Spaces:
Sleeping
Sleeping
| import logging | |
| import time | |
| from pathlib import Path | |
| import gradio as gr | |
| import nltk | |
| from cleantext import clean | |
| from summarize import load_model_and_tokenizer, summarize_via_tokenbatches | |
| from utils import load_example_filenames, truncate_word_count | |
| _here = Path(__file__).parent | |
| nltk.download("stopwords") # TODO=find where this requirement originates from | |
| logging.basicConfig( | |
| level=logging.INFO, format="%(asctime)s - %(levelname)s - %(message)s" | |
| ) | |
| def proc_submission( | |
| input_text: str, | |
| model_size: str, | |
| num_beams, | |
| token_batch_length, | |
| length_penalty, | |
| repetition_penalty, | |
| no_repeat_ngram_size, | |
| max_input_length: int = 768, | |
| ): | |
| """ | |
| proc_submission - a helper function for the gradio module | |
| Parameters | |
| ---------- | |
| input_text : str, required, the text to be processed | |
| max_input_length : int, optional, the maximum length of the input text, default=512 | |
| Returns | |
| ------- | |
| str of HTML, the interactive HTML form for the model | |
| """ | |
| 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, | |
| } | |
| st = time.perf_counter() | |
| history = {} | |
| clean_text = clean(input_text, lower=False) | |
| max_input_length = 1024 if model_size == "base" else max_input_length | |
| processed = truncate_word_count(clean_text, max_input_length) | |
| if processed["was_truncated"]: | |
| tr_in = processed["truncated_text"] | |
| msg = f"Input text was truncated to {max_input_length} words (based on whitespace)" | |
| logging.warning(msg) | |
| history["WARNING"] = msg | |
| else: | |
| tr_in = input_text | |
| msg = None | |
| _summaries = summarize_via_tokenbatches( | |
| tr_in, | |
| model_sm if model_size == "base" else model, | |
| tokenizer_sm if model_size == "base" else tokenizer, | |
| batch_length=token_batch_length, | |
| **settings, | |
| ) | |
| sum_text = [f"Section {i}: " + s["summary"][0] for i, s in enumerate(_summaries)] | |
| sum_scores = [ | |
| f" - Section {i}: {round(s['summary_score'],4)}" | |
| for i, s in enumerate(_summaries) | |
| ] | |
| sum_text_out = "\n".join(sum_text) | |
| history["Summary Scores"] = "<br><br>" | |
| scores_out = "\n".join(sum_scores) | |
| rt = round((time.perf_counter() - st) / 60, 2) | |
| print(f"Runtime: {rt} minutes") | |
| html = "" | |
| html += f"<p>Runtime: {rt} minutes on CPU</p>" | |
| if msg is not None: | |
| html += f"<h2>WARNING:</h2><hr><b>{msg}</b><br><br>" | |
| html += "" | |
| return html, sum_text_out, scores_out | |
| def load_single_example_text( | |
| example_path: str or Path, | |
| ): | |
| """ | |
| load_single_example - a helper function for the gradio module to load examples | |
| Returns: | |
| list of str, the examples | |
| """ | |
| global name_to_path | |
| full_ex_path = name_to_path[example_path] | |
| full_ex_path = Path(full_ex_path) | |
| # load the examples into a list | |
| with open(full_ex_path, "r", encoding="utf-8", errors="ignore") as f: | |
| raw_text = f.read() | |
| text = clean(raw_text, lower=False) | |
| return text | |
| if __name__ == "__main__": | |
| model, tokenizer = load_model_and_tokenizer("pszemraj/led-large-book-summary") | |
| model_sm, tokenizer_sm = load_model_and_tokenizer("pszemraj/led-base-book-summary") | |
| name_to_path = load_example_filenames(_here / "examples") | |
| logging.info(f"Loaded {len(name_to_path)} examples") | |
| demo = gr.Blocks() | |
| with demo: | |
