my_model/tabs/run_inference.py

import streamlit as st
import torch
import bitsandbytes
import accelerate
import scipy
import copy
import time
from PIL import Image
import torch.nn as nn
import pandas as pd
from my_model.object_detection import detect_and_draw_objects
from my_model.captioner.image_captioning import get_caption
from my_model.utilities.gen_utilities import free_gpu_resources
from my_model.state_manager import StateManager
from my_model.config import inference_config as config


class InferenceRunner(StateManager):

    """
    InferenceRunner manages the user interface and interactions for a Streamlit-based
    Knowledge-Based Visual Question Answering (KBVQA) application. It handles image uploads,
    displays sample images, and facilitates the question-answering process using the KBVQA model.
    it inherits the StateManager class.
    """
    
    def __init__(self):
        """
        Initializes the InferenceRunner instance, setting up the necessary state.
        """
        
        super().__init__()


    def answer_question(self, caption, detected_objects_str, question):
        """
        Generates an answer to a given question based on the image's caption and detected objects.

        Args:
            caption (str): The caption generated for the image.
            detected_objects_str (str): String representation of objects detected in the image.
            question (str): The user's question about the image.
        Returns:
            str: The generated answer to the question.
        """
        free_gpu_resources()
        answer = st.session_state.kbvqa.generate_answer(question, caption, detected_objects_str)
        prompt_length  = st.session_state.kbvqa.current_prompt_length
        free_gpu_resources()
        return answer, prompt_length


    def display_sample_images(self):
        self.col1.write("Choose from sample images:")
        cols = self.col1.columns(len(config.SAMPLE_IMAGES))
        for idx, sample_image_path in enumerate(config.SAMPLE_IMAGES):
            with cols[idx]:
                image = Image.open(sample_image_path)
                image_for_display = self.resize_image(sample_image_path, 80, 80)
                st.image(image_for_display)
                if st.button(f'Select Sample Image {idx + 1}', key=f'sample_{idx}'):
                    self.process_new_image(sample_image_path, image)

    def handle_image_upload(self):
        uploaded_image = self.col1.file_uploader("Or upload an Image", type=["png", "jpg", "jpeg"])
        if uploaded_image is not None:
            self.process_new_image(uploaded_image.name, Image.open(uploaded_image))

    def display_image_and_analysis(self, image_key, image_data, nested_col21, nested_col22):
        
        image_for_display = self.resize_image(image_data['image'], 600)
        nested_col21.image(image_for_display, caption=f'Uploaded Image: {image_key[-11:]}')
        self.handle_analysis_button(image_key, image_data, nested_col22)

    def handle_analysis_button(self, image_key, image_data, nested_col22):
        if not image_data['analysis_done'] or self.settings_changed or self.confidance_change:
            nested_col22.text("Please click 'Analyze Image'..")
            analyze_button_key = f'analyze_{image_key}_{st.session_state.detection_model}_{st.session_state.confidence_level}'
            if nested_col22.button('Analyze Image', key=analyze_button_key, on_click=self.disable_widgets, disabled=self.is_widget_disabled):
                caption, detected_objects_str, image_with_boxes = self.analyze_image(image_data['image'])
                self.update_image_data(image_key, caption, detected_objects_str, True)
            st.session_state['loading_in_progress'] = False

    def handle_question_answering(self, image_key, image_data, nested_col22):
        # Initialize qa_history for each image
        #qa_history = image_data.get('qa_history', [])
        if image_data['analysis_done']:
            self.display_question_answering_interface(image_key, image_data, nested_col22)

        if self.settings_changed or self.confidance_change:
            nested_col22.warning("Confidence level changed, please click 'Analyze Image'.")

