EngineX-Cambricon/enginex-mlu370-any2any
10
0
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity

init

Browse Source
This commit is contained in:
luopingyi
2025-10-09 16:47:16 +08:00
parent c8feb4deb5
commit e27e3f16bb
5248 changed files with 1778505 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

0
transformers/tests/models/nougat/__init__.py Normal file
View File

323
transformers/tests/models/nougat/test_image_processing_nougat.py Normal file
View File

@@ -0,0 +1,323 @@
# Copyright 2023 HuggingFace Inc.
#
# 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 unittest
from functools import cached_property
import numpy as np
from huggingface_hub import hf_hub_download
from transformers.image_utils import SizeDict, load_image
from transformers.testing_utils import require_torch, require_vision
from transformers.utils import is_torch_available, is_torchvision_available, is_vision_available
from ...test_image_processing_common import ImageProcessingTestMixin, prepare_image_inputs
from ...test_processing_common import url_to_local_path
if is_torch_available():
import torch
if is_vision_available():
from PIL import Image
from transformers import NougatImageProcessor
if is_torchvision_available():
from transformers import NougatImageProcessorFast
class NougatImageProcessingTester:
def __init__(
self,
parent,
batch_size=7,
num_channels=3,
image_size=18,
min_resolution=30,
max_resolution=400,
do_crop_margin=True,
do_resize=True,
size=None,
do_thumbnail=True,
do_align_long_axis: bool = False,
do_pad=True,
do_normalize: bool = True,
image_mean=[0.5, 0.5, 0.5],
image_std=[0.5, 0.5, 0.5],
):
size = size if size is not None else {"height": 20, "width": 20}
self.parent = parent
self.batch_size = batch_size
self.num_channels = num_channels
self.image_size = image_size
self.min_resolution = min_resolution
self.max_resolution = max_resolution
self.do_crop_margin = do_crop_margin
self.do_resize = do_resize
self.size = size
self.do_thumbnail = do_thumbnail
self.do_align_long_axis = do_align_long_axis
self.do_pad = do_pad
self.do_normalize = do_normalize
self.image_mean = image_mean
self.image_std = image_std
self.data_format = "channels_first"
self.input_data_format = "channels_first"
def prepare_image_processor_dict(self):
return {
"do_crop_margin": self.do_crop_margin,
"do_resize": self.do_resize,
"size": self.size,
"do_thumbnail": self.do_thumbnail,
"do_align_long_axis": self.do_align_long_axis,
"do_pad": self.do_pad,
"do_normalize": self.do_normalize,
"image_mean": self.image_mean,
"image_std": self.image_std,
}
def expected_output_image_shape(self, images):
return self.num_channels, self.size["height"], self.size["width"]
def prepare_dummy_image(self):
revision = "ec57bf8c8b1653a209c13f6e9ee66b12df0fc2db"
filepath = hf_hub_download(
repo_id="hf-internal-testing/fixtures_docvqa",
filename="nougat_pdf.png",
repo_type="dataset",
revision=revision,
)
image = Image.open(filepath).convert("RGB")
return image
def prepare_image_inputs(self, equal_resolution=False, numpify=False, torchify=False):
return prepare_image_inputs(
batch_size=self.batch_size,
num_channels=self.num_channels,
min_resolution=self.min_resolution,
max_resolution=self.max_resolution,
equal_resolution=equal_resolution,
numpify=numpify,
torchify=torchify,
)
@require_torch
@require_vision
class NougatImageProcessingTest(ImageProcessingTestMixin, unittest.TestCase):
image_processing_class = NougatImageProcessor if is_vision_available() else None
fast_image_processing_class = NougatImageProcessorFast if is_torchvision_available() else None
def setUp(self):
super().setUp()
self.image_processor_tester = NougatImageProcessingTester(self)
@property
def image_processor_dict(self):
return self.image_processor_tester.prepare_image_processor_dict()
@cached_property
def image_processor(self):
return self.image_processing_class(**self.image_processor_dict)
@unittest.skip(reason="FIXME: @yoni.")
