Spaces:
Running
Running
File size: 26,378 Bytes
1d3a103 e414859 1d3a103 435abb4 7e0dde7 73feb19 1d3a103 9c491e8 5f2c216 565eabb 87e5c9c b8e1b99 c1cba4f 0f39362 665f924 435abb4 87e5c9c b8e1b99 73feb19 b8e1b99 bd3ba15 80098ed b8e1b99 87e5c9c b8e1b99 87e5c9c 5f2c216 7e0dde7 b8e1b99 87e5c9c c430753 9e8f29e e414859 c430753 7e0dde7 c1cba4f c430753 87e5c9c 9d26661 e9ed1f2 87e5c9c 435abb4 b8e1b99 1d3a103 435abb4 b8e1b99 e414859 7e0dde7 e414859 7e0dde7 73feb19 435abb4 73feb19 435abb4 e414859 9c491e8 e414859 7e0dde7 e414859 9c491e8 e414859 7e0dde7 e414859 7e0dde7 e414859 b8e1b99 1d3a103 b8e1b99 8312087 565eabb 8312087 b8e1b99 87e5c9c b8e1b99 3927544 80098ed 0bd3372 87e5c9c b8e1b99 87e5c9c f84fce9 0bd3372 73feb19 0bd3372 73feb19 87e5c9c f84fce9 87e5c9c e414859 87e5c9c 73feb19 80098ed 73feb19 87e5c9c 0407083 9e8f29e 0407083 9e8f29e 87e5c9c 9e8f29e 665f924 0407083 665f924 c13ffb4 665f924 f84fce9 0407083 665f924 87e5c9c 9e8f29e 87e5c9c 3ddba7d 839eda3 3ddba7d 50d040d 3ddba7d 50d040d 3ddba7d 50d040d d717178 50d040d b8e1b99 9e8f29e b8e1b99 87e5c9c 15c9486 87e5c9c 1d3a103 87e5c9c c430753 87e5c9c 1d3a103 87e5c9c 1d3a103 87e5c9c c13ffb4 87e5c9c 80098ed 03e9034 73feb19 c430753 87e5c9c 3927544 1d3a103 87e5c9c 1d3a103 87e5c9c 9c491e8 87e5c9c 716199b 5f2c216 87e5c9c 5f2c216 defd486 73feb19 5f2c216 87e5c9c 1d3a103 87e5c9c 1d3a103 87e5c9c 1d3a103 87e5c9c 9c491e8 bd3ba15 87e5c9c bd3ba15 73feb19 5f2c216 1d3a103 5f2c216 7e0dde7 defd486 7e0dde7 5f2c216 73feb19 5f2c216 87e5c9c bd3ba15 7e0dde7 87e5c9c 9c491e8 7e0dde7 9c491e8 435abb4 9c491e8 435abb4 9c491e8 435abb4 9c491e8 435abb4 9e8f29e 435abb4 9e8f29e 9c491e8 dcce2ac 9c491e8 435abb4 9c491e8 435abb4 9c491e8 87e5c9c 9c491e8 bd3ba15 9c491e8 435abb4 9c491e8 435abb4 dcce2ac bd3ba15 5f2c216 435abb4 87e5c9c bd3ba15 73feb19 0f39362 435abb4 87e5c9c 5f2c216 87e5c9c 7e0dde7 87e5c9c 4874c8d 87e5c9c ba90de1 b8e1b99 3927544 72c7da5 435abb4 72c7da5 435abb4 72c7da5 3927544 7e0dde7 ba90de1 1413bdf 55b49e6 1413bdf 0f39362 7452863 73feb19 7e0dde7 7452863 0f39362 6a1173d 87e5c9c 7e0dde7 87e5c9c c430753 7e0dde7 895976b 15c9486 4874c8d 15c9486 7e0dde7 15c9486 7e0dde7 e414859 7e0dde7 e414859 7e0dde7 f84fce9 7e0dde7 9e8f29e 7e0dde7 34de38e 22fd690 c13ffb4 4874c8d ba90de1 3927544 1d116fb c13ffb4 bd3ba15 c13ffb4 435abb4 c13ffb4 435abb4 c13ffb4 87e5c9c ba90de1 3927544 c13ffb4 bd3ba15 c13ffb4 1d116fb c13ffb4 80098ed 87e5c9c 7e0dde7 87e5c9c 7e0dde7 3927544 7e0dde7 87e5c9c 73feb19 7e0dde7 73feb19 87e5c9c b8e1b99 87e5c9c 80098ed 87e5c9c 34de38e 87e5c9c e414859 9c491e8 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 |
"""
