app.py

import gradio as gr
import torch
import torch.nn.functional as F
from transformers import Blip2Processor, Blip2ForConditionalGeneration
from PIL import Image
from peft import LoraConfig, get_peft_model


# Initialize the processor and model
processor = Blip2Processor.from_pretrained("Salesforce/blip2-flan-t5-xl")
# model_path = "full-blip2-deit-config-yes-no-2.pth"
# model = torch.load("./full-blip2-deit-config-2.pth")
# model = torch.load("./full-blip2-deit.pth") # not working - error
# model = torch.load("./full-blip2-deit-config-free-form-4-ver-2.pth")
model = torch.load("./full_config_blip2-deit-05")

model.eval()  # Set the model to evaluation mode
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.to(device)

def preprocess_image(image):
    """Preprocess the image to match the model's input requirements."""
    # Convert PIL image to tensor
    pixel_values = processor(images=image, return_tensors="pt").pixel_values.to(device)

    # Apply specific model's preprocessing
    patch_embeddings = model.vision_model.embeddings.patch_embeddings.projection(pixel_values)
    patch_embeddings_flat = patch_embeddings.view(1, -1, 1408)

    cls_token = model.vision_model.embeddings.cls_token.expand(1, -1, -1)
    dist_token = model.vision_model.embeddings.distillation_token.expand(1, -1, -1)
    full_embeddings = torch.cat([cls_token, dist_token, patch_embeddings_flat], dim=1)

    encoder_outputs = model.vision_model.encoder(full_embeddings)
    image_outputs = encoder_outputs.last_hidden_state

    image_outputs = F.adaptive_avg_pool2d(image_outputs, (3, 50176))
    image_outputs = image_outputs.view(1, 3, 224, 224)  # Adjusted dimensions
    return image_outputs

def generate_answer_blip2(image, question):
    """Generate answers based on an image and a question using a BLIP2 model."""
    image_outputs = preprocess_image(image)
    
    # Prepare question
    question_formatted = "Question: " + question + " Answer:"
    inputs = processor(text=question_formatted, return_tensors="pt")
    inputs['pixel_values'] = image_outputs.to(device)  # Ensure image tensor is on the correct device
    
    # Generate response using the model
    generated_ids = model.generate(**inputs, max_length=50)
    generated_answer = processor.batch_decode(generated_ids, skip_special_tokens=True)

    return generated_answer[0]  # Return the first (and typically only) generated answer

# Function to display the demo interface
def show_demo():
    return (
        gr.update(visible=True),
        gr.update(visible=True),
        gr.update(visible=True),
        gr.update(visible=True),
        gr.update(visible=True)
    )

# Setting up the Gradio interface with Blocks
with gr.Blocks() as landing_page:
    gr.Markdown("# Welcome to the Visual Question Answering Demo")
    gr.Markdown("This demo uses the customized BLIP2 model to answer questions about images.")
    gr.Markdown("### How to Use: ")
    gr.Markdown("1. Upload an image. \n2. Enter a question related to the image. \n3. Receive the generated answer.")
    gr.Markdown("### Model Information: ")
    gr.Markdown("The BLIP2 model combines vision and language understanding to generate answers based on the provided image and question.")
    
    with gr.Column() as demo_column:
        start_demo_button = gr.Button("Start Demo")
        image_input = gr.Image(label="Upload Image", visible=False)
        question_input = gr.Textbox(label="Enter your question", visible=False)
        submit_button = gr.Button("Submit", visible=False)
        clear_button = gr.Button("Clear", visible=False)
        answer_output = gr.Textbox(label="Generated Answer", visible=False)

    start_demo_button.click(fn=show_demo, inputs=None, outputs=[image_input, question_input, submit_button, clear_button, answer_output])

    def generate_and_show_answer(image, question):
        return generate_answer_blip2(image, question)

    submit_button.click(fn=generate_and_show_answer, inputs=[image_input, question_input], outputs=answer_output)
    clear_button.click(fn=lambda: (None, "", "", ""), inputs=None, outputs=[image_input, question_input, answer_output, answer_output])

if __name__ == "__main__":
    landing_page.launch()