def test_slow_fast_equivalence_batched(self):
pass
def test_image_processor_properties(self):
for image_processing_class in self.image_processor_list:
image_processing = image_processing_class(**self.image_processor_dict)
self.assertTrue(hasattr(image_processing, "do_resize"))
self.assertTrue(hasattr(image_processing, "size"))
self.assertTrue(hasattr(image_processing, "do_normalize"))
self.assertTrue(hasattr(image_processing, "image_mean"))
self.assertTrue(hasattr(image_processing, "image_std"))
def test_image_processor_from_dict_with_kwargs(self):
for image_processing_class in self.image_processor_list:
image_processor = image_processing_class(**self.image_processor_dict)
self.assertEqual(image_processor.size, {"height": 20, "width": 20})
kwargs = dict(self.image_processor_dict)
kwargs.pop("size", None)
image_processor = self.image_processing_class(**kwargs, size=42)
self.assertEqual(image_processor.size, {"height": 42, "width": 42})
def test_expected_output(self):
dummy_image = self.image_processor_tester.prepare_dummy_image()
for image_processing_class in self.image_processor_list:
image_processor = image_processing_class(**self.image_processor_dict)
inputs = image_processor(dummy_image, return_tensors="pt")
torch.testing.assert_close(inputs["pixel_values"].mean(), torch.tensor(0.4906), rtol=1e-3, atol=1e-3)
def test_crop_margin_all_white(self):
image = np.uint8(np.ones((3, 100, 100)) * 255)
for image_processing_class in self.image_processor_list:
if image_processing_class == NougatImageProcessorFast:
image = torch.from_numpy(image)
image_processor = image_processing_class(**self.image_processor_dict)
cropped_image = image_processor.crop_margin(image)
self.assertTrue(torch.equal(image, cropped_image))
else:
image_processor = image_processing_class(**self.image_processor_dict)
cropped_image = image_processor.crop_margin(image)
self.assertTrue(np.array_equal(image, cropped_image))
def test_crop_margin_centered_black_square(self):
image = np.ones((3, 100, 100), dtype=np.uint8) * 255
image[:, 45:55, 45:55] = 0
expected_cropped = image[:, 45:55, 45:55]
for image_processing_class in self.image_processor_list:
if image_processing_class == NougatImageProcessorFast:
image = torch.from_numpy(image)
expected_cropped = torch.from_numpy(expected_cropped)
image_processor = image_processing_class(**self.image_processor_dict)
cropped_image = image_processor.crop_margin(image)
self.assertTrue(torch.equal(expected_cropped, cropped_image))
else:
image_processor = image_processing_class(**self.image_processor_dict)
cropped_image = image_processor.crop_margin(image)
self.assertTrue(np.array_equal(expected_cropped, cropped_image))
def test_align_long_axis_no_rotation(self):
image = np.uint8(np.ones((3, 100, 200)) * 255)
for image_processing_class in self.image_processor_list:
if image_processing_class == NougatImageProcessorFast:
image = torch.from_numpy(image)
size = SizeDict(height=200, width=300)
image_processor = image_processing_class(**self.image_processor_dict)
aligned_image = image_processor.align_long_axis(image, size)
self.assertEqual(image.shape, aligned_image.shape)
else:
size = {"height": 200, "width": 300}
image_processor = image_processing_class(**self.image_processor_dict)
aligned_image = image_processor.align_long_axis(image, size)
self.assertEqual(image.shape, aligned_image.shape)
def test_align_long_axis_with_rotation(self):
image = np.uint8(np.ones((3, 200, 100)) * 255)
for image_processing_class in self.image_processor_list:
image_processor = image_processing_class(**self.image_processor_dict)
if image_processing_class == NougatImageProcessorFast:
image = torch.from_numpy(image)
size = SizeDict(height=300, width=200)
image_processor = image_processing_class(**self.image_processor_dict)
aligned_image = image_processor.align_long_axis(image, size)
self.assertEqual(torch.Size([3, 200, 100]), aligned_image.shape)
else:
size = {"height": 300, "width": 200}
image_processor = image_processing_class(**self.image_processor_dict)
aligned_image = image_processor.align_long_axis(image, size)
self.assertEqual((3, 200, 100), aligned_image.shape)
def test_align_long_axis_data_format(self):
image = np.uint8(np.ones((3, 100, 200)) * 255)