app.py - the main module for the gradio app for summarization
Usage:
app.py [-h] [--share] [-m MODEL] [-nb ADD_BEAM_OPTION] [-batch TOKEN_BATCH_OPTION]
[-level {DEBUG,INFO,WARNING,ERROR}]
Details:
python app.py --help
Environment Variables:
USE_TORCH (str): whether to use torch (1) or not (0)
TOKENIZERS_PARALLELISM (str): whether to use parallelism (true) or not (false)
Optional Environment Variables:
APP_MAX_WORDS (int): the maximum number of words to use for summarization
APP_OCR_MAX_PAGES (int): the maximum number of pages to use for OCR
"""
import argparse
import contextlib
import gc
import logging
import os
import pprint as pp
import random
import re
import sys
import time
from pathlib import Path
os.environ["USE_TORCH"] = "1"
os.environ["TOKENIZERS_PARALLELISM"] = "false"
logging.basicConfig(
level=logging.INFO,
format="%(asctime)s [%(levelname)s] %(name)s - %(message)s",
datefmt="%Y-%b-%d %H:%M:%S",
)
import gradio as gr
import nltk
import torch
from cleantext import clean
from doctr.models import ocr_predictor
from aggregate import BatchAggregator
from pdf2text import convert_PDF_to_Text
from summarize import load_model_and_tokenizer, summarize_via_tokenbatches
from utils import (
contraction_aware_tokenize,
extract_batches,
load_example_filenames,
remove_stagnant_files,
remove_stopwords,
saves_summary,
textlist2html,
truncate_word_count,
)
_here = Path(__file__).parent
nltk.download("punkt", force=True, quiet=True)
nltk.download("popular", force=True, quiet=True)
# Constants & Globals
MODEL_OPTIONS = [
"pszemraj/long-t5-tglobal-base-16384-book-summary",
"pszemraj/long-t5-tglobal-base-sci-simplify",
"pszemraj/long-t5-tglobal-base-sci-simplify-elife",
"pszemraj/long-t5-tglobal-base-16384-booksci-summary-v1",
"pszemraj/pegasus-x-large-book-summary",
] # models users can choose from
BEAM_OPTIONS = [2, 3, 4] # beam sizes users can choose from
TOKEN_BATCH_OPTIONS = [
1024,
1536,
2048,
2560,
3072,
] # token batch sizes users can choose from
SUMMARY_PLACEHOLDER = "<p><em>Output will appear below:</em></p>"
AGGREGATE_MODEL = "MBZUAI/LaMini-Flan-T5-783M" # model to use for aggregation
# if duplicating space: uncomment this line to adjust the max words
# os.environ["APP_MAX_WORDS"] = str(2048) # set the max words to 2048
# os.environ["APP_OCR_MAX_PAGES"] = str(40) # set the max pages to 40
# os.environ["APP_AGG_FORCE_CPU"] = str(1) # force cpu for aggregation
aggregator = BatchAggregator(
AGGREGATE_MODEL, force_cpu=os.environ.get("APP_AGG_FORCE_CPU", False)
)
def aggregate_text(
summary_text: str,
text_file: gr.inputs.File = None,
) -> str:
"""
Aggregate the text from the batches.
