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

fix document qa bf16 pipeline (#35456)

Browse Source 
* fix document qa bf16 pipeline

Signed-off-by: jiqing-feng <jiqing.feng@intel.com>

* add test

Signed-off-by: jiqing-feng <jiqing.feng@intel.com>

* fix test

Signed-off-by: jiqing-feng <jiqing.feng@intel.com>

---------

Signed-off-by: jiqing-feng <jiqing.feng@intel.com>
This commit is contained in:
jiqing-feng
2025-01-21 00:18:07 +08:00
committed by GitHub
parent ec97417827
commit 729b569531
2 changed files with 51 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

47
tests/pipelines/test_pipelines_document_question_answering.py
View File

@@ -14,7 +14,12 @@
import unittest
from transformers import MODEL_FOR_DOCUMENT_QUESTION_ANSWERING_MAPPING, AutoTokenizer, is_vision_available
from transformers import (
MODEL_FOR_DOCUMENT_QUESTION_ANSWERING_MAPPING,
AutoTokenizer,
is_torch_available,
is_vision_available,
)
from transformers.pipelines import DocumentQuestionAnsweringPipeline, pipeline
from transformers.pipelines.document_question_answering import apply_tesseract
from transformers.testing_utils import (
@@ -24,6 +29,7 @@ from transformers.testing_utils import (
require_pytesseract,
require_tf,
require_torch,
require_torch_bf16,
require_vision,
slow,
)
@@ -31,6 +37,9 @@ from transformers.testing_utils import (
from .test_pipelines_common import ANY
if is_torch_available():
import torch
if is_vision_available():
from PIL import Image
@@ -145,6 +154,42 @@ class DocumentQuestionAnsweringPipelineTests(unittest.TestCase):
outputs = dqa_pipeline(image=image, question=question, words=words, boxes=boxes, top_k=2)
self.assertEqual(outputs, [])
@require_torch
@require_torch_bf16
@require_detectron2
@require_pytesseract
def test_small_model_pt_bf16(self):
dqa_pipeline = pipeline(
"document-question-answering",
model="hf-internal-testing/tiny-random-layoutlmv2-for-dqa-test",
torch_dtype=torch.bfloat16,
)
image = INVOICE_URL
question = "How many cats are there?"
expected_output = [
{"score": 0.0001, "answer": "oy 2312/2019", "start": 38, "end": 39},
{"score": 0.0001, "answer": "oy 2312/2019 DUE", "start": 38, "end": 40},
]
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):