for image_processing_class in self.image_processor_list:
if image_processing_class == NougatImageProcessorFast:
image = torch.from_numpy(image)
image_processor = image_processing_class(**self.image_processor_dict)
size = SizeDict(height=200, width=300)
aligned_image = image_processor.align_long_axis(image, size)
self.assertEqual(torch.Size([3, 100, 200]), aligned_image.shape)
else:
size = {"height": 200, "width": 300}
data_format = "channels_first"
image_processor = image_processing_class(**self.image_processor_dict)
aligned_image = image_processor.align_long_axis(image, size, data_format)
self.assertEqual((3, 100, 200), aligned_image.shape)
def prepare_dummy_np_image(self):
revision = "ec57bf8c8b1653a209c13f6e9ee66b12df0fc2db"
filepath = hf_hub_download(
repo_id="hf-internal-testing/fixtures_docvqa",
filename="nougat_pdf.png",
repo_type="dataset",
revision=revision,
)
image = Image.open(filepath).convert("RGB")
return np.array(image).transpose(2, 0, 1)
def test_crop_margin_equality_cv2_python(self):
image = self.prepare_dummy_np_image()
for image_processing_class in self.image_processor_list:
if image_processing_class == NougatImageProcessorFast:
image = torch.from_numpy(image)
image_processor = image_processing_class(**self.image_processor_dict)
image_cropped_python = image_processor.crop_margin(image)
self.assertEqual(image_cropped_python.shape, torch.Size([3, 850, 685]))
self.assertAlmostEqual(image_cropped_python.float().mean().item(), 237.43881150708458, delta=0.001)
else:
image_processor = image_processing_class(**self.image_processor_dict)
image_cropped_python = image_processor.crop_margin(image)
self.assertEqual(image_cropped_python.shape, (3, 850, 685))
self.assertAlmostEqual(image_cropped_python.mean(), 237.43881150708458, delta=0.001)
def test_call_numpy_4_channels(self):
for image_processing_class in self.image_processor_list:
if image_processing_class == NougatImageProcessor:
# Test that can process images which have an arbitrary number of channels
# Initialize image_processing
image_processor = image_processing_class(**self.image_processor_dict)
# create random numpy tensors
self.image_processor_tester.num_channels = 4
image_inputs = self.image_processor_tester.prepare_image_inputs(equal_resolution=False, numpify=True)
# Test not batched input
encoded_images = image_processor(
image_inputs[0],
return_tensors="pt",
input_data_format="channels_last",
image_mean=0,
image_std=1,
).pixel_values
expected_output_image_shape = self.image_processor_tester.expected_output_image_shape(
[image_inputs[0]]
)
self.assertEqual(tuple(encoded_images.shape), (1, *expected_output_image_shape))
# Test batched
encoded_images = image_processor(
image_inputs,
return_tensors="pt",
input_data_format="channels_last",
image_mean=0,
image_std=1,
).pixel_values
expected_output_image_shape = self.image_processor_tester.expected_output_image_shape(image_inputs)
self.assertEqual(
tuple(encoded_images.shape), (self.image_processor_tester.batch_size, *expected_output_image_shape)
)
def test_slow_fast_equivalence(self):
if not self.test_slow_image_processor or not self.test_fast_image_processor:
self.skipTest(reason="Skipping slow/fast equivalence test")
if self.image_processing_class is None or self.fast_image_processing_class is None:
self.skipTest(reason="Skipping slow/fast equivalence test as one of the image processors is not defined")
dummy_image = load_image(url_to_local_path("http://images.cocodataset.org/val2017/000000039769.jpg"))
image_processor_slow = self.image_processing_class(**self.image_processor_dict)
image_processor_fast = self.fast_image_processing_class(**self.image_processor_dict)
encoding_slow = image_processor_slow(dummy_image, return_tensors="pt")
encoding_fast = image_processor_fast(dummy_image, return_tensors="pt")
# Adding a larger than usual tolerance because the slow processor uses reducing_gap=2.0 during resizing.
torch.testing.assert_close(encoding_slow.pixel_values, encoding_fast.pixel_values, atol=2e-1, rtol=0)
self.assertLessEqual(
torch.mean(torch.abs(encoding_slow.pixel_values - encoding_fast.pixel_values)).item(), 2e-2
)

202
transformers/tests/models/nougat/test_tokenization_nougat.py Normal file
View File

