PaddleX
PaddleX copied to clipboard
Published
1 week ago
•
PaddlePaddle
PaddlePaddle
使用PP-StructureV3产线出现内存滞留现象
使用Paddlex PP-StructureV3产线解析pdf文档时发现显存占用持续升高且不释放现象，最终会导致显存溢出，增加paddle.device.cuda.empty_cache()主动释放显存未起作用。 通过修改产线yaml文件中的limit_type为max并设置limit_side_len为合理值，可以避免显存升高直到溢出，但是也依然有显存持续升高且不释放的问题，想问下这个是什么原因？如何正确释放显存？
我也遇到了这个问题，5090D显卡，使用Paddlex PP-StructureV3产线解析pdf以及图片，显存占用持续升高，分批次对server发了60个文件测试最终显存占到了如下，但是OCR结束后也无法释放。
|=========================================+========================+======================|
| 0 NVIDIA GeForce RTX 5090 D Off | 00000000:01:00.0 Off | N/A |
| 31% 41C P8 22W / 600W | 28927MiB / 32607MiB | 0% Default |
| | | N/A |
+-----------------------------------------+------------------------+----------------------+
那么在限制图片最大尺寸后，是否有观察到在处理一定数量的文件后，显存不再上升的情况？还是说，在整个处理过程中，显存始终在不断上升。
整个处理过程中，显存始终在不断上升(限制图片最大尺寸后显存上升变慢)
请提供使用的配置和最小可复现脚本，我们这边尝试复现一下。
from fastapi import FastAPI, File, UploadFile, HTTPException
from fastapi.responses import JSONResponse
import tempfile
import os
from paddlex import create_pipeline
# 初始化 FastAPI 应用
app = FastAPI()
max_file_size = 20 * 1024 * 1024
OCR_DEVICE = ''
os.environ['PADDLE_PDX_CACHE_HOME'] = r'./models'
structurev3_yaml_path = r"./PP-StructureV3.yaml"
pipeline_layout = create_pipeline(pipeline=structurev3_yaml_path)
@app.post("/parse-pdf/")
async def parse_pdf(file: UploadFile = File(...)):
try:
if file.size > max_file_size:
raise HTTPException(status_code=400, detail="File size exceeds the limit of 20MB.")
# 创建临时文件保存上传的 PDF
with tempfile.NamedTemporaryFile(delete=False, suffix=".pdf") as temp_pdf:
temp_pdf.write(await file.read())
temp_pdf_path = temp_pdf.name
print(f"temp_pdf_path:{temp_pdf_path}")
markdown_texts = await get_pdf_content(temp_pdf_path)
# 删除临时文件
# os.unlink(temp_pdf_path)
# 返回解析结果
return JSONResponse(content={"status": "success", "data": markdown_texts})
except Exception as e:
# 捕获异常并返回错误信息
return JSONResponse(content={"status": "error", "message": str(e)}, status_code=500)
async def get_pdf_content(input_file: str):
if not os.path.exists(input_file):
raise Exception("file_path is empty...")
try:
output = pipeline_layout.predict(input=input_file)
markdown_list = []
markdown_images = []
for res in output:
md_info = res.markdown
markdown_list.append(md_info)
markdown_images.append(md_info.get("markdown_images", {}))
markdown_texts = pipeline_layout.concatenate_markdown_pages(markdown_list)
return markdown_texts
except Exception as e:
raise Exception(f"get_pdf_content error:{e}")
finally:
release_gpu_memory()
def release_gpu_memory():
"""手动释放GPU显存"""
if 'gpu' in OCR_DEVICE:
import gc
import paddle
# logger.info("gpu release memory")
# cuda.empty_cache()
gc.collect()
paddle.device.cuda.empty_cache()
yaml配置
pipeline_name: PP-StructureV3
batch_size: 8
use_doc_preprocessor: True
use_seal_recognition: True
use_table_recognition: True
use_formula_recognition: True
use_chart_recognition: True
use_region_detection: True
SubModules:
LayoutDetection:
module_name: layout_detection
model_name: PP-DocLayout_plus-L
model_dir: null
batch_size: 8
threshold:
0: 0.3 # paragraph_title
1: 0.5 # image
2: 0.4 # text
3: 0.5 # number
4: 0.5 # abstract
5: 0.5 # content
6: 0.5 # figure_table_chart_title
7: 0.3 # formula
8: 0.5 # table
9: 0.5 # reference
10: 0.5 # doc_title
11: 0.5 # footnote
12: 0.5 # header
13: 0.5 # algorithm
14: 0.5 # footer
15: 0.45 # seal
16: 0.5 # chart
17: 0.5 # formula_number
18: 0.5 # aside_text
19: 0.5 # reference_content
layout_nms: True
layout_unclip_ratio: [1.02, 1.02]
layout_merge_bboxes_mode:
0: "large" # paragraph_title
1: "large" # image
2: "union" # text
3: "union" # number
4: "union" # abstract
5: "union" # content
6: "union" # figure_table_chart_title
7: "large" # formula
