app.py

import torch 
import re 
import gradio as gr
import streamlit as st
# st.title("Image Caption Generator")
from transformers import AutoTokenizer, ViTFeatureExtractor, VisionEncoderDecoderModel 
import os
import tensorflow as tf
os.environ['TF_ENABLE_ONEDNN_OPTS'] = '0'

device='cpu'
encoder_checkpoint = "nlpconnect/vit-gpt2-image-captioning"
decoder_checkpoint = "nlpconnect/vit-gpt2-image-captioning"
model_checkpoint = "nlpconnect/vit-gpt2-image-captioning"

feature_extractor = ViTFeatureExtractor.from_pretrained(encoder_checkpoint)
tokenizer = AutoTokenizer.from_pretrained(decoder_checkpoint)
model = VisionEncoderDecoderModel.from_pretrained(model_checkpoint).to(device)


def predict(image, max_length=64, num_beams=4):
  image = image.convert('RGB')
  image = feature_extractor(image, return_tensors="pt").pixel_values.to(device)
  clean_text = lambda x: x.replace('<|endoftext|>','').split('\n')[0]
  caption_ids = model.generate(image, max_length = max_length)[0]
  caption_text = clean_text(tokenizer.decode(caption_ids))
  return caption_text 

input = gr.inputs.Image(label="Upload any Image", type = 'pil', optional=True)
output = gr.outputs.Textbox(type="text",label="Captions")
examples = ["example1.jpg"]
print("------------------------- 6 -------------------------\n")
title = "Image to Text ViT with LORA"

# interface = gr.Interface(
            
#         fn=predict,
#         description=description,
#         inputs = input,
#         theme="grass",
#         outputs=output,
#         examples=examples,
#         title=title,
#     )
# interface.launch(debug=True)

with gr.Blocks() as demo:
    

    gr.HTML(
        """
        <div style="text-align: center; max-width: 1200px; margin: 20px auto;">
        <h1 style="font-weight: 900; font-size: 3rem; margin: 0rem">
            ViT Image-to-Text with LORA
        </h1>   
        <h2 style="text-align: left; font-weight: 450; font-size: 1rem; margin-top: 2rem; margin-bottom: 1.5rem">
        In the field of large language models, the challenge of fine-tuning has long perplexed researchers. Microsoft, however, has unveiled an innovative solution called <b>Low-Rank Adaptation (LoRA)</b>. With the emergence of behemoth models like GPT-3 boasting billions of parameters, the cost of fine-tuning them for specific tasks or domains has become exorbitant.
        LoRA offers a groundbreaking approach by freezing the weights of pre-trained models and introducing trainable layers known as <b>rank-decomposition matrices in each transformer block</b>. This ingenious technique significantly reduces the number of trainable parameters and minimizes GPU memory requirements, as gradients no longer need to be computed for the majority of model weights.        
        <br>
        <br>
        You can find more info here: <a href="https://www.linkedin.com/pulse/fine-tuning-image-to-text-algorithms-with-lora-daniel-puente-viejo" target="_blank";>Linkedin article</a>
        </h2>
        
        </div>
        """)
    with gr.Row():
            with gr.Column(scale=1):
                img = gr.inputs.Image(label="Upload any Image", type = 'pil', optional=True)
                button = gr.Button(value="Describe")
            with gr.Column(scale=1):
                out = gr.outputs.Textbox(type="text",label="Captions")   
                
    button.click(predict, inputs=[img], outputs=[out])
    gr.Examples(
        examples=[os.path.join(os.path.dirname(__file__), "example1.jpg")],
        inputs=img,
        outputs=out,
        fn=predict,
        cache_examples=True,
    )
demo.launch(debug=True)