complete_poem.py

import gradio as gr
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("vinai/phobert-base")
# Define your models
models = {
    "Lục Bát": AutoModelForCausalLM.from_pretrained(
        "Libosa2707/vietnamese-poem-luc-bat-gpt2"
    ),
    "Bảy Chữ": AutoModelForCausalLM.from_pretrained(
        "Libosa2707/vietnamese-poem-bay-chu-gpt2"
    ),
    "Tám Chữ": AutoModelForCausalLM.from_pretrained(
        "Libosa2707/vietnamese-poem-tam-chu-gpt2"
    ),
    "Năm Chữ": AutoModelForCausalLM.from_pretrained(
        "Libosa2707/vietnamese-poem-nam-chu-gpt2"
    ),
}


def complete_poem(text, style):
    # Preprocess the input text
    text = text.strip()
    text = text.lower()

    # Choose the model based on the selected style
    model = models[style]

    # Tokenize the input line
    input_ids = tokenizer.encode(text, return_tensors="pt")[:, :-1]

    # Generate text
    output = model.generate(input_ids, max_length=100, do_sample=True, temperature=0.7)

    # Decode the output
    generated_text = tokenizer.decode(
        output[:, input_ids.shape[-1] :][0], skip_special_tokens=True
    )

    text = text + " " + generated_text

    # Post-process the output
    text = text.replace("<unk>", "\n")
    pretty_text = ""
    for idx, line in enumerate(text.split("\n")):
        line = line.strip()
        if not line:
            continue
        line = line[0].upper() + line[1:]
        pretty_text += line + "\n"

    return pretty_text


complete_poem_interface = gr.Interface(
    title="Viết tiếp áng thơ hay...",
    fn=complete_poem,
    inputs=[
        gr.components.Textbox(
            lines=1,
            placeholder="Tôi đâu có biết làm thơ",
            label="Những áng thơ đầu tiên",
        ),
        gr.components.Dropdown(
            choices=["Lục Bát", "Bảy Chữ", "Tám Chữ", "Năm Chữ"],
            label="Kiểu thơ",
            value="Lục Bát",
        ),
    ],
    outputs="text",
)