NOTE: you should probably include the BatchAggregator object as a fn arg if using this code
:param batches_html: The batches to aggregate, in html format
:param text_file: The text file to append the aggregate summary to
:return: The aggregate summary in html format
"""
if summary_text is None or summary_text == SUMMARY_PLACEHOLDER:
logging.error("No text provided. Make sure a summary has been generated first.")
return "Error: No text provided. Make sure a summary has been generated first."
try:
extracted_batches = extract_batches(summary_text)
except Exception as e:
logging.info(summary_text)
logging.info(f"the batches html is: {type(summary_text)}")
return f"Error: unable to extract batches - check input: {e}"
if not extracted_batches:
logging.error("unable to extract batches - check input")
return "Error: unable to extract batches - check input"
out_path = None
if text_file is not None:
out_path = text_file.name # assuming name attribute stores the file path
content_batches = [batch["content"] for batch in extracted_batches]
full_summary = aggregator.infer_aggregate(content_batches)
# if a path that exists is provided, append the summary with markdown formatting
if out_path:
out_path = Path(out_path)
try:
with open(out_path, "a", encoding="utf-8") as f:
f.write("\n\n## Aggregate Summary\n\n")
f.write(
"- This is an instruction-based LLM aggregation of the previous 'summary batches'.\n"
)
f.write(f"- Aggregation model: {aggregator.model_name}\n\n")
f.write(f"{full_summary}\n\n")
logging.info(f"Updated {out_path} with aggregate summary")
except Exception as e:
logging.error(f"unable to update {out_path} with aggregate summary: {e}")
full_summary_html = f"""
<div style="
margin-bottom: 20px;
font-size: 18px;
line-height: 1.5em;
color: #333;
">
<h2 style="font-size: 22px; color: #555;">Aggregate Summary:</h2>
<p style="white-space: pre-line;">{full_summary}</p>
</div>
"""
return full_summary_html
def predict(
input_text: str,
model_name: str,
token_batch_length: int = 1024,
empty_cache: bool = True,
**settings,
) -> list:
"""
predict - helper fn to support multiple models for summarization at once
:param str input_text: the input text to summarize
:param str model_name: model name to use
:param int token_batch_length: the length of the token batches to use
:param bool empty_cache: whether to empty the cache before loading a new= model
:return: list of dicts with keys "summary" and "score"
"""
if torch.cuda.is_available() and empty_cache:
torch.cuda.empty_cache()
model, tokenizer = load_model_and_tokenizer(model_name)
summaries = summarize_via_tokenbatches(
input_text,
model,
tokenizer,
batch_length=token_batch_length,
**settings,
)
del model
del tokenizer
gc.collect()
return summaries
def proc_submission(
input_text: str,
model_name: str,
num_beams: int,
token_batch_length: int,
length_penalty: float,
repetition_penalty: float,
no_repeat_ngram_size: int,
predrop_stopwords: bool,
max_input_length: int = 6144,
):
"""
proc_submission - a helper function for the gradio module to process submissions
Args:
input_text (str): the input text to summarize
model_name (str): the hf model tag of the model to use
num_beams (int): the number of beams to use
token_batch_length (int): the length of the token batches to use
length_penalty (float): the length penalty to use
repetition_penalty (float): the repetition penalty to use
no_repeat_ngram_size (int): the no repeat ngram size to use
predrop_stopwords (bool): whether to pre-drop stopwords before truncating/summarizing
max_input_length (int, optional): the maximum input length to use. Defaults to 6144.
Note:
the max_input_length is set to 6144 by default, but can be changed by setting the
environment variable APP_MAX_WORDS to a different value.