    def display_question_answering_interface(self, image_key, image_data, nested_col22):
        
        sample_questions = config.SAMPLE_QUESTIONS.get(image_key, [])
        selected_question = nested_col22.selectbox("Select a sample question or type your own:", ["Custom question..."] + sample_questions, key=f'sample_question_{image_key}')
        custom_question = nested_col22.text_input("Or ask your own question:", key=f'custom_question_{image_key}')
        question = custom_question if selected_question == "Custom question..." else selected_question
        self.process_question(image_key, question, image_data, nested_col22)

        qa_history = image_data.get('qa_history', [])
        for num, (q, a, p) in enumerate(qa_history):
            nested_col22.text(f"Q{num+1}: {q}\nA{num+1}: {a}\nPrompt Length: {p}\n")

            
    def process_question(self, image_key, question, image_data, nested_col22):
        qa_history = image_data.get('qa_history', [])
        if question and (question not in [q for q, _, _ in qa_history] or self.settings_changed or self.confidance_change):
            if nested_col22.button('Get Answer', key=f'answer_{image_key}', disabled=self.is_widget_disabled):
                answer, prompt_length = self.answer_question(image_data['caption'], image_data['detected_objects_str'], question)
                self.add_to_qa_history(image_key, question, answer, prompt_length)
               # nested_col22.text(f"Q: {question}\nA: {answer}\nPrompt Length: {prompt_length}")

    def image_qa_app(self):
        self.display_sample_images()
        self.handle_image_upload()
        self.display_session_state()
        with self.col2:
            for image_key, image_data in self.get_images_data().items():
                with st.container():
                    nested_col21, nested_col22 = st.columns([0.65, 0.35])
                    self.display_image_and_analysis(image_key, image_data, nested_col21, nested_col22)
                    self.handle_question_answering(image_key, image_data, nested_col22)
                    

        
    def run_inference(self):
        """
        Sets up the widgets and manages the inference process. This method handles model loading,
        reloading, and the overall flow of the inference process based on user interactions.

        """
        
        self.set_up_widgets()
  
        load_fine_tuned_model = False
        fine_tuned_model_already_loaded = False
        reload_detection_model = False
        force_reload_full_model = False
        
     
        if self.is_model_loaded and self.settings_changed:
            self.col1.warning("Model settings have changed, please reload the model, this will take a second .. ")
            self.update_prev_state()
         
           
        st.session_state.button_label = "Reload Model" if self.is_model_loaded and st.session_state.kbvqa.detection_model != st.session_state['detection_model'] else "Load Model"
      
        with self.col1:
            if st.session_state.method == "Fine-Tuned Model":
                with st.container():
                    nested_col11, nested_col12 = st.columns([0.5, 0.5])
                    if nested_col11.button(st.session_state.button_label, on_click=self.disable_widgets, disabled=self.is_widget_disabled):
                        if st.session_state.button_label == "Load Model":
                            if self.is_model_loaded:
                                free_gpu_resources()
                                fine_tuned_model_already_loaded = True
                            else:
                                load_fine_tuned_model = True
                        else:
                            reload_detection_model = True
                    if nested_col12.button("Force Reload", on_click=self.disable_widgets, disabled=self.is_widget_disabled):
                        force_reload_full_model = True
                        

                if load_fine_tuned_model:
                    t1=time.time()
                    free_gpu_resources()
                    self.load_model()
                    st.session_state['time_taken_to_load_model'] = int(time.time()-t1)
                    st.session_state['loading_in_progress'] = False
                    
                elif fine_tuned_model_already_loaded:
                    free_gpu_resources()
                    self.col1.text("Model already loaded and no settings were changed:)")
                    st.session_state['loading_in_progress'] = False
                    
                elif reload_detection_model:
                    free_gpu_resources()
                    self.reload_detection_model()
                    st.session_state['loading_in_progress'] = False
                    
                elif force_reload_full_model:
                    free_gpu_resources()
                    t1=time.time()
                    self.force_reload_model()
                    st.session_state['time_taken_to_load_model'] = int(time.time()-t1)
                    st.session_state['loading_in_progress'] = False
                    st.session_state['model_loaded'] = True
                    
            elif st.session_state.method == "In-Context Learning (n-shots)":
                self.col1.warning(f'Model using {st.session_state.method} is not deployed yet, will be ready later.')
                st.session_state['loading_in_progress'] = False
        
        if self.is_model_loaded:
            free_gpu_resources()
            st.session_state['loading_in_progress'] = False
            self.image_qa_app()