app.py

import logging
import pathlib
import gradio as gr
import pandas as pd
from gt4sd.algorithms.generation.hugging_face import (
    HuggingFaceSeq2SeqGenerator,
    HuggingFaceGenerationAlgorithm,
)
from transformers import AutoTokenizer

logger = logging.getLogger(__name__)
logger.addHandler(logging.NullHandler())

task2prefix = {
    "forward": "Predict the product of the following reaction: ",
    "retrosynthesis": "Predict the reaction that produces the following product: ",
    "paragraph to actions": "Which actions are described in the following paragraph: ",
    "molecular captioning": "Caption the following smile: ",
    "text-conditional de novo generation": "Write in SMILES the described molecule: ",
}


def run_inference(
    model_name_or_path: str,
    task: str,
    prompt: str,
    num_beams: int,
):
    instruction = task2prefix[task]

    config = HuggingFaceSeq2SeqGenerator(
        algorithm_version=model_name_or_path,
        prefix=instruction,
        prompt=prompt,
        num_beams=num_beams,
    )

    model = HuggingFaceGenerationAlgorithm(config)
    tokenizer = AutoTokenizer.from_pretrained("t5-small")

    text = list(model.sample(1))[0]

    text = text.replace(instruction + prompt, "")
    text = text.split(tokenizer.eos_token)[0]
    text = text.replace(tokenizer.pad_token, "")
    text = text.strip()

    return text


if __name__ == "__main__":

    models = [
        "multitask-text-and-chemistry-t5-small-standard",
        "multitask-text-and-chemistry-t5-small-augm"
    ]

    metadata_root = pathlib.Path(__file__).parent.joinpath("model_cards")

    examples = pd.read_csv(metadata_root.joinpath("examples.csv"), header=None).fillna(
        ""
    )
    print("Examples: ", examples.values.tolist())

    with open(metadata_root.joinpath("article.md"), "r") as f:
        article = f.read()
    with open(metadata_root.joinpath("description.md"), "r") as f:
        description = f.read()

    demo = gr.Interface(
        fn=run_inference,
        title="Multitask Text and Chemistry T5",
        inputs=[
            gr.Dropdown(
                models,
                label="Language model",
                value="multitask-text-and-chemistry-t5-small-augm",
            ),
            gr.Radio(
                choices=[
                    "forward",
                    "retrosynthesis",
                    "paragraph to actions",
                    "molecular captioning",
                    "text-conditional de novo generation",
                ],
                label="Task",
                value="paragraph to actions",
            ),
            gr.Textbox(
                label="Text prompt",
                placeholder="I'm a stochastic parrot.",
                lines=1,
            ),
            gr.Slider(minimum=1, maximum=50, value=10, label="num_beams", step=1),
        ],
        outputs=gr.Textbox(label="Output"),
        article=article,
        description=description,
        examples=examples.values.tolist(),
    )
    demo.launch(debug=True, show_error=True)