8: "union" # table
9: "union" # reference
10: "union" # doc_title
11: "union" # footnote
12: "union" # header
13: "union" # algorithm
14: "union" # footer
15: "union" # seal
16: "large" # chart
17: "union" # formula_number
18: "union" # aside_text
19: "union" # reference_content
ChartRecognition:
module_name: chart_recognition
model_name: PP-Chart2Table
model_dir: null
batch_size: 1
RegionDetection:
module_name: layout_detection
model_name: PP-DocBlockLayout
model_dir: null
layout_nms: True
layout_merge_bboxes_mode: "small"
SubPipelines:
DocPreprocessor:
pipeline_name: doc_preprocessor
batch_size: 8
use_doc_orientation_classify: True
use_doc_unwarping: True
SubModules:
DocOrientationClassify:
module_name: doc_text_orientation
model_name: PP-LCNet_x1_0_doc_ori
model_dir: null
batch_size: 8
DocUnwarping:
module_name: image_unwarping
model_name: UVDoc
model_dir: null
GeneralOCR:
pipeline_name: OCR
batch_size: 8
text_type: general
use_doc_preprocessor: False
use_textline_orientation: True
SubModules:
TextDetection:
module_name: text_detection
model_name: PP-OCRv5_server_det
model_dir: null
limit_side_len: 736
limit_type: max
max_side_limit: 4000
thresh: 0.3
box_thresh: 0.6
unclip_ratio: 1.5
TextLineOrientation:
module_name: textline_orientation
model_name: PP-LCNet_x0_25_textline_ori
model_dir: null
batch_size: 8
TextRecognition:
module_name: text_recognition
model_name: PP-OCRv5_server_rec
model_dir: null
batch_size: 8
score_thresh: 0.0
TableRecognition:
pipeline_name: table_recognition_v2
use_layout_detection: False
use_doc_preprocessor: False
use_ocr_model: False
SubModules:
TableClassification:
module_name: table_classification
model_name: PP-LCNet_x1_0_table_cls
model_dir: null
WiredTableStructureRecognition:
module_name: table_structure_recognition
model_name: SLANeXt_wired
model_dir: null
WirelessTableStructureRecognition:
module_name: table_structure_recognition
model_name: SLANet_plus
model_dir: null
WiredTableCellsDetection:
module_name: table_cells_detection
model_name: RT-DETR-L_wired_table_cell_det
model_dir: null
WirelessTableCellsDetection:
module_name: table_cells_detection
model_name: RT-DETR-L_wireless_table_cell_det
model_dir: null
TableOrientationClassify:
module_name: doc_text_orientation
model_name: PP-LCNet_x1_0_doc_ori
model_dir: null
SubPipelines:
GeneralOCR:
pipeline_name: OCR
text_type: general
use_doc_preprocessor: False
use_textline_orientation: True
SubModules:
TextDetection:
module_name: text_detection
model_name: PP-OCRv5_server_det
model_dir: null
limit_side_len: 736
limit_type: max
max_side_limit: 4000
thresh: 0.3
box_thresh: 0.4
unclip_ratio: 1.5
TextLineOrientation:
module_name: textline_orientation
model_name: PP-LCNet_x0_25_textline_ori
model_dir: null
batch_size: 8
TextRecognition:
module_name: text_recognition
model_name: PP-OCRv5_server_rec
model_dir: null
batch_size: 8
score_thresh: 0.0
SealRecognition:
pipeline_name: seal_recognition
batch_size: 8
use_layout_detection: False
use_doc_preprocessor: False
SubPipelines:
SealOCR:
pipeline_name: OCR
batch_size: 8
text_type: seal
use_doc_preprocessor: False
use_textline_orientation: False
SubModules:
TextDetection:
module_name: seal_text_detection
model_name: PP-OCRv4_server_seal_det
model_dir: null
limit_side_len: 736
limit_type: max
max_side_limit: 4000
thresh: 0.2
box_thresh: 0.6
unclip_ratio: 0.5
TextRecognition:
module_name: text_recognition
model_name: PP-OCRv5_server_rec
model_dir: null
batch_size: 8
score_thresh: 0
FormulaRecognition:
pipeline_name: formula_recognition
batch_size: 8
use_layout_detection: False
use_doc_preprocessor: False
SubModules:
FormulaRecognition:
module_name: formula_recognition
model_name: PP-FormulaNet_plus-L
model_dir: null
batch_size: 8
