertTrue(
torch.allclose(
_pad(items, "offset_mappings", 0, "right"),
torch.zeros((2, 11, 2), dtype=torch.long),
),
)