Returns:
tuple (4): a tuple containing the following:
"""
remove_stagnant_files() # clean up old files
settings = {
"length_penalty": float(length_penalty),
"repetition_penalty": float(repetition_penalty),
"no_repeat_ngram_size": int(no_repeat_ngram_size),
"encoder_no_repeat_ngram_size": 4,
"num_beams": int(num_beams),
"min_length": 4,
"max_length": int(token_batch_length // 4),
"early_stopping": True,
"do_sample": False,
}
max_input_length = int(os.environ.get("APP_MAX_WORDS", max_input_length))
logging.info(
f"max_input_length set to: {max_input_length}. pre-drop stopwords: {predrop_stopwords}"
)
st = time.perf_counter()
history = {}
cln_text = clean(input_text, lower=False)
parsed_cln_text = remove_stopwords(cln_text) if predrop_stopwords else cln_text
logging.info(
f"pre-truncation word count: {len(contraction_aware_tokenize(parsed_cln_text))}"
)
truncation_validated = truncate_word_count(
parsed_cln_text, max_words=max_input_length
)
if truncation_validated["was_truncated"]:
model_input_text = truncation_validated["processed_text"]
# create elaborate HTML warning
input_wc = len(contraction_aware_tokenize(parsed_cln_text))
msg = f"""
<div style="background-color: #FFA500; color: white; padding: 20px;">
<h3>Warning</h3>
<p>Input text was truncated to {max_input_length} words. That's about {100*max_input_length/input_wc:.2f}% of the original text.</p>
<p>Dropping stopwords is set to {predrop_stopwords}. If this is not what you intended, please validate the advanced settings.</p>
</div>
"""
logging.warning(msg)
history["WARNING"] = msg
else:
model_input_text = truncation_validated["processed_text"]
msg = None
if len(input_text) < 50:
# this is essentially a different case from the above
msg = f"""
<div style="background-color: #880808; color: white; padding: 20px;">
<br>
<img src="https://i.imgflip.com/7kadd9.jpg" alt="no text">
<br>
<h3>Error</h3>
<p>Input text is too short to summarize. Detected {len(input_text)} characters.
Please load text by selecting an example from the dropdown menu or by pasting text into the text box.</p>
</div>
"""
logging.warning(msg)
logging.warning("RETURNING EMPTY STRING")
history["WARNING"] = msg
return msg, "<strong>No summary generated.</strong>", "", []
_summaries = predict(
input_text=model_input_text,
model_name=model_name,
token_batch_length=token_batch_length,
**settings,
)
sum_text = [s["summary"][0].strip() + "\n" for s in _summaries]
sum_scores = [
f" - Batch Summary {i}: {round(s['summary_score'],4)}"
for i, s in enumerate(_summaries)
]
full_summary = textlist2html(sum_text)
history["Summary Scores"] = "<br><br>"
scores_out = "\n".join(sum_scores)
rt = round((time.perf_counter() - st) / 60, 2)
logging.info(f"Runtime: {rt} minutes")
html = ""
html += f"<p>Runtime: {rt} minutes with model: {model_name}</p>"
if msg is not None:
html += msg
html += ""
settings["remove_stopwords"] = predrop_stopwords
settings["model_name"] = model_name
saved_file = saves_summary(summarize_output=_summaries, outpath=None, **settings)
return html, full_summary, scores_out, saved_file
def load_single_example_text(
example_path: str or Path,
max_pages: int = 20,
) -> str:
"""
load_single_example_text - loads a single example text file
:param strorPath example_path: name of the example to load
:param int max_pages: the maximum number of pages to load from a PDF
:return str: the text of the example
"""
global name_to_path, ocr_model
full_ex_path = name_to_path[example_path]
full_ex_path = Path(full_ex_path)
if full_ex_path.suffix in [".txt", ".md"]:
with open(full_ex_path, "r", encoding="utf-8", errors="ignore") as f:
raw_text = f.read()
