Add DocumentQuestionAnswering pipeline (#18414)
* [WIP] Skeleton of VisualQuestionAnweringPipeline extended to support LayoutLM-like models * Fixup * Use the full encoding * Basic refactoring to DocumentQuestionAnsweringPipeline * Cleanup * Improve args, docs, and implement preprocessing * Integrate OCR * Refactor question_answering pipeline * Use refactored QA code in the document qa pipeline * Fix tests * Some small cleanups * Use a string type annotation for Image.Image * Update encoding with image features * Wire through the basic docs * Handle invalid response * Handle empty word_boxes properly * Docstring fix * Integrate Donut model * Fixup * Incorporate comments * Address comments * Initial incorporation of tests * Address Comments * Change assert to ValueError * Comments * Wrap `score` in float to make it JSON serializable * Incorporate AutoModeLForDocumentQuestionAnswering changes * Fixup * Rename postprocess function * Fix auto import * Applying comments * Improve docs * Remove extra assets and add copyright * Address comments Co-authored-by: Ankur Goyal <ankur@impira.com>
This commit is contained in:
Ankur Goyal
280
tests/pipelines/test_pipelines_document_question_answering.py
Normal file
280
tests/pipelines/test_pipelines_document_question_answering.py
Normal file
@@ -0,0 +1,280 @@
|
||||
# Copyright 2022 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 unittest
|
||||
|
||||
from transformers import MODEL_FOR_DOCUMENT_QUESTION_ANSWERING_MAPPING, AutoTokenizer, is_vision_available
|
||||
from transformers.pipelines import pipeline
|
||||
from transformers.pipelines.document_question_answering import apply_tesseract
|
||||
from transformers.testing_utils import (
|
||||
is_pipeline_test,
|
||||
nested_simplify,
|
||||
require_detectron2,
|
||||
require_pytesseract,
|
||||
require_tf,
|
||||
require_torch,
|
||||
require_vision,
|
||||
slow,
|
||||
)
|
||||
|
||||
from .test_pipelines_common import ANY, PipelineTestCaseMeta
|
||||
|
||||
|
||||
if is_vision_available():
|
||||
from PIL import Image
|
||||
|
||||
from transformers.image_utils import load_image
|
||||
else:
|
||||
|
||||
class Image:
|
||||
@staticmethod
|
||||
def open(*args, **kwargs):
|
||||
pass
|
||||
|
||||
def load_image(_):
|
||||
return None
|
||||
|
||||
|
||||
# This is a pinned image from a specific revision of a document question answering space, hosted by HuggingFace,
|
||||
# so we can expect it to be available.
|
||||
INVOICE_URL = (
|
||||
"https://huggingface.co/spaces/impira/docquery/resolve/2f6c96314dc84dfda62d40de9da55f2f5165d403/invoice.png"
|
||||
)
|
||||
|
||||
|
||||
@is_pipeline_test
|
||||
@require_torch
|
||||
@require_vision
|
||||
class DocumentQuestionAnsweringPipelineTests(unittest.TestCase, metaclass=PipelineTestCaseMeta):
|
||||
model_mapping = MODEL_FOR_DOCUMENT_QUESTION_ANSWERING_MAPPING
|
||||
|
||||
@require_pytesseract
|
||||
@require_vision
|
||||
def get_test_pipeline(self, model, tokenizer, feature_extractor):
|
||||
dqa_pipeline = pipeline(
|
||||
"document-question-answering", model=model, tokenizer=tokenizer, feature_extractor=feature_extractor
|
||||
)
|
||||
|
||||
image = INVOICE_URL
|
||||
word_boxes = list(zip(*apply_tesseract(load_image(image), None, "")))
|
||||
question = "What is the placebo?"
|
||||
examples = [
|
||||
{
|
||||
"image": load_image(image),
|
||||
"question": question,
|
||||
},
|
||||
{
|
||||
"image": image,
|
||||
"question": question,
|
||||
},
|
||||
{
|
||||
"image": image,
|
||||
"question": question,
|
||||
"word_boxes": word_boxes,
|
||||
},
|
||||
{
|
||||
"image": None,
|
||||
"question": question,
|
||||
"word_boxes": word_boxes,
|
||||
},
|
||||
]
|
||||
return dqa_pipeline, examples
|
||||
|
||||
def run_pipeline_test(self, dqa_pipeline, examples):
|
||||
outputs = dqa_pipeline(examples, top_k=2)
|
||||
self.assertEqual(
|
||||
outputs,
|
||||
[
|
||||
[
|
||||
{"score": ANY(float), "answer": ANY(str), "start": ANY(int), "end": ANY(int)},
|
||||
{"score": ANY(float), "answer": ANY(str), "start": ANY(int), "end": ANY(int)},
|
||||
]
|
||||
]
|
||||
* 4,
|
||||
)
|
||||
|
||||
@require_torch
|
||||
@require_detectron2
|
||||
@require_pytesseract
|
||||
def test_small_model_pt(self):
|
||||
dqa_pipeline = pipeline("document-question-answering", model="hf-internal-testing/tiny-random-layoutlmv2")
|
||||
image = INVOICE_URL
|
||||
question = "How many cats are there?"