| gr.Markdown("# Long-Form Summarization: LED & BookSum") | |
| gr.Markdown( | |
| "A simple demo using a fine-tuned LED model to summarize long-form text. See [model card](https://huggingface.co/pszemraj/led-large-book-summary) for a notebook with GPU inference (much faster) on Colab." | |
| ) | |
| with gr.Column(): | |
| gr.Markdown("## Load Inputs & Select Parameters") | |
| gr.Markdown( | |
| "Enter your text below or choose an example, and select the model size and parameters. Press the button to load examples." | |
| ) | |
| model_size = gr.inputs.Radio( | |
| choices=["base", "large"], label="model size", default="large" | |
| ) | |
| num_beams = gr.inputs.Slider( | |
| minimum=2, maximum=4, label="num_beams", default=2, step=1 | |
| ) | |
| token_batch_length = gr.inputs.Slider( | |
| minimum=512, | |
| maximum=1024, | |
| label="token_batch_length", | |
| default=512, | |
| step=256, | |
| ) | |
| length_penalty = gr.inputs.Slider( | |
| minimum=0.5, maximum=1.0, label="length penalty", default=0.7, step=0.05 | |
| ) | |
| repetition_penalty = gr.inputs.Slider( | |
| minimum=1.0, | |
| maximum=5.0, | |
| label="repetition penalty", | |
| default=3.5, | |
| step=0.1, | |
| ) | |
| no_repeat_ngram_size = gr.inputs.Slider( | |
| minimum=2, maximum=4, label="no repeat ngram size", default=3, step=1 | |
| ) | |
| example_name = gr.Dropdown( | |
| list(name_to_path.keys()), | |
| label="Load Example", | |
| ) | |
| load_examples_button = gr.Button( | |
| "Load Example", | |
| ) | |
| input_text = gr.Textbox( | |
| lines=6, | |
| label="input text", | |
| 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 :)", | |
| ) | |
| with gr.Column(): | |
| gr.Markdown("## Generate Summary") | |
| gr.Markdown("Summary generation should take approximately 1-2 minutes for most settings.") | |
| summarize_button = gr.Button("Summarize!") | |
| output_text = gr.HTML("<p><em>Output will appear below:</em></p>") | |
| gr.Markdown("### Summary Output") | |
| summary_text = gr.Textbox( | |
| label="Summary", placeholder="The generated summary will appear here" | |
| ) | |
| gr.Markdown( | |
| "The summary scores can be thought of as representing the quality of the summary. less-negative numbers (closer to 0) are better:" | |
| ) | |
| summary_scores = gr.Textbox( | |
| label="Summary Scores", placeholder="Summary scores will appear here" | |
| ) | |
| with gr.Column(): | |
| gr.Markdown("## About the Model") | |
| gr.Markdown( | |
| "- [This model](https://huggingface.co/pszemraj/led-large-book-summary) is a fine-tuned checkpoint of [allenai/led-large-16384](https://huggingface.co/allenai/led-large-16384) on the [BookSum dataset](https://arxiv.org/abs/2105.08209).The goal was to create a model that can generalize well and is useful in summarizing lots of text in academic and daily usage." | |
| ) | |
| gr.Markdown( | |
| "- The model can be used with tag [pszemraj/led-large-book-summary](https://huggingface.co/pszemraj/led-large-book-summary). See the model card for details on usage & a notebook for a tutorial." | |
| ) | |
| load_examples_button.click( | |
| fn=load_single_example_text, inputs=[example_name], outputs=[input_text] | |
| ) | |
| summarize_button.click( | |
| fn=proc_submission, | |
| inputs=[ | |
| input_text, | |
| model_size, | |
| num_beams, | |
| token_batch_length, | |
| length_penalty, | |
| repetition_penalty, | |
| no_repeat_ngram_size, | |
| ], | |
| outputs=[output_text, summary_text, summary_scores], | |
| ) | |
| demo.launch(enable_queue=True, prevent_thread_lock=True) | |