shorter.py

from transformers import pipeline

def generate_summary(text: str, device: str="cpu") -> str:
    """ summarize decoded text using flan-t5-small """
    global summarizer # to not inicialize every time function is called

    def summary():
        # chunking if file is large
        if len(text) < 200:
            return text

        max_chunk = 512
        chunks = [text[i:i+max_chunk] for i in range(0, len(text), max_chunk)]

        summaries = [
            summarizer(chunk, max_length=100, min_length=25, do_sample=False)[0]['summary_text']
            for chunk in chunks
        ]

        combined_text = " ".join(summaries)
        input_len = len(combined_text.split())
        # set max_length to be ~50% of input length but capped

        adaptive_max_length = min(int(input_len * 0.5), 300)
        adaptive_min_length = min(50, adaptive_max_length)
        return summarizer(
            combined_text, max_length=adaptive_max_length, min_length=adaptive_min_length, do_sample=False
        )[0]["summary_text"]

    try:
        if summarizer:
            final_summary = summary()
    except NameError:
        summarizer = pipeline("summarization", model="google/flan-t5-small", device=device)
        final_summary = summary()

    return final_summary