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"
print("------------------------- 1 -------------------------\n")
feature_extractor = ViTFeatureExtractor.from_pretrained(encoder_checkpoint)
print("------------------------- 2 -------------------------\n")
tokenizer = AutoTokenizer.from_pretrained(decoder_checkpoint)
print("------------------------- 3 -------------------------\n")
model = VisionEncoderDecoderModel.from_pretrained(model_checkpoint).to(device)
print("------------------------- 4 -------------------------\n")


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 

print("------------------------- 5 -------------------------\n")
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">
            TextDiffuser: Diffusion Models as Text Painters
        </h1>   
        <h2 style="text-align: left; font-weight: 450; font-size: 1rem; margin-top: 0.5rem; margin-bottom: 0.5rem">
        We propose <b>Image to Text</b>, with ViT model but with LORA fine-tuning. 
        </h2>
        </div>
        """)
    gr.inputs.Image(label="Upload any Image", type = 'pil', optional=True)
    gr.outputs.Textbox(type="text",label="Captions")
    # gr.Image(label="Upload any Image", type = 'pil', optional=True)
    # gr.Textbox(type="text",label="Captions")
    

demo.launch(debug=True)