Upload scrpt27.py with huggingface_hub
Browse files- scrpt27.py +435 -0
scrpt27.py
ADDED
@@ -0,0 +1,435 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import json
|
3 |
+
import shutil
|
4 |
+
import argparse
|
5 |
+
import logging
|
6 |
+
import multiprocessing as mp
|
7 |
+
from concurrent.futures import ProcessPoolExecutor, as_completed
|
8 |
+
import torch
|
9 |
+
import psutil
|
10 |
+
import numpy as np
|
11 |
+
from tqdm import tqdm
|
12 |
+
from magic_pdf.pipe.UNIPipe import UNIPipe
|
13 |
+
from magic_pdf.libs.commons import read_file
|
14 |
+
from magic_pdf.libs.config_reader import get_device
|
15 |
+
from magic_pdf.tools.common import do_parse
|
16 |
+
from magic_pdf.libs.pdf_image_tools import cut_image
|
17 |
+
from magic_pdf.rw.DiskReaderWriter import DiskReaderWriter
|
18 |
+
from magic_pdf.filter.pdf_meta_scan import pdf_meta_scan
|
19 |
+
from magic_pdf.filter.pdf_classify_by_type import classify
|
20 |
+
import fitz # PyMuPDF
|
21 |
+
import time
|
22 |
+
import signal
|
23 |
+
|
24 |
+
# Set up logging
|
25 |
+
logging.basicConfig(level=logging.DEBUG, format='%(asctime)s - %(levelname)s - %(message)s')
|
26 |
+
logger = logging.getLogger(__name__)
|
27 |
+
|
28 |
+
# Minimum batch size
|
29 |
+
MIN_BATCH_SIZE = 1
|
30 |
+
|
31 |
+
def parse_arguments():
|
32 |
+
parser = argparse.ArgumentParser(description="Process multiple PDFs using Magic PDF")
|
33 |
+
parser.add_argument("--input", default="input", help="Input folder containing PDF files")
|
34 |
+
parser.add_argument("--output", default="output", help="Output folder for processed files")
|
35 |
+
parser.add_argument("--config", default="magic-pdf.template.json", help="Path to configuration file")
|
36 |
+
parser.add_argument("--timeout", type=int, default=240, help="Timeout for processing each PDF (in seconds)")
|
37 |
+
parser.add_argument("--max-workers", type=int, default=None, help="Maximum number of worker processes")
|
38 |
+
parser.add_argument("--use-bf16", action="store_true", help="Use bfloat16 precision for model inference")
|
39 |
+
parser.add_argument("--initial-batch-size", type=int, default=1, help="Initial batch size for processing")
|
40 |
+
return parser.parse_args()
|
41 |
+
|
42 |
+
def load_config(config_path):
|
43 |
+
with open(config_path, 'r') as f:
|
44 |
+
return json.load(f)
|
45 |
+
|
46 |
+
def get_available_memory(gpu_id):
|
47 |
+
return torch.cuda.get_device_properties(gpu_id).total_memory - torch.cuda.memory_allocated(gpu_id)
|
48 |
+
|
49 |
+
def extract_images(pdf_path, output_folder):
|
50 |
+
doc = fitz.open(pdf_path)
|
51 |
+
pdf_name = os.path.splitext(os.path.basename(pdf_path))[0]
|
52 |
+
images_folder = os.path.join(output_folder, 'images')
|
53 |
+
os.makedirs(images_folder, exist_ok=True)
|
54 |
+
|
55 |
+
for page_num, page in enumerate(doc):
|
56 |
+
for img_index, img in enumerate(page.get_images(full=True)):
|
57 |
+
xref = img[0]
|
58 |
+
base_image = doc.extract_image(xref)
|
59 |
+
image_bytes = base_image["image"]
|
60 |
+
image_ext = base_image["ext"]
|
61 |
+
image_filename = f'{pdf_name}_{page_num+1:03d}_{img_index+1:03d}.{image_ext}'
|
62 |
+
image_path = os.path.join(images_folder, image_filename)
|
63 |
+
with open(image_path, "wb") as image_file:
|
64 |
+
image_file.write(image_bytes)
|
65 |
+
doc.close()
|
66 |
+
|
67 |
+
class MagicModel:
|
68 |
+
def __init__(self, config):
|
69 |
+
self.config = config
|
70 |
+
|
71 |
+