text = clean(raw_text, lower=False)
elif full_ex_path.suffix == ".pdf":
logging.info(f"Loading PDF file {full_ex_path}")
max_pages = int(os.environ.get("APP_OCR_MAX_PAGES", max_pages))
logging.info(f"max_pages set to: {max_pages}")
conversion_stats = convert_PDF_to_Text(
full_ex_path,
ocr_model=ocr_model,
max_pages=max_pages,
)
text = conversion_stats["converted_text"]
else:
logging.error(f"Unknown file type {full_ex_path.suffix}")
text = "ERROR - check example path"
return text
def load_uploaded_file(file_obj, max_pages: int = 20, lower: bool = False) -> str:
"""
load_uploaded_file - loads a file uploaded by the user
:param file_obj (POTENTIALLY list): Gradio file object inside a list
:param int max_pages: the maximum number of pages to load from a PDF
:param bool lower: whether to lowercase the text
:return str: the text of the file
"""
global ocr_model
logger = logging.getLogger(__name__)
# check if mysterious file object is a list
if isinstance(file_obj, list):
file_obj = file_obj[0]
file_path = Path(file_obj.name)
try:
logger.info(f"Loading file:\t{file_path}")
if file_path.suffix in [".txt", ".md"]:
with open(file_path, "r", encoding="utf-8", errors="ignore") as f:
raw_text = f.read()
text = clean(raw_text, lower=lower)
elif file_path.suffix == ".pdf":
logger.info(f"loading a PDF file: {file_path.name}")
max_pages = int(os.environ.get("APP_OCR_MAX_PAGES", max_pages))
logger.info(f"max_pages is: {max_pages}. Starting conversion...")
conversion_stats = convert_PDF_to_Text(
file_path,
ocr_model=ocr_model,
max_pages=max_pages,
)
text = conversion_stats["converted_text"]
else:
logger.error(f"Unknown file type:\t{file_path.suffix}")
text = "ERROR - check file - unknown file type. PDF, TXT, and MD are supported."
return text
except Exception as e:
logger.error(f"Trying to load file:\t{file_path},\nerror:\t{e}")
return f"Error: Could not read file {file_path.name}. Make sure it is a PDF, TXT, or MD file."
def parse_args():
"""arguments for the command line interface"""
parser = argparse.ArgumentParser(
description="Document Summarization with Long-Document Transformers - Demo",
formatter_class=argparse.ArgumentDefaultsHelpFormatter,
epilog="Runs a local-only web UI to summarize documents. pass --share for a public link to share.",
)
parser.add_argument(
"--share",
dest="share",
action="store_true",
help="Create a public link to share",
)
parser.add_argument(
"-m",
"--model",
type=str,
default=None,
help=f"Add a custom model to the list of models: {pp.pformat(MODEL_OPTIONS, compact=True)}",
)
parser.add_argument(
"-nb",
"--add_beam_option",
type=int,
default=None,
help=f"Add a beam search option to the demo UI options, default: {pp.pformat(BEAM_OPTIONS, compact=True)}",
)
parser.add_argument(
"-batch",
"--token_batch_option",
type=int,
default=None,
help=f"Add a token batch size to the demo UI options, default: {pp.pformat(TOKEN_BATCH_OPTIONS, compact=True)}",
)
parser.add_argument(
"-max_agg",
"-2x",
"--aggregator_beam_boost",
dest="aggregator_beam_boost",
action="store_true",
help="Double the number of beams for the aggregator during beam search",
)
parser.add_argument(
"-level",
"--log_level",
type=str,
default="INFO",
choices=["DEBUG", "INFO", "WARNING", "ERROR"],
help="Set the logging level",
)
return parser.parse_args()
if __name__ == "__main__":
"""main - the main function of the app"""
logger = logging.getLogger(__name__)
args = parse_args()
logger.setLevel(args.log_level)
logger.info(f"args: {pp.pformat(args.__dict__, compact=True)}")
# add any custom options
if args.model is not None:
logger.info(f"Adding model {args.model} to the list of models")