@@ -0,0 +1,202 @@
# Copyright 2023 The HuggingFace Team. All rights reserved.
#
# 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 unittest
from transformers import NougatTokenizerFast
from transformers.models.nougat.tokenization_nougat_fast import markdown_compatible, normalize_list_like_lines
from transformers.testing_utils import require_levenshtein, require_nltk, require_tokenizers
from ...test_tokenization_common import TokenizerTesterMixin
@require_tokenizers
class NougatTokenizationTest(TokenizerTesterMixin, unittest.TestCase):
from_pretrained_id = "facebook/nougat-base"
slow_tokenizer_class = None
rust_tokenizer_class = NougatTokenizerFast
tokenizer_class = NougatTokenizerFast
test_rust_tokenizer = True
test_slow_tokenizer = False
from_pretrained_vocab_key = "tokenizer_file"
special_tokens_map = {"bos_token": "<s>", "eos_token": "</s>", "unk_token": "<unk>", "pad_token": "<pad>"}
@classmethod
def setUpClass(cls):
super().setUpClass()
tokenizer = NougatTokenizerFast.from_pretrained("facebook/nougat-base")
tokenizer.save_pretrained(cls.tmpdirname)
@classmethod
def get_rust_tokenizer(cls, pretrained_name=None, **kwargs):
_kwargs = copy.deepcopy(cls.special_tokens_map)
_kwargs.update(kwargs)
kwargs = _kwargs
pretrained_name = pretrained_name or cls.tmpdirname
return NougatTokenizerFast.from_pretrained(pretrained_name, **kwargs)
def test_padding(self, max_length=6):
for tokenizer, pretrained_name, kwargs in self.tokenizers_list:
with self.subTest(f"{tokenizer.__class__.__name__} ({pretrained_name})"):
tokenizer_r = self.get_rust_tokenizer(pretrained_name, **kwargs)
# Simple input
sentence1 = "This is a simple input"
sentence2 = ["This is a simple input 1", "This is a simple input 2"]
pair1 = ("This is a simple input", "This is a pair")
pair2 = [
("This is a simple input 1", "This is a simple input 2"),
("This is a simple pair 1", "This is a simple pair 2"),
]
# Simple input tests
try:
tokenizer_r.encode(sentence1, max_length=max_length)
tokenizer_r.encode_plus(sentence1, max_length=max_length)
tokenizer_r.batch_encode_plus(sentence2, max_length=max_length)
tokenizer_r.encode(pair1, max_length=max_length)
tokenizer_r.batch_encode_plus(pair2, max_length=max_length)
except ValueError:
self.fail("Nougat Tokenizer should be able to deal with padding")
tokenizer_r.pad_token = None # Hotfixing padding = None
self.assertRaises(
ValueError, tokenizer_r.encode, sentence1, max_length=max_length, padding="max_length"
)
# Simple input
self.assertRaises(
ValueError, tokenizer_r.encode_plus, sentence1, max_length=max_length, padding="max_length"
)
# Simple input
self.assertRaises(
ValueError,
tokenizer_r.batch_encode_plus,
sentence2,
max_length=max_length,
padding="max_length",
)
# Pair input
self.assertRaises(ValueError, tokenizer_r.encode, pair1, max_length=max_length, padding="max_length")
# Pair input
self.assertRaises(
ValueError, tokenizer_r.encode_plus, pair1, max_length=max_length, padding="max_length"
)
# Pair input
self.assertRaises(
ValueError,
tokenizer_r.batch_encode_plus,
pair2,
max_length=max_length,
padding="max_length",
)
@unittest.skip(reason="NougatTokenizerFast does not have tokenizer_file in its signature")
def test_rust_tokenizer_signature(self):
pass
@unittest.skip(reason="NougatTokenizerFast does not support pretokenized inputs")
def test_pretokenized_inputs(self):
pass
@unittest.skip(reason="NougatTokenizerFast directly inherits from PreTrainedTokenizerFast")
def test_prepare_for_model(self):
pass
@unittest.skip(reason="This needs a slow tokenizer. Nougat does not have one!")
def test_encode_decode_with_spaces(self):
pass
class MarkdownCompatibleTest(unittest.TestCase):
def test_equation_tag(self):
input_text = "(3.2) \\[Equation Text\\]"
excepted_output = "\\[Equation Text \\tag{3.2}\\]"
self.assertEqual(markdown_compatible(input_text), excepted_output)
def test_equation_tag_letters(self):
input_text = "(18a) \\[Equation Text\\]"
excepted_output = "\\[Equation Text \\tag{18a}\\]"
self.assertEqual(markdown_compatible(input_text), excepted_output)
def test_bold_formatting(self):
input_text = r"This is \bm{bold} text."
