app.py

import torch
import re
import gradio as gr
from pathlib import Path
from transformers import AutoTokenizer, AutoFeatureExtractor, VisionEncoderDecoderModel
# Pattern to ignore all the text after 2 or more full stops
regex_pattern = "[.]{2,}"
#sample = val_dataset[800]
#model = model.cuda()
#print(tokenizer.decode(model.generate(sample['pixel_values'].unsqueeze(0).cuda())[0]).replace('<|endoftext|>', '').split('\n')[0],'\n\n\n')


def post_process(text):
    try:
        text = text.strip()
        text = re.split(regex_pattern, text)[0]
    except Exception as e:
        print(e)
        pass
    return text
def predict(image, max_length=64, num_beams=4):
    pixel_values = feature_extractor(images=image, return_tensors="pt").pixel_values
    pixel_values = pixel_values.to(device)
    with torch.no_grad():
        text = model.generate(pixel_values.unsqueeze(0).cpu())
        text = tokenizer.decode(text.replace('<|endoftext|>', '').split('\n')[0],'\n\n\n')
       # output_ids = model.generate(
        #    pixel_values,
        #    max_length=max_length,
         #   num_beams=num_beams,
         #   return_dict_in_generate=True,
        #).sequences
        
    #preds = tokenizer.batch_decode(output_ids, skip_special_tokens=True)
    #pred = post_process(preds[0])
    return text
    
model_path = "team-indain-image-caption/hindi-image-captioning"
device = torch.device("cuda:0" if torch.cuda.is_available() else"cpu")
# Load model.
model = VisionEncoderDecoderModel.from_pretrained(model_path)
model.to(device)
print("Loaded model")
feature_extractor = AutoFeatureExtractor.from_pretrained("google/vit-base-patch16-224-in21k")
print("Loaded feature_extractor")
tokenizer = AutoTokenizer.from_pretrained(model_path)
#if model.decoder.name_or_path == "gpt2":
 #   tokenizer.pad_token = tokenizer.bos_token
print("Loaded tokenizer")
title = "Hindi Image Captioning"
description = ""
input = gr.inputs.Image(type="pil")
#example_images = sorted([f.as_posix() for f in Path("examples").glob("*.jpg")])
#print(f"Loaded {len(example_images)} example images")
interface = gr.Interface(
    fn=predict,
    inputs=input,
    outputs="textbox",
    title=title,
    description=description,
    #examples=example_images,
    live=True,
  
)
interface.launch()