MODEL_OPTIONS.append(args.model)
if args.add_beam_option is not None:
logger.info(f"Adding beam search option {args.add_beam_option} to the list")
BEAM_OPTIONS.append(args.add_beam_option)
if args.token_batch_option is not None:
logger.info(f"Adding token batch option {args.token_batch_option} to the list")
TOKEN_BATCH_OPTIONS.append(args.token_batch_option)
if args.aggregator_beam_boost:
logger.info("Doubling aggregator num_beams")
_agg_cfg = aggregator.get_generation_config()
_agg_cfg["num_beams"] = _agg_cfg["num_beams"] * 2
aggregator.update_generation_config(**_agg_cfg)
logger.info("Loading OCR model")
with contextlib.redirect_stdout(None):
ocr_model = ocr_predictor(
"db_resnet50",
"crnn_mobilenet_v3_large",
pretrained=True,
assume_straight_pages=True,
)
# load the examples
name_to_path = load_example_filenames(_here / "examples")
logger.info(f"Loaded {len(name_to_path)} examples")
demo = gr.Blocks(title="Document Summarization with Long-Document Transformers")
_examples = list(name_to_path.keys())
logger.info("Starting app instance")
with demo:
gr.Markdown("# Document Summarization with Long-Document Transformers")
gr.Markdown(
"An example use case for fine-tuned long document transformers. Model(s) are trained on [book summaries](https://hf.co/datasets/kmfoda/booksum). Architectures [in this demo](https://hf.co/spaces/pszemraj/document-summarization) are [LongT5-base](https://hf.co/pszemraj/long-t5-tglobal-base-16384-book-summary) and [Pegasus-X-Large](https://hf.co/pszemraj/pegasus-x-large-book-summary)."
)
with gr.Column():
gr.Markdown("## Load Inputs & Select Parameters")
gr.Markdown(
"""Enter/paste text below, or upload a file. Pick a model & adjust params (_optional_), and press **Summarize!**
See [the guide doc](https://gist.github.com/pszemraj/722a7ba443aa3a671b02d87038375519) for details.
"""
)
with gr.Row(variant="compact"):
with gr.Column(scale=0.5, variant="compact"):
model_name = gr.Dropdown(
choices=MODEL_OPTIONS,
value=MODEL_OPTIONS[0],
label="Model Name",
)
num_beams = gr.Radio(
choices=BEAM_OPTIONS,
label="Beam Search: # of Beams",
value=BEAM_OPTIONS[0],
)
load_examples_button = gr.Button(
"Load Example in Dropdown",
)
load_file_button = gr.Button("Load & Process File")
with gr.Column(variant="compact"):
example_name = gr.Dropdown(
_examples,
label="Examples",
value=random.choice(_examples),
)
uploaded_file = gr.File(
label="File Upload",
file_count="single",
file_types=[".txt", ".md", ".pdf"],
type="file",
)
with gr.Row():
input_text = gr.Textbox(
lines=4,
max_lines=12,
label="Text to Summarize",
placeholder="Enter text to summarize, the text will be cleaned and truncated on Spaces. Narrative, academic (both papers and lecture transcription), and article text work well. May take a bit to generate depending on the input text :)",
)
gr.Markdown("---")
with gr.Column():
gr.Markdown("## Generate Summary")
with gr.Row():
summarize_button = gr.Button(
"Summarize!",
variant="primary",
)
gr.Markdown(
"_Summarization should take ~1-2 minutes for most settings, but may extend up to 5-10 minutes in some scenarios._"
)
output_text = gr.HTML("<p><em>Output will appear below:</em></p>")
with gr.Column():
gr.Markdown("### Results & Scores")
with gr.Row():
with gr.Column(variant="compact"):
gr.Markdown(
"Download the summary as a text file, with parameters and scores."
)
text_file = gr.File(
label="Download as Text File",
file_count="single",
type="file",
interactive=False,
)
with gr.Column(variant="compact"):
gr.Markdown(
"Scores **roughly** represent the summary quality as a measure of the model's 'confidence'. less-negative numbers (closer to 0) are better."