def process_pdf(self, pdf_data, parse_type, layout_info, log_file_path):
|
72 |
+
processed_pages = []
|
73 |
+
with open(log_file_path, 'a') as log_file:
|
74 |
+
log_file.write(f"Entering process_pdf\n")
|
75 |
+
log_file.write(f" parse_type: {parse_type}, (expected: str)\n")
|
76 |
+
log_file.write(f" layout_info (length: {len(layout_info)}), (expected: list of dicts): {layout_info}\n")
|
77 |
+
for page_index, page_info in enumerate(layout_info):
|
78 |
+
try:
|
79 |
+
with open(log_file_path, 'a') as log_file:
|
80 |
+
log_file.write(f"Processing page {page_index}\n")
|
81 |
+
log_file.write(f" Page info (expected: dict): {page_info}\n")
|
82 |
+
processed_page = self.process_page(page_info, parse_type)
|
83 |
+
processed_pages.append(processed_page)
|
84 |
+
except Exception as e:
|
85 |
+
with open(log_file_path, 'a') as log_file:
|
86 |
+
log_file.write(f"Error processing page {page_index} in process_pdf: {str(e)}\n")
|
87 |
+
log_file.write(f"Page info (expected: dict): {page_info}\n")
|
88 |
+
with open(log_file_path, 'a') as log_file:
|
89 |
+
log_file.write(f"Exiting process_pdf\n")
|
90 |
+
return {
|
91 |
+
"processed_pages": processed_pages,
|
92 |
+
"parse_type": parse_type,
|
93 |
+
}
|
94 |
+
|
95 |
+
def process_page(self, page_info, parse_type):
|
96 |
+
with open(log_file_path, 'a') as log_file:
|
97 |
+
log_file.write(f"Entering process_page\n")
|
98 |
+
log_file.write(f" page_info (expected: dict): {page_info}\n")
|
99 |
+
log_file.write(f" parse_type (expected: str): {parse_type}\n")
|
100 |
+
result = {
|
101 |
+
"page_no": page_info.get("page_info", {}).get("page_no", "unknown"),
|
102 |
+
"content": "Processed page content",
|
103 |
+
"parse_type": parse_type
|
104 |
+
}
|
105 |
+
with open(log_file_path, 'a') as log_file:
|
106 |
+
log_file.write(f"Exiting process_page\n")
|
107 |
+
return result
|
108 |
+
|
109 |
+
def process_single_pdf(input_file, output_folder, gpu_id, config, timeout, use_bf16, model, log_file_path):
|
110 |
+
start_time = time.time()
|
111 |
+
pdf_name = os.path.splitext(os.path.basename(input_file))[0]
|
112 |
+
output_subfolder = os.path.join(output_folder, pdf_name, 'auto')
|
113 |
+
os.makedirs(output_subfolder, exist_ok=True)
|
114 |
+
|
115 |
+
def timeout_handler(signum, frame):
|
116 |
+
raise TimeoutError("PDF processing timed out")
|
117 |
+
|
118 |
+
try:
|
119 |
+
signal.signal(signal.SIGALRM, timeout_handler)
|
120 |
+
signal.alarm(timeout)
|
121 |
+
|
122 |
+
if gpu_id >= 0:
|
123 |
+
torch.cuda.set_device(gpu_id)
|
124 |
+
if use_bf16 and torch.cuda.is_bf16_supported():
|
125 |
+
torch.set_default_dtype(torch.bfloat16)
|
126 |
+
else:
|
127 |
+
torch.set_default_dtype(torch.float32)
|
128 |
+
torch.set_default_device(f'cuda:{gpu_id}')
|
129 |
+
else:
|
130 |
+
if use_bf16:
|
131 |
+
torch.set_default_dtype(torch.bfloat16)
|
132 |
+
else:
|
133 |
+
torch.set_default_dtype(torch.float32)
|
134 |
+
torch.set_default_device('cpu')
|
135 |
+
|
136 |
+
pdf_data = read_file(input_file, 'rb')
|
137 |
+
|
138 |
+
# Perform PDF metadata scan
|
139 |
+
metadata = pdf_meta_scan(pdf_data)
|
140 |
+
with open(log_file_path, 'a') as log_file:
|
141 |
+
log_file.write(f"Processing PDF: {input_file}\n")
|
142 |
+
log_file.write(f"Metadata (expected: dict): {json.dumps(metadata, indent=2)}\n")
|
143 |
+
|
144 |
+
# Check if metadata indicates the PDF should be dropped
|
145 |
+
if metadata.get("_need_drop", False):
|
146 |