|
||||
|
||||
expected_output = [
|
||||
{
|
||||
"score": 0.0001,
|
||||
"answer": "2312/2019 DUE DATE 26102/2019 ay DESCRIPTION UNIT PRICE",
|
||||
"start": 38,
|
||||
"end": 45,
|
||||
},
|
||||
{"score": 0.0001, "answer": "2312/2019 DUE", "start": 38, "end": 39},
|
||||
]
|
||||
outputs = dqa_pipeline(image=image, question=question, top_k=2)
|
||||
self.assertEqual(nested_simplify(outputs, decimals=4), expected_output)
|
||||
|
||||
outputs = dqa_pipeline({"image": image, "question": question}, top_k=2)
|
||||
self.assertEqual(nested_simplify(outputs, decimals=4), expected_output)
|
||||
|
||||
# This image does not detect ANY text in it, meaning layoutlmv2 should fail.
|
||||
# Empty answer probably
|
||||
image = "./tests/fixtures/tests_samples/COCO/000000039769.png"
|
||||
outputs = dqa_pipeline(image=image, question=question, top_k=2)
|
||||
self.assertEqual(outputs, [])
|
||||
|
||||
# We can optionnally pass directly the words and bounding boxes
|
||||
image = "./tests/fixtures/tests_samples/COCO/000000039769.png"
|
||||
words = []
|
||||
boxes = []
|
||||
outputs = dqa_pipeline(image=image, question=question, words=words, boxes=boxes, top_k=2)
|
||||
self.assertEqual(outputs, [])
|
||||
|
||||
# TODO: Enable this once hf-internal-testing/tiny-random-donut is implemented
|
||||
# @require_torch
|
||||
# def test_small_model_pt_donut(self):
|
||||
# dqa_pipeline = pipeline("document-question-answering", model="hf-internal-testing/tiny-random-donut")
|
||||
# # dqa_pipeline = pipeline("document-question-answering", model="../tiny-random-donut")
|
||||
# image = "https://templates.invoicehome.com/invoice-template-us-neat-750px.png"
|
||||
# question = "How many cats are there?"
|
||||
#
|
||||
# outputs = dqa_pipeline(image=image, question=question, top_k=2)
|
||||
# self.assertEqual(
|
||||
# nested_simplify(outputs, decimals=4), [{"score": 0.8799, "answer": "2"}, {"score": 0.296, "answer": "1"}]
|
||||
# )
|
||||
|
||||
@slow
|
||||
@require_torch
|
||||
@require_detectron2
|
||||
@require_pytesseract
|
||||
def test_large_model_pt(self):
|
||||
dqa_pipeline = pipeline(
|
||||
"document-question-answering",
|
||||
model="tiennvcs/layoutlmv2-base-uncased-finetuned-docvqa",
|
||||
revision="9977165",
|
||||
)
|
||||
image = INVOICE_URL
|
||||
question = "What is the invoice number?"