)
summary_scores = gr.Textbox(
label="Summary Scores",
placeholder="Summary scores will appear here",
)
with gr.Column(variant="panel"):
gr.Markdown("### **Summary Output**")
summary_text = gr.HTML(
label="Summary",
value="<center><i>Summary will appear here!</i></center>",
)
with gr.Column():
gr.Markdown("### **Aggregate Summary Batches**")
gr.Markdown(
"_Note: this is an experimental feature. Feedback welcome in the [discussions](https://hf.co/spaces/pszemraj/document-summarization/discussions)!_"
)
with gr.Row():
aggregate_button = gr.Button(
"Aggregate!",
variant="primary",
)
gr.Markdown(
f"""Aggregate the above batches into a cohesive summary.
- A secondary instruct-tuned LM consolidates info
- Current model: [{AGGREGATE_MODEL}](https://hf.co/{AGGREGATE_MODEL})
"""
)
with gr.Column(variant="panel"):
aggregated_summary = gr.HTML(
label="Aggregate Summary",
value="<center><i>Aggregate summary will appear here!</i></center>",
)
gr.Markdown(
"\n\n_Aggregate summary is also appended to the bottom of the `.txt` file._"
)
gr.Markdown("---")
with gr.Column():
gr.Markdown("### Advanced Settings")
gr.Markdown(
"Refer to [the guide doc](https://gist.github.com/pszemraj/722a7ba443aa3a671b02d87038375519) for what these are, and how they impact _quality_ and _speed_."
)
with gr.Row(variant="compact"):
length_penalty = gr.Slider(
minimum=0.3,
maximum=1.1,
label="length penalty",
value=0.7,
step=0.05,
)
token_batch_length = gr.Radio(
choices=TOKEN_BATCH_OPTIONS,
label="token batch length",
# select median option
value=TOKEN_BATCH_OPTIONS[len(TOKEN_BATCH_OPTIONS) // 2],
)
with gr.Row(variant="compact"):
repetition_penalty = gr.Slider(
minimum=1.0,
maximum=5.0,
label="repetition penalty",
value=1.5,
step=0.1,
)
no_repeat_ngram_size = gr.Radio(
choices=[2, 3, 4, 5],
label="no repeat ngram size",
value=3,
)
predrop_stopwords = gr.Checkbox(
label="Drop Stopwords (Pre-Truncation)",
value=False,
)
with gr.Column():
gr.Markdown("## About")
gr.Markdown(
"- Models are fine-tuned on the [🅱️ookSum dataset](https://arxiv.org/abs/2105.08209). The goal was to create a model that generalizes well and is useful for summarizing text in academic and everyday use."
)
gr.Markdown(
"- _Update April 2023:_ Additional models fine-tuned on the [PLOS](https://hf.co/datasets/pszemraj/scientific_lay_summarisation-plos-norm) and [ELIFE](https://hf.co/datasets/pszemraj/scientific_lay_summarisation-elife-norm) subsets of the [scientific lay summaries](https://arxiv.org/abs/2210.09932) dataset are available (see dropdown at the top)."
)
gr.Markdown(
"Adjust the max input words & max PDF pages for OCR by duplicating this space and [setting the environment variables](https://hf.co/docs/hub/spaces-overview#managing-secrets) `APP_MAX_WORDS` and `APP_OCR_MAX_PAGES` to the desired integer values."
)
gr.Markdown("---")
load_examples_button.click(
fn=load_single_example_text, inputs=[example_name], outputs=[input_text]
)
load_file_button.click(
fn=load_uploaded_file, inputs=uploaded_file, outputs=[input_text]
)
summarize_button.click(
fn=proc_submission,
inputs=[
input_text,
model_name,
num_beams,
token_batch_length,
length_penalty,
repetition_penalty,
no_repeat_ngram_size,
predrop_stopwords,
],
outputs=[output_text, summary_text, summary_scores, text_file],
)
aggregate_button.click(
fn=aggregate_text,
inputs=[summary_text, text_file],
outputs=[aggregated_summary],
)
demo.launch(enable_queue=True, share=args.share)
|