+
with open(log_file_path, 'a') as log_file:
|
147 |
+
log_file.write(f"Dropping PDF {input_file}: {metadata.get('_drop_reason', 'Unknown reason')}\n")
|
148 |
+
return input_file, "Dropped", None
|
149 |
+
|
150 |
+
# Check if all required fields are present in metadata
|
151 |
+
required_fields = ['total_page', 'page_width_pts', 'page_height_pts', 'image_info_per_page',
|
152 |
+
'text_len_per_page', 'imgs_per_page', 'text_layout_per_page', 'invalid_chars']
|
153 |
+
for field in required_fields:
|
154 |
+
if field not in metadata:
|
155 |
+
raise ValueError(f"Required field '{field}' not found in metadata for {input_file}")
|
156 |
+
|
157 |
+
# Extract required fields for classify function
|
158 |
+
total_page = metadata['total_page']
|
159 |
+
page_width = metadata['page_width_pts']
|
160 |
+
page_height = metadata['page_height_pts']
|
161 |
+
img_sz_list = metadata['image_info_per_page']
|
162 |
+
text_len_list = metadata['text_len_per_page']
|
163 |
+
img_num_list = metadata['imgs_per_page']
|
164 |
+
text_layout_list = metadata['text_layout_per_page']
|
165 |
+
invalid_chars = metadata['invalid_chars']
|
166 |
+
|
167 |
+
with open(log_file_path, 'a') as log_file:
|
168 |
+
log_file.write(f"Classify parameters:\n")
|
169 |
+
log_file.write(f" total_page (expected: int): {total_page}\n")
|
170 |
+
log_file.write(f" page_width (expected: int): {page_width}\n")
|
171 |
+
log_file.write(f" page_height (expected: int): {page_height}\n")
|
172 |
+
log_file.write(f" img_sz_list (expected: list of lists): {img_sz_list[:5]}...\n")
|
173 |
+
log_file.write(f" text_len_list (expected: list of ints): {text_len_list[:5]}...\n")
|
174 |
+
log_file.write(f" img_num_list (expected: list of ints): {img_num_list[:5]}...\n")
|
175 |
+
log_file.write(f" text_layout_list (expected: list of strs): {text_layout_list[:5]}...\n")
|
176 |
+
log_file.write(f" invalid_chars (expected: bool): {invalid_chars}\n")
|
177 |
+
|
178 |
+
# Classify PDF
|
179 |
+
try:
|
180 |
+
is_text_pdf, classification_results = classify(
|
181 |
+
total_page, page_width, page_height, img_sz_list[:total_page],
|
182 |
+
text_len_list[:total_page], img_num_list[:total_page],
|
183 |
+
text_layout_list[:len(text_layout_list)], invalid_chars
|
184 |
+
)
|
185 |
+
with open(log_file_path, 'a') as log_file:
|
186 |
+
log_file.write(f"Classification Results:\n")
|
187 |
+
log_file.write(f" is_text_pdf (expected: bool): {is_text_pdf}\n")
|
188 |
+
log_file.write(f" classification_results (expected: dict): {classification_results}\n")
|
189 |
+
|
190 |
+
except Exception as e:
|
191 |
+
with open(log_file_path, 'a') as log_file:
|
192 |
+
log_file.write(f"Error in classify function for {input_file}: {str(e)}\n")
|
193 |
+
return input_file, f"Classification Error: {str(e)}", None
|
194 |
+
|
195 |
+
image_writer = DiskReaderWriter(output_subfolder)
|
196 |
+
with open(log_file_path, 'a') as log_file:
|
197 |
+
log_file.write(f"Image writer initialized: {image_writer}\n")
|
198 |
+
|
199 |
+
# Create jso_useful_key as a dictionary
|
200 |
+
model_json = [] # Or load your model data here
|
201 |
+
jso_useful_key = {"_pdf_type": "", "model_list": model_json}
|
202 |
+
|
203 |
+
unipipe = UNIPipe(pdf_data, jso_useful_key, image_writer)
|
204 |
+
with open(log_file_path, 'a') as log_file:
|
205 |
+
log_file.write(f"UNIPipe initialized: {unipipe}\n")
|
206 |
+
|
207 |
+
parse_type = unipipe.pipe_classify()
|
208 |
+
