import os
import sys

import streamlit as st
from huggingface_hub import snapshot_download
from transformers import AutoTokenizer


LOCAL_PATH = snapshot_download("flax-community/clip-spanish")
sys.path.append(LOCAL_PATH)

from modeling_hybrid_clip import FlaxHybridCLIP
from test_on_image import run_inference


def save_file_to_disk(uplaoded_file):
    temp_file = os.path.join("/tmp", uplaoded_file.name)
    with open(temp_file,"wb") as f:
        f.write(uploaded_file.getbuffer())
    return temp_file

# load the saved model
tokenizer = AutoTokenizer.from_pretrained("dccuchile/bert-base-spanish-wwm-cased")
model = FlaxHybridCLIP.from_pretrained(LOCAL_PATH)


st.title("Image-Caption Matching")
uploaded_file = st.file_uploader("Choose an image...", type="jpg")
text_input = st.text_input("Type a caption")

if uploaded_file is not None and text_input:
    local_image_path = None
    try:
        local_image_path = save_file_to_disk(uploaded_file)
        score = run_inference(local_image_path, text_input, model)[0][0].tolist()
        st.write('%s (%.2f%%)' % score)
    finally:
        if local_image_path:
            os.remove(local_image_path)