app.py

# -*- coding: utf-8 -*-
"""Untitled1.ipynb

Automatically generated by Colab.

Original file is located at
    https://colab.research.google.com/drive/1vA1O3q8yuzV8Hi3O8LhNuLGWS18yVEkb
"""

import streamlit as st
import PIL.Image
import base64
import time
import os
import torch
from transformers import AutoProcessor, LlavaForConditionalGeneration, BitsAndBytesConfig
from peft import PeftModel
from deep_translator import GoogleTranslator


@st.cache_resource
def load_model():
    model_id = "HuggingFaceH4/vsft-llava-1.5-7b-hf-trl"
    quantization_config = BitsAndBytesConfig(load_in_4bit=True)
    base_model = LlavaForConditionalGeneration.from_pretrained(model_id, quantization_config=quantization_config, torch_dtype=torch.float16)

    # Load the PEFT Lora adapter
    peft_lora_adapter_path = "Praveen0309/llava-1.5-7b-hf-ft-mix-vsft-3"
    peft_lora_adapter = PeftModel.from_pretrained(base_model, peft_lora_adapter_path, adapter_name="lora_adapter")
    base_model.load_adapter(peft_lora_adapter_path, adapter_name="lora_adapter")

    processor = AutoProcessor.from_pretrained("HuggingFaceH4/vsft-llava-1.5-7b-hf-trl")

    return base_model, processor

base_model, processor = load_model()

# Function to translate text from Bengali to English
def deep_translator_bn_en(input_sentence):
    english_translation = GoogleTranslator(source="bn", target="en").translate(input_sentence)
    return english_translation

# Function to translate text from English to Bengali
def deep_translator_en_bn(input_sentence):
    bengali_translation = GoogleTranslator(source="en", target="bn").translate(input_sentence)
    return bengali_translation

def inference(image, image_prompt):
    prompt = f"USER: <image>\n{image_prompt} ASSISTANT:"

    # Assuming your model can handle PIL images
    image = image.convert("RGB")  # Ensure image is RGB mode

    inputs = processor(text=prompt, images=image, return_tensors="pt")
    generate_ids = base_model.generate(**inputs, max_new_tokens=1024)
    decoded_response = processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
    return decoded_response

def image_to_base64(image_path):
    with open(image_path, 'rb') as img:
        encoded_string = base64.b64encode(img.read())
    return encoded_string.decode('utf-8')

# Function that takes User Inputs and displays it on ChatUI
def query_message(history,txt,img):
    image_prompt = deep_translator_bn_en(txt)
    history += [(image_prompt,None)]
    base64 = image_to_base64(img)
    data_url = f"data:image/jpeg;base64,{base64}"
    history += [(f"{image_prompt} ![]({data_url})", None)]
    return history

# Function that takes User Inputs, generates Response and displays on Chat UI
def llm_response(history,text,img):
        image_prompt = deep_translator_bn_en(text)
        response = inference(img,image_prompt)
        assistant_index = response.find("ASSISTANT:")
        extracted_string = response[assistant_index + len("ASSISTANT:"):].strip()
        output = deep_translator_en_bn(extracted_string)
        history += [(text,output)]
        return history

# Interface Code
st.title('My_BoT')

# Create a sidebar
sidebar = st.sidebar
sidebar.header('User Inputs')

# Add a file uploader to the sidebar
uploaded_file = sidebar.file_uploader("Upload an Image", type=['png', 'jpg', 'jpeg'])

# Add a text input to the sidebar
text_input = sidebar.text_input("Enter text and press enter")

# Initialize session state for history if it doesn't exist
if 'history' not in st.session_state:
    st.session_state.history = []

# Check if text is entered and no image is uploaded
if text_input and uploaded_file is None:
    st.write("Please upload an image.")

# Add a button to the sidebar
submit_button = sidebar.button("Submit")

# When the button is clicked, generate the response and display it
if submit_button:
    if uploaded_file is not None:
        image = PIL.Image.open(uploaded_file)
        st.session_state.history = llm_response(st.session_state.history, text_input, image)
        for text, output in st.session_state.history:
            st.write(f"User: {text}")
            if output is not None:
                st.write(f"Assistant: {output}")