with open(log_file_path, 'a') as log_file:
|
209 |
+
log_file.write(f"pipe_classify result (expected: str): {parse_type}\n")
|
210 |
+
|
211 |
+
# Add detailed logging for pipe_analyze inputs and output
|
212 |
+
with open(log_file_path, 'a') as log_file:
|
213 |
+
log_file.write(f"Detailed pipe_analyze Inputs for {input_file}:\n")
|
214 |
+
log_file.write(f" parse_type (expected: str): {parse_type}\n")
|
215 |
+
try:
|
216 |
+
layout_info = unipipe.pipe_analyze()
|
217 |
+
with open(log_file_path, 'a') as log_file:
|
218 |
+
log_file.write(f"pipe_analyze Results (expected: list of dicts, length: {len(layout_info)}): {layout_info}\n")
|
219 |
+
except Exception as e:
|
220 |
+
with open(log_file_path, 'a') as log_file:
|
221 |
+
log_file.write(f"Error in pipe_analyze for {input_file}: {str(e)}\n")
|
222 |
+
return input_file, f"pipe_analyze Error: {str(e)}", None
|
223 |
+
|
224 |
+
# Use OCR if it's not classified as a text PDF
|
225 |
+
if not is_text_pdf:
|
226 |
+
parse_type = 'ocr'
|
227 |
+
with open(log_file_path, 'a') as log_file:
|
228 |
+
log_file.write(f"parse_type after OCR check (expected: str): {parse_type}\n")
|
229 |
+
|
230 |
+
# Process the PDF using the model
|
231 |
+
try:
|
232 |
+
parse_result = model.process_pdf(pdf_data, parse_type, layout_info, log_file_path)
|
233 |
+
with open(log_file_path, 'a') as log_file:
|
234 |
+
log_file.write(f"Model process_pdf result (expected: dict): {parse_result}\n")
|
235 |
+
except Exception as e:
|
236 |
+
with open(log_file_path, 'a') as log_file:
|
237 |
+
log_file.write(f"Error in model processing for {input_file}: {str(e)}\n")
|
238 |
+
return input_file, f"Model Processing Error: {str(e)}", None
|
239 |
+
|
240 |
+
try:
|
241 |
+
markdown_content = unipipe.pipe_mk_markdown(parse_result)
|
242 |
+
with open(log_file_path, 'a') as log_file:
|
243 |
+
log_file.write(f"pipe_mk_markdown result (expected: str, length: {len(markdown_content)}): {markdown_content}\n")
|
244 |
+
except Exception as e:
|
245 |
+
with open(log_file_path, 'a') as log_file:
|
246 |
+
log_file.write(f"Error in pipe_mk_markdown for {input_file}: {str(e)}\n")
|
247 |
+
log_file.write(f" parse_result (expected: dict): {parse_result}\n")
|
248 |
+
return input_file, f"pipe_mk_markdown Error: {str(e)}", None
|
249 |
+
|
250 |
+
try:
|
251 |
+
uni_format = unipipe.pipe_mk_uni_format(parse_result)
|
252 |
+
with open(log_file_path, 'a') as log_file:
|
253 |
+
log_file.write(f"pipe_mk_uni_format result (expected: dict): {uni_format}\n")
|
254 |
+
except Exception as e:
|
255 |
+
with open(log_file_path, 'a') as log_file:
|
256 |
+
log_file.write(f"Error in pipe_mk_uni_format for {input_file}: {str(e)}\n")
|
257 |
+
log_file.write(f" parse_result (expected: dict): {parse_result}\n")
|
258 |
+
return input_file, f"pipe_mk_uni_format Error: {str(e)}", None
|
259 |
+
|
260 |
+
# Write markdown content
|
261 |
+
with open(os.path.join(output_subfolder, f'{pdf_name}.md'), 'w', encoding='utf-8') as f:
|
262 |
+
f.write(markdown_content)
|
263 |
+
|
264 |
+
# Write middle.json
|
265 |
+
with open(os.path.join(output_subfolder, 'middle.json'), 'w', encoding='utf-8') as f:
|
266 |
+
json.dump(parse_result, f, ensure_ascii=False, indent=2)
|
267 |
+
|
268 |
+
# Write model.json
|
269 |
+
with open(os.path.join(output_subfolder, 'model.json'), 'w', encoding='utf-8') as f:
|
270 |
+
json.dump(uni_format, f, ensure_ascii=False, indent=2)
|
271 |
+
|
272 |
+