expected_output = r"This is \mathbf{bold} text."
self.assertEqual(markdown_compatible(input_text), expected_output)
def test_url_conversion(self):
input_text = "Visit my website at https://www.example.com"
expected_output = "Visit my website at [https://www.example.com](https://www.example.com)"
self.assertEqual(markdown_compatible(input_text), expected_output)
def test_algorithm_code_block(self):
input_text = "```python\nprint('Hello, world!')\n```"
expected_output = "```\npython\nprint('Hello, world!')\n```"
self.assertEqual(markdown_compatible(input_text), expected_output)
def test_escape_characters(self):
input_text = r"Escaped characters like \n should not be \\[affected\\]"
expected_output = r"Escaped characters like \n should not be \\[affected\\]"
self.assertEqual(markdown_compatible(input_text), expected_output)
def test_nested_tags(self):
input_text = r"This is a super nested \bm{\bm{\bm{\bm{\bm{bold}}}}} tag."
expected_output = r"This is a super nested \mathbf{\mathbf{\mathbf{\mathbf{\mathbf{bold}}}}} tag."
self.assertEqual(markdown_compatible(input_text), expected_output)
class TestNormalizeListLikeLines(unittest.TestCase):
def test_two_level_lines(self):
input_str = "* Item 1 * Item 2"
expected_output = "* Item 1\n* Item 2\n"
self.assertEqual(normalize_list_like_lines(input_str), expected_output)
def test_three_level_lines(self):
input_str = "- I. Item 1 - II. Item 2 - III. Item 3"
expected_output = "- I. Item 1\n- II. Item 2\n- III. Item 3\n"
self.assertEqual(normalize_list_like_lines(input_str), expected_output)
def test_nested_lines(self):
input_str = "- I. Item 1 - I.1 Sub-item 1 - I.1.1 Sub-sub-item 1 - II. Item 2"
expected_output = "- I. Item 1\n\t- I.1 Sub-item 1\n\t\t- I.1.1 Sub-sub-item 1\n- II. Item 2\n"
self.assertEqual(normalize_list_like_lines(input_str), expected_output)
@require_tokenizers
class NougatPostProcessingTest(unittest.TestCase):
def setUp(self):
super().setUp()
self.tokenizer = NougatTokenizerFast.from_pretrained("facebook/nougat-base")
def test_correct_tables_basic(self):
input_str = "\\begin{table} \\begin{tabular}{l l} & \\ \\end{tabular} \\end{table}"
expected_output = "\\begin{table}\n\\begin{tabular}{l l} & \\ \\end{tabular}\n\\end{table}"
self.assertEqual(self.tokenizer.correct_tables(input_str), expected_output)
def test_correct_tables_high_count(self):
input_str = "\\begin{tabular}" * 20
expected_output = ""
self.assertEqual(self.tokenizer.correct_tables(input_str), expected_output)
@require_levenshtein
@require_nltk
def test_postprocess_as_nougat_no_markdown(self):
input_str = "# Nougat: Neural Optical Understanding for Academic Documents\n\n Lukas Blecher\n\nCorrespondence to: lblecher@meta.com\n\nGuillem Cucurull\n\nThomas Scialom\n\nRobert Stojnic\n\nMeta AI\n\nThe paper reports 8.1M papers but the authors recently updated the numbers on the GitHub page https://github.com/allenai/s2orc\n\n###### Abstract\n\nScientific knowledge is predominantly stored in books and scientific journals, often in the form of PDFs. However, the PDF format leads to a loss of semantic information, particularly for mathematical expressions. We propose Nougat (**N**eural **O**ptical **U**nderstanding for **A**cademic Documents), a Visual Transformer model that performs an _Optical Character Recognition_ (OCR) task for processing scientific documents into a markup language, and demonstrate the effectiveness of our model on a new dataset of scientific documents. The proposed approach offers a promising solution to enhance the accessibility of scientific knowledge in the digital age, by bridging the gap between human-readable documents and machine-readable text. We release the models and code to accelerate future work on scientific text recognition.\n\n## 1 Introduction\n\nThe majority of scientific knowledge is stored in books or published in scientific journals, most commonly in the Portable Document Format (PDF). Next to HTML, PDFs are the second most prominent data format on the internet, making up 2.4% of common crawl [1]. However, the information stored in these files is very difficult to extract into any other formats. This is especially true for highly specialized documents, such as scientific research papers, where the semantic information of mathematical expressions is lost.\n\nExisting Optical Character Recognition (OCR) engines, such as Tesseract OCR [2], excel at detecting and classifying individual characters and words in an image, but fail to understand the relationship between them due to their line-by-line approach. This means that they treat superscripts and subscripts in the same way as the surrounding text, which is a significant drawback for mathematical expressions. In mathematical notations like fractions, exponents, and matrices, relative positions of characters are crucial.