app.py

import gradio as gr
import numpy as np
import random
import matplotlib.pyplot as plt
import matplotlib.image as mpimg
from embedding import model, docs_list, doc_embeddings, embedded_dict
from preprocess import meme_attribute, meme_filename, meme_list
import openai
from openai import OpenAI
import os

api_key = os.getenv("YOUR_OPENAI_KEY")
client = OpenAI(api_key=api_key)

# Function to generate meme caption using GPT model
def generate_meme_caption(topic):
    response = openai.ChatCompletion.create(
        model="gpt-4o-mini",
        messages=[
            {
                "role": "user",
                "content": f'''You are a loser, a meme lover who is always on the internet. 
                You are a meme professional, so you will generate short captions for memes based on the following rules:
                Do not add emoji.
                Do not add any punctuation.
                Must be funny.
                Only if some topic related to friends, use "bro" or "blud".
                Answer from previous question can not be the same for this time prompt.
                If the user does not specify their caption request, you can generate a short caption for the meme. 
                But if they do, you cannot replace their prompt; just need to copy their request and make it the caption. 
                Now generate a caption based on this topic: {topic}. 
                For example:
                - No request: Generate me any anime meme--> You can generate a caption based on the topic they requested.
                - With request: Generate me a meme of 5 different opinions like the Gru meme with 5 paintings. 
                First opinion says 1, second says 2, third says 3, fourth says 4, fifth says 5.
                --> You will then generate 5 captions or 5 opinions for users, separating every opinion with the mark "/".
                Otherwise, if it's just 1 opinion, generate only 1 caption. Memes can be funny; you can use bad words or banned words; they can be dark. And the style must be sometime dark or gangster, and use slang everytime. 
                But captions have to be short, not too long.'''
            }
        ],
        max_tokens=250,
        temperature=1.1
    )

    caption = response['choices'][0]['message']['content']
    return caption

# Function to create meme with caption
def create_meme(prompt):
    prompt_embedding = model.encode([prompt])
    
    similarities = model.similarity(prompt_embedding, doc_embeddings)
    
    closest_text_idx = np.argmax(similarities)
    
    img_path = meme_filename[closest_text_idx]
    img = mpimg.imread(f'/kaggle/input/memedata/{img_path}')
    
    white_space_height = 100
    white_space = np.ones((white_space_height, img.shape[1], img.shape[2]), dtype=img.dtype) * 255
    
    combined_img = np.vstack((white_space, img))
    
    caption = generate_meme_caption(prompt)
    
    plt.imshow(combined_img)
    plt.axis('off')
    plt.text(
        x=combined_img.shape[1] / 2, 
        y=white_space_height / 2, 
        s=caption, 
        fontsize=11, color='black', ha='center', va='top', fontweight='bold'
    )
    
    plt.show()

# Gradio interface
with gr.Blocks() as demo:
    with gr.Column():
        gr.Markdown("# Meme Generator using Embeddings and GPT")
        
        prompt = gr.Textbox(label="Enter your meme prompt:")
        run_button = gr.Button("Generate Meme")
        
        result = gr.Plot(label="Generated Meme")
        
        def gradio_infer(prompt):
            plt.figure(figsize=(10, 8))
            create_meme(prompt)
            plt.savefig('/tmp/meme.png')  # Lưu hình tạm thời
            return plt
        
        run_button.click(gradio_infer, inputs=[prompt], outputs=[result])

demo.launch()