# Copy original PDF
|
273 |
+
shutil.copy(input_file, os.path.join(output_subfolder, f'{pdf_name}.pdf'))
|
274 |
+
|
275 |
+
# Generate layout.pdf and spans.pdf
|
276 |
+
try:
|
277 |
+
do_parse(input_file, parse_type, output_subfolder, draw_bbox=True)
|
278 |
+
except Exception as e:
|
279 |
+
with open(log_file_path, 'a') as log_file:
|
280 |
+
log_file.write(f"Error in do_parse for {input_file}: {str(e)}\n")
|
281 |
+
return input_file, f"do_parse Error: {str(e)}", None
|
282 |
+
|
283 |
+
# Extract images
|
284 |
+
extract_images(input_file, output_subfolder)
|
285 |
+
|
286 |
+
processing_time = time.time() - start_time
|
287 |
+
with open(log_file_path, 'a') as log_file:
|
288 |
+
log_file.write(f"Successfully processed {input_file} on GPU {gpu_id} in {processing_time:.2f} seconds\n")
|
289 |
+
|
290 |
+
# Prepare result for JSONL output
|
291 |
+
result = {
|
292 |
+
"file_name": pdf_name,
|
293 |
+
"processing_time": processing_time,
|
294 |
+
"parse_type": parse_type,
|
295 |
+
"metadata": metadata,
|
296 |
+
"classification": classification_results,
|
297 |
+
"is_text_pdf": is_text_pdf
|
298 |
+
}
|
299 |
+
|
300 |
+
return input_file, "Success", result
|
301 |
+
|
302 |
+
except ValueError as ve:
|
303 |
+
with open(log_file_path, 'a') as log_file:
|
304 |
+
log_file.write(f"Metadata error: {str(ve)}\n")
|
305 |
+
return input_file, f"Metadata Error: {str(ve)}", None
|
306 |
+
|
307 |
+
except TimeoutError:
|
308 |
+
with open(log_file_path, 'a') as log_file:
|
309 |
+
log_file.write(f"Processing timed out after {timeout} seconds\n")
|
310 |
+
return input_file, "Timeout", None
|
311 |
+
|
312 |
+
except Exception as e:
|
313 |
+
with open(log_file_path, 'a') as log_file:
|
314 |
+
log_file.write(f"Error occurred: {str(e)}\n")
|
315 |
+
return input_file, f"Error: {str(e)}", None
|
316 |
+
|
317 |
+
finally:
|
318 |
+
signal.alarm(0) # Cancel the alarm
|
319 |
+
if gpu_id >= 0:
|
320 |
+
torch.cuda.empty_cache()
|
321 |
+
|
322 |
+
def process_pdf_batch(batch, output_folder, gpu_id, config, timeout, use_bf16, model, log_file_path):
|
323 |
+
results = []
|
324 |
+
for pdf_file in batch:
|
325 |
+
result = process_single_pdf(pdf_file, output_folder, gpu_id, config, timeout, use_bf16, model, log_file_path)
|
326 |
+
results.append(result)
|
327 |
+
return results
|
328 |
+
|
329 |
+
def write_to_jsonl(results, output_file):
|
330 |
+
with open(output_file, 'a') as f:
|
331 |
+
for result in results:
|
332 |
+
if result[2]: # Check if result is not None
|
333 |
+
json.dump(result[2], f)
|
334 |
+
f.write('\n')
|
335 |
+
|
336 |
+
def get_gpu_memory_usage(gpu_id):
|
337 |
+
if gpu_id < 0:
|
338 |
+
return 0, 0 # CPU mode
|
339 |
+
total_memory = torch.cuda.get_device_properties(gpu_id).total_memory
|
340 |
+
allocated_memory = torch.cuda.memory_allocated(gpu_id)
|
341 |
+
return allocated_memory, total_memory
|
342 |
+
|
343 |
+
def main():
|
344 |
+
mp.set_start_method('spawn', force=True)
|
345 |
+
|
346 |
+
args = parse_arguments()
|
347 |
+
config = load_config(args.config)
|
348 |
+
|
349 |
+
input_folder = args.input
|
350 |
+
output_folder = args.output
|
351 |
+
os.makedirs(output_folder, exist_ok=True)
|
352 |
+
|
353 |
+
pdf_files = [os.path.join(input_folder, f) for f in os.listdir(input_folder) if f.endswith('.pdf')]
|
354 |
+
|
355 |
+
num_gpus = torch.cuda.device_count()
|
356 |
+
if num_gpus == 0:
|
357 |
+
print("No GPUs available. Using CPU.")