|
||||
|
||||
outputs = dqa_pipeline(image=image, question=question, top_k=2)
|
||||
self.assertEqual(
|
||||
nested_simplify(outputs, decimals=4),
|
||||
[
|
||||
{"score": 0.9966, "answer": "us-001", "start": 15, "end": 15},
|
||||
{"score": 0.0009, "answer": "us-001", "start": 15, "end": 15},
|
||||
],
|
||||
)
|
||||
|
||||
outputs = dqa_pipeline({"image": image, "question": question}, top_k=2)
|
||||
self.assertEqual(
|
||||
nested_simplify(outputs, decimals=4),
|
||||
[
|
||||
{"score": 0.9966, "answer": "us-001", "start": 15, "end": 15},
|
||||
{"score": 0.0009, "answer": "us-001", "start": 15, "end": 15},
|
||||
],
|
||||
)
|
||||
|
||||
outputs = dqa_pipeline(
|
||||
[{"image": image, "question": question}, {"image": image, "question": question}], top_k=2
|
||||
)
|
||||
self.assertEqual(
|
||||
nested_simplify(outputs, decimals=4),
|
||||
[
|
||||
[
|
||||
{"score": 0.9966, "answer": "us-001", "start": 15, "end": 15},
|
||||
{"score": 0.0009, "answer": "us-001", "start": 15, "end": 15},
|
||||
],
|
||||
]
|
||||
* 2,
|
||||
)
|
||||
|
||||
@slow
|
||||
@require_torch
|
||||
@require_pytesseract
|
||||
@require_vision
|
||||
def test_large_model_pt_layoutlm(self):
|
||||
tokenizer = AutoTokenizer.from_pretrained(
|
||||
"impira/layoutlm-document-qa", revision="3dc6de3", add_prefix_space=True
|
||||
)
|
||||
dqa_pipeline = pipeline(
|
||||
"document-question-answering",
|
||||
model="impira/layoutlm-document-qa",
|
||||
tokenizer=tokenizer,
|
||||
revision="3dc6de3",
|
||||
)
|
||||
image = INVOICE_URL
|
||||
question = "What is the invoice number?"
|
||||
|
||||
outputs = dqa_pipeline(image=image, question=question, top_k=2)
|
||||
self.assertEqual(
|
||||
nested_simplify(outputs, decimals=4),
|
||||
[
|
||||
{"score": 0.9998, "answer": "us-001", "start": 15, "end": 15},
|
||||
{"score": 0.0, "answer": "INVOICE # us-001", "start": 13, "end": 15},
|
||||
],
|
||||
)
|
||||
|
||||
outputs = dqa_pipeline({"image": image, "question": question}, top_k=2)
|
||||
self.assertEqual(
|
||||
nested_simplify(outputs, decimals=4),
|
||||
[
|
||||
{"score": 0.9998, "answer": "us-001", "start": 15, "end": 15},
|
||||
{"score": 0.0, "answer": "INVOICE # us-001", "start": 13, "end": 15},
|
||||
],
|
||||
)
|
||||
|
||||
outputs = dqa_pipeline(
|
||||
[{"image": image, "question": question}, {"image": image, "question": question}], top_k=2
|
||||
)
|
||||
self.assertEqual(
|
||||
nested_simplify(outputs, decimals=4),
|
||||
[
|
||||
[
|
||||
{"score": 0.9998, "answer": "us-001", "start": 15, "end": 15},
|
||||
{"score": 0.0, "answer": "INVOICE # us-001", "start": 13, "end": 15},
|
||||
]
|
||||
]
|
||||
* 2,
|
||||
)
|
||||
|
||||
word_boxes = list(zip(*apply_tesseract(load_image(image), None, "")))
|
||||
|
||||
# This model should also work if `image` is set to None
|
||||
outputs = dqa_pipeline({"image": None, "word_boxes": word_boxes, "question": question}, top_k=2)
|
||||
self.assertEqual(
|
||||
nested_simplify(outputs, decimals=4),
|
||||
[
|
||||
{"score": 0.9998, "answer": "us-001", "start": 15, "end": 15},
|
||||
{"score": 0.0, "answer": "INVOICE # us-001", "start": 13, "end": 15},
|
||||
],
|
||||
)
|
||||
|
||||
@slow
|
||||
@require_torch
|
||||
def test_large_model_pt_donut(self):
|
||||
dqa_pipeline = pipeline(
|
||||
"document-question-answering",
|
||||
model="naver-clova-ix/donut-base-finetuned-docvqa",
|
||||
tokenizer=AutoTokenizer.from_pretrained("naver-clova-ix/donut-base-finetuned-docvqa"),
|
||||
feature_extractor="naver-clova-ix/donut-base-finetuned-docvqa",
|
||||
)
|
||||
|
||||
image = INVOICE_URL
|
||||
question = "What is the invoice number?"
|
||||
outputs = dqa_pipeline(image=image, question=question, top_k=2)
|
||||
self.assertEqual(nested_simplify(outputs, decimals=4), {"answer": "us-001"})
|
||||
|
||||
@require_tf
|
||||
@unittest.skip("Document question answering not implemented in TF")
|
||||
def test_small_model_tf(self):
|
||||
pass
|
||||
Reference in New Issue
Block a user