app.py

import os.path

import gdown
import gradio as gr
import torch


from Model import TRCaptionNet, clip_transform



device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# device = "cpu"

preprocess_tasviret = clip_transform(336)
model_tasviret = TRCaptionNet({
    "max_length": 35,
    "clip": "ViT-L/14@336px",
    "bert": "dbmdz/bert-base-turkish-cased",
    "proj": True,
    "proj_num_head": 16
})
model_ckpt = "./checkpoints/TRCaptionNet-TasvirEt_L14_334_berturk.pth"
model_tasviret.load_state_dict(torch.load(model_ckpt, map_location=device)["model"], strict=True)
model_tasviret = model_tasviret.to(device)
model_tasviret.eval()

preprocess = clip_transform(224)
model = TRCaptionNet({
    "max_length": 35,
    "clip": "ViT-L/14",
    "bert": "dbmdz/bert-base-turkish-cased",
    "proj": True,
    "proj_num_head": 16
})
model_ckpt = "./checkpoints/TRCaptionNet_L14_berturk.pth"
model.load_state_dict(torch.load(model_ckpt, map_location=device)["model"], strict=True)
model = model.to(device)
model.eval()



def inference(raw_image, min_length, repetition_penalty):
    batch = preprocess_tasviret(raw_image).unsqueeze(0).to(device)
    caption_tasviret = model_tasviret.generate(batch, min_length=min_length, repetition_penalty=repetition_penalty)[0]
    
    batch = preprocess(raw_image).unsqueeze(0).to(device)
    caption = model.generate(batch, min_length=min_length, repetition_penalty=repetition_penalty)[0]
    
    return [caption, caption_tasviret]


inputs = [gr.Image(type='pil', interactive=True,),
          gr.Slider(minimum=4, maximum=22, value=8, label="MINIMUM CAPTION LENGTH", step=1),
          gr.Slider(minimum=1, maximum=2, value=1.6, label="REPETITION PENALTY")]
          
outputs = [gr.components.Textbox(label="Caption"), gr.components.Textbox(label="Caption-TasvirEt")]
title = "TRCaptionNet-TasvirEt"
paper_link = ""
github_link = "https://github.com/serdaryildiz/TRCaptionNet"
IEEE_link = "https://github.com/serdaryildiz/TRCaptionNet"

description = f"<p style='text-align: center'><a href='{IEEE_link}' target='_blank'> SIU2024: Turkish Image Captioning with Vision Transformer Based Encoders and Text Decoders</a> "
description += f"<p style='text-align: center'><a href='{github_link}' target='_blank'>TRCaptionNet</a> : A novel and accurate deep Turkish image captioning model with vision transformer based image encoders and deep linguistic text decoders"

examples = [
    ["images/test1.jpg"],
    ["images/test2.jpg"],
    ["images/test3.jpg"],
    ["images/test4.jpg"],
    ["images/test5.jpg"],
    ["images/test6.jpg"],
    ["images/test7.jpg"],
    ["images/test8.jpg"],
    ["images/test9.jpg"],
    ["images/test10.jpg"],
    ["images/test11.jpg"],
]
article = f"<p style='text-align: center'><a href='{paper_link}' target='_blank'>Paper</a> | <a href='{github_link}' target='_blank'>Github Repo</a></p>"
css = ".output-image, .input-image, .image-preview {height: 600px !important}"

iface = gr.Interface(fn=inference,
                     inputs=inputs,
                     outputs=outputs,
                     title=title,
                     description=description,
                     examples=examples,
                     article=article,
                     css=css)
iface.launch()