|
358 |
+
num_gpus = 1
|
359 |
+
gpu_ids = [-1]
|
360 |
+
else:
|
361 |
+
gpu_ids = list(range(num_gpus))
|
362 |
+
|
363 |
+
num_workers = args.max_workers or min(num_gpus, os.cpu_count())
|
364 |
+
|
365 |
+
main_jsonl = os.path.join(output_folder, 'processing_results.jsonl')
|
366 |
+
temp_jsonl = os.path.join(output_folder, 'temp_results.jsonl')
|
367 |
+
log_file_path = os.path.join(output_folder, 'processing_log.txt')
|
368 |
+
|
369 |
+
# Enable deterministic mode
|
370 |
+
torch.backends.cudnn.deterministic = True
|
371 |
+
torch.backends.cudnn.benchmark = False
|
372 |
+
|
373 |
+
# Load the model
|
374 |
+
model = MagicModel(config)
|
375 |
+
|
376 |
+
results = []
|
377 |
+
with ProcessPoolExecutor(max_workers=num_workers) as executor:
|
378 |
+
for gpu_id in gpu_ids:
|
379 |
+
batch_size = args.initial_batch_size
|
380 |
+
pdf_index = 0
|
381 |
+
oom_occurred = False
|
382 |
+
while pdf_index < len(pdf_files):
|
383 |
+
batch = pdf_files[pdf_index:pdf_index + batch_size]
|
384 |
+
try:
|
385 |
+
future = executor.submit(process_pdf_batch, batch, output_folder, gpu_id, config, args.timeout, args.use_bf16, model, log_file_path)
|
386 |
+
batch_results = future.result()
|
387 |
+
results.extend(batch_results)
|
388 |
+
for result in batch_results:
|
389 |
+
write_to_jsonl([result], temp_jsonl)
|
390 |
+
|
391 |
+
# Print VRAM usage
|
392 |
+
allocated, total = get_gpu_memory_usage(gpu_id)
|
393 |
+
with open(log_file_path, 'a') as log_file:
|
394 |
+
log_file.write(f"GPU {gpu_id} - Batch size: {batch_size}, VRAM usage: {allocated/1024**3:.2f}GB / {total/1024**3:.2f}GB\n")
|
395 |
+
# If successful and OOM hasn't occurred yet, increase batch size
|
396 |
+
if not oom_occurred:
|
397 |
+
batch_size += 1
|
398 |
+
pdf_index += len(batch)
|
399 |
+
except torch.cuda.OutOfMemoryError:
|
400 |
+
# If OOM occurs, reduce batch size
|
401 |
+
oom_occurred = True
|
402 |
+
batch_size = max(MIN_BATCH_SIZE, batch_size - 1)
|
403 |
+
with open(log_file_path, 'a') as log_file:
|
404 |
+
log_file.write(f"OOM error occurred. Reducing batch size to {batch_size}\n")
|
405 |
+
torch.cuda.empty_cache()
|
406 |
+
continue
|
407 |
+
|
408 |
+
# After processing each batch, move temp JSONL to main JSONL
|
409 |
+
if os.path.exists(temp_jsonl):
|
410 |
+
with open(temp_jsonl, 'r') as temp, open(main_jsonl, 'a') as main:
|
411 |
+
shutil.copyfileobj(temp, main)
|
412 |
+
os.remove(temp_jsonl)
|
413 |
+
|
414 |
+
# Clear GPU cache after each batch
|
415 |
+
if gpu_id >= 0:
|
416 |
+
torch.cuda.empty_cache()
|
417 |
+
|
418 |
+
success_count = sum(1 for _, status, _ in results if status == "Success")
|
419 |
+
timeout_count = sum(1 for _, status, _ in results if status == "Timeout")
|
420 |
+
error_count = len(results) - success_count - timeout_count
|
421 |
+
|
422 |
+
with open(log_file_path, 'a') as log_file:
|
423 |
+
log_file.write(f"Processed {len(results)} PDFs. {success_count} succeeded, {timeout_count} timed out, {error_count} failed.\n")
|
424 |
+
|
425 |
+
with open(os.path.join(output_folder, 'processing_summary.txt'), 'w') as summary:
|
426 |
+
summary.write(f"Total PDFs processed: {len(results)}\n")
|
427 |
+
summary.write(f"Successful: {success_count}\n")
|
428 |
+
summary.write(f"Timed out: {timeout_count}\n")
|
429 |
+
summary.write(f"Failed: {error_count}\n\n")
|
430 |
+
summary.write("Failed PDFs:\n")
|
431 |
+
for pdf, status, _ in [result for result in results if result[1] != "Success"]:
|
432 |
+
summary.write(f" - {pdf}: {status}\n")
|
433 |
+
|
434 |
+
if __name__ == '__main__':
|
435 |
+
main()
|