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

Fast image processor (#28847)

Browse Source 
* Draft fast image processors

* Draft working fast version

* py3.8 compatible cache

* Enable loading fast image processors through auto

* Tidy up; rescale behaviour based on input type

* Enable tests for fast image processors

* Smarter rescaling

* Don't default to Fast

* Safer imports

* Add necessary Pillow requirement

* Woops

* Add AutoImageProcessor test

* Fix up

* Fix test for imagegpt

* Fix test

* Review comments

* Add warning for TF and JAX input types

* Rearrange

* Return transforms

* NumpyToTensor transformation

* Rebase - include changes from upstream in ImageProcessingMixin

* Safe typing

* Fix up

* convert mean/std to tesnor to rescale

* Don't store transforms in state

* Fix up

* Update src/transformers/image_processing_utils_fast.py

Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

* Update src/transformers/models/auto/image_processing_auto.py

Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

* Update src/transformers/models/auto/image_processing_auto.py

Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

* Update src/transformers/models/auto/image_processing_auto.py

Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

* Warn if fast image processor available

* Update src/transformers/models/vit/image_processing_vit_fast.py

* Transpose incoming numpy images to be in CHW format

* Update mapping names based on packages, auto set fast to None

* Fix up

* Fix

* Add AutoImageProcessor.from_pretrained(checkpoint, use_fast=True) test

* Update src/transformers/models/vit/image_processing_vit_fast.py

Co-authored-by: Pavel Iakubovskii <qubvel@gmail.com>

* Add equivalence and speed tests

* Fix up

---------

Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>
Co-authored-by: Pavel Iakubovskii <qubvel@gmail.com>
This commit is contained in:
amyeroberts
2024-06-11 15:47:38 +01:00
committed by GitHub
parent edc1dffd00
commit f53fe35b29
64 changed files with 1645 additions and 813 deletions
Download Patch File Download Diff File Expand all files Collapse all files

21
tests/models/auto/test_image_processing_auto.py
View File

@@ -27,8 +27,10 @@ from transformers import (
AutoImageProcessor,
CLIPConfig,
CLIPImageProcessor,
ViTImageProcessor,
ViTImageProcessorFast,
)
from transformers.testing_utils import DUMMY_UNKNOWN_IDENTIFIER
from transformers.testing_utils import DUMMY_UNKNOWN_IDENTIFIER, require_torchvision, require_vision
sys.path.append(str(Path(__file__).parent.parent.parent.parent / "utils"))
@@ -133,6 +135,23 @@ class AutoImageProcessorTest(unittest.TestCase):
):
_ = AutoImageProcessor.from_pretrained("hf-internal-testing/config-no-model")
@require_vision
@require_torchvision
def test_use_fast_selection(self):
checkpoint = "hf-internal-testing/tiny-random-vit"
# Slow image processor is selected by default
image_processor = AutoImageProcessor.from_pretrained(checkpoint)
self.assertIsInstance(image_processor, ViTImageProcessor)
# Fast image processor is selected when use_fast=True
image_processor = AutoImageProcessor.from_pretrained(checkpoint, use_fast=True)
self.assertIsInstance(image_processor, ViTImageProcessorFast)
# Slow image processor is selected when use_fast=False
image_processor = AutoImageProcessor.from_pretrained(checkpoint, use_fast=False)
self.assertIsInstance(image_processor, ViTImageProcessor)
def test_from_pretrained_dynamic_image_processor(self):
# If remote code is not set, we will time out when asking whether to load the model.
with self.assertRaises(ValueError):