vision.py

import math

import streamlit as st
from transformers import (
    PreTrainedTokenizerFast,
    VisionEncoderDecoderModel,
    ViTImageProcessor,
)


MODEL_NAME = "grascii/gregg-vision-v0.2.1"
MIN_LOG_PROB = math.log(0.5)
NUM_BEAMS = 3


@st.cache_resource(show_spinner=f"Loading {MODEL_NAME}")
def load_model():
    model = VisionEncoderDecoderModel.from_pretrained(
        MODEL_NAME, token=st.secrets.HF_TOKEN
    )
    tokenizer = PreTrainedTokenizerFast.from_pretrained(
        MODEL_NAME,
        token=st.secrets.HF_TOKEN,
    )
    processor = ViTImageProcessor.from_pretrained(MODEL_NAME, token=st.secrets.HF_TOKEN)
    return model, tokenizer, processor


@st.cache_data(ttl=3600, show_spinner=f"Running {MODEL_NAME}")
def run_vision(image):
    model, tokenizer, processor = load_model()
    pixel_values = processor(image, return_tensors="pt").pixel_values
    generated = model.generate(
        pixel_values,
        max_new_tokens=12,
        num_beams=NUM_BEAMS,
        num_return_sequences=NUM_BEAMS,
        output_scores=True,
        return_dict_in_generate=True,
    )
    return [
        tokenizer.convert_ids_to_tokens(
            generated["sequences"][0], skip_special_tokens=True
        )
    ] + [
        tokenizer.convert_ids_to_tokens(seq, skip_special_tokens=True)
        for seq, score in zip(
            generated["sequences"][1:], generated["sequences_scores"][1:]
        )
        if score > MIN_LOG_PROB
    ]