app.py

import gradio as gr
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer, GenerationConfig

# Load the model and tokenizer
model = AutoModelForSeq2SeqLM.from_pretrained("pszemraj/flan-t5-large-grammar-synthesis")
tokenizer = AutoTokenizer.from_pretrained("pszemraj/flan-t5-large-grammar-synthesis")


def correct_text(text, genConfig):
    inputs = tokenizer.encode("" + text, return_tensors="pt")
    outputs = model.generate(inputs, **genConfig.to_dict())

    corrected_text = tokenizer.decode(outputs[0], skip_special_tokens=True)
    return corrected_text

def respond(text, max_new_tokens, min_new_tokens, num_beams, num_beam_groups, temperature, top_k, top_p, no_repeat_ngram_size, guidance_scale, do_sample: bool):
    config = GenerationConfig(
        max_new_tokens=max_new_tokens,
        min_new_tokens=min_new_tokens,
        num_beams=num_beams,
        num_beam_groups=num_beam_groups,
        temperature=float(temperature),
        top_k=top_k,
        top_p=float(top_p),
        no_repeat_ngram_size=no_repeat_ngram_size,
        early_stopping=True,
        do_sample=do_sample
    )

    if guidance_scale > 0:
        config.guidance_scale = float(guidance_scale)
    
    corrected = correct_text(text, config)
    yield corrected

def update_prompt(prompt):
    return prompt

# Create the Gradio interface
with gr.Blocks() as demo:
    gr.Markdown("""# Grammar Correction App""")
    prompt_box = gr.Textbox(placeholder="Enter your prompt here...")
    output_box = gr.Textbox()

    # Sample prompts
    with gr.Row():
        samp1 = gr.Button("we shood buy an car")
        samp2 = gr.Button("she is more taller")
        samp3 = gr.Button("John and i saw a sheep over their.")
        
        samp1.click(update_prompt, samp1, prompt_box)
        samp2.click(update_prompt, samp2, prompt_box)
        samp3.click(update_prompt, samp3, prompt_box)
    submitBtn = gr.Button("Submit")
    
    with gr.Accordion("Generation Parameters:", open=False):
        max_tokens  = gr.Slider(minimum=1,   maximum=256,   value=50,  step=1,    label="Max New Tokens")
        min_tokens  = gr.Slider(minimum=0,   maximum=256,   value=0,   step=1,    label="Min New Tokens")
        num_beams   = gr.Slider(minimum=1,   maximum=20,    value=5,   step=1,    label="Num Beams")
        beam_groups = gr.Slider(minimum=1,   maximum=20,    value=1,   step=1,    label="Num Beams Groups")
        temperature = gr.Slider(minimum=0.1, maximum=100.0, value=0.7, step=0.1,  label="Temperature")
        top_k       = gr.Slider(minimum=0,   maximum=200,   value=50,  step=1,    label="Top-k")
        top_p       = gr.Slider(minimum=0.1, maximum=1.0,   value=1.0, step=0.05, label="Top-p (nucleus sampling)")
        guideScale  = gr.Slider(minimum=0.1, maximum=50.0,  value=1.0, step=0.1, label="Guidance Scale")
        no_repeat_ngram_size = gr.Slider(0, 20, value=0, step=1, label="Limit N-grams of given Size")
        do_sample = gr.Checkbox(value=True, label="Do Sampling")
        
    submitBtn.click(respond, [prompt_box, max_tokens, min_tokens, num_beams, beam_groups, temperature, top_k, top_p, no_repeat_ngram_size, guideScale, do_sample], output_box)

demo.launch()