\n\nConverting academic research papers into machine-readable text also enables accessibility and searchability of science as a whole. The information of millions of academic papers can not be fully accessed because they are locked behind an unreadable format. Existing corpora, such as the S2ORC dataset [3], capture the text of 12M2 papers using GROBID [4], but are missing meaningful representations of the mathematical equations.\n\nFootnote 2: The paper reports 8.1M papers but the authors recently updated the numbers on the GitHub page https://github.com/allenai/s2orc\n\nTo this end, we introduce Nougat, a transformer based model that can convert images of document pages to formatted markup text.\n\nThe primary contributions in this paper are\n\n* Release of a pre-trained model capable of converting a PDF to a lightweight markup language. We release the code and the model on GitHub3 Footnote 3: https://github.com/facebookresearch/nougat\n* We introduce a pipeline to create dataset for pairing PDFs to source code\n* Our method is only dependent on the image of a page, allowing access to scanned papers and books" # noqa: E231
expected_output = "\n\n# Nougat: Neural Optical Understanding for Academic Documents\n\n Lukas Blecher\n\nCorrespondence to: lblecher@meta.com\n\nGuillem Cucurull\n\nThomas Scialom\n\nRobert Stojnic\n\nMeta AI\n\nThe paper reports 8.1M papers but the authors recently updated the numbers on the GitHub page https://github.com/allenai/s2orc\n\n###### Abstract\n\nScientific knowledge is predominantly stored in books and scientific journals, often in the form of PDFs. However, the PDF format leads to a loss of semantic information, particularly for mathematical expressions. We propose Nougat (**N**eural **O**ptical **U**nderstanding for **A**cademic Documents), a Visual Transformer model that performs an _Optical Character Recognition_ (OCR) task for processing scientific documents into a markup language, and demonstrate the effectiveness of our model on a new dataset of scientific documents. The proposed approach offers a promising solution to enhance the accessibility of scientific knowledge in the digital age, by bridging the gap between human-readable documents and machine-readable text. We release the models and code to accelerate future work on scientific text recognition.\n\n## 1 Introduction\n\nThe majority of scientific knowledge is stored in books or published in scientific journals, most commonly in the Portable Document Format (PDF). Next to HTML, PDFs are the second most prominent data format on the internet, making up 2.4% of common crawl [1]. However, the information stored in these files is very difficult to extract into any other formats. This is especially true for highly specialized documents, such as scientific research papers, where the semantic information of mathematical expressions is lost.\n\nExisting Optical Character Recognition (OCR) engines, such as Tesseract OCR [2], excel at detecting and classifying individual characters and words in an image, but fail to understand the relationship between them due to their line-by-line approach. This means that they treat superscripts and subscripts in the same way as the surrounding text, which is a significant drawback for mathematical expressions. In mathematical notations like fractions, exponents, and matrices, relative positions of characters are crucial.\n\nConverting academic research papers into machine-readable text also enables accessibility and searchability of science as a whole. The information of millions of academic papers can not be fully accessed because they are locked behind an unreadable format. Existing corpora, such as the S2ORC dataset [3], capture the text of 12M2 papers using GROBID [4], but are missing meaningful representations of the mathematical equations.\n\nFootnote 2: The paper reports 8.1M papers but the authors recently updated the numbers on the GitHub page https://github.com/allenai/s2orc\n\nTo this end, we introduce Nougat, a transformer based model that can convert images of document pages to formatted markup text.\n\nThe primary contributions in this paper are\n\n* Release of a pre-trained model capable of converting a PDF to a lightweight markup language. We release the code and the model on GitHub3 Footnote 3: https://github.com/facebookresearch/nougat\n* We introduce a pipeline to create dataset for pairing PDFs to source code\n* Our method is only dependent on the image of a page, allowing access to scanned papers and books" # noqa: E231
self.assertEqual(self.tokenizer.post_process_single(input_str, fix_markdown=False), expected_output)