app.py

import gradio as gr
from transformers import AutoProcessor, Blip2ForConditionalGeneration
import torch
from PIL import Image

# Check for GPU availability and set the device variable accordingly
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

# Load the BLIP-2 model and processor
processor = AutoProcessor.from_pretrained("Salesforce/blip2-opt-2.7b")

# Load model with additional parameters if GPU is available, else load without additional parameters
if torch.cuda.is_available():
    device_map = {0: 'cpu', 1: 'cpu'}  # Define a custom device map if needed
    model = Blip2ForConditionalGeneration.from_pretrained(
        "Salesforce/blip2-opt-2.7b",
        load_in_8bit=True,
        device_map=device_map,
        load_in_8bit_fp32_cpu_offload=True
    )
else:
    model = Blip2ForConditionalGeneration.from_pretrained("Salesforce/blip2-opt-2.7b")


def blip2_interface(image, prompted_caption_text, vqa_question, chat_context):
    # Prepare image input
    image_input = Image.fromarray(image).convert('RGB')
    inputs = processor(image_input, return_tensors="pt").to(device)  # Remove torch.float16 dtype conversion
    
    
    # Image Captioning
    generated_ids = model.generate(**inputs, max_new_tokens=20)
    image_caption = processor.batch_decode(generated_ids, skip_special_tokens=True)[0].strip()

    # Prompted Image Captioning
    inputs = processor(image_input, text=prompted_caption_text, return_tensors="pt").to(device)
    generated_ids = model.generate(**inputs, max_new_tokens=20)
    prompted_caption = processor.batch_decode(generated_ids, skip_special_tokens=True)[0].strip()
    
    # Visual Question Answering (VQA)
    prompt = f"Question: {vqa_question} Answer:"
    inputs = processor(image_input, text=prompt, return_tensors="pt").to(device)
    generated_ids = model.generate(**inputs, max_new_tokens=10)
    vqa_answer = processor.batch_decode(generated_ids, skip_special_tokens=True)[0].strip()
    
    # Chat-based Prompting
    prompt = chat_context + " Answer:"
    inputs = processor(image_input, text=prompt, return_tensors="pt").to(device)
    generated_ids = model.generate(**inputs, max_new_tokens=10)
    chat_response = processor.batch_decode(generated_ids, skip_special_tokens=True)[0].strip()

    return image_caption, prompted_caption, vqa_answer, chat_response

# Define Gradio input components
image_input = gr.inputs.Image(type="numpy", label="Image Input")
prompted_caption_input = gr.inputs.Textbox(label="Prompted Caption Text")
vqa_question_input = gr.inputs.Textbox(label="VQA Question")
chat_context = gr.inputs.Textbox(label="Chat Context")

# Define Gradio output components with labels corresponding to the inputs
image_caption_result = gr.outputs.Textbox(label="Image Caption")
prompted_caption_result = gr.outputs.Textbox(label="Prompted Image Caption")
vqa_answer = gr.outputs.Textbox(label="VQA Answer")
chat_response = gr.outputs.Textbox(label="Chat Response")

# Create Gradio interface
iface = gr.Interface(
    fn=blip2_interface,
    inputs=[image_input, prompted_caption_input, vqa_question_input, chat_context],
    outputs=[image_caption_result, prompted_caption_result, vqa_answer, chat_response],
    title="BLIP-2 Image Captioning and VQA",
    description="Interact with the BLIP-2 model for image captioning, prompted image captioning, visual question answering, and chat-based prompting.",
)

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