app.py

import gradio as gr
from model import create_vit_instance
from pathlib import Path
import torch
import torchvision
from PIL import Image
from typing import List, Dict, Tuple
from timeit import default_timer as timer


# Reading all available classes
with open('class_names.txt', 'r') as f:
  all_classes = [name.replace('\n', '') for name in f.readlines()]

demo_vit_model, demo_vit_transforms = create_vit_instance(num_classes=len(all_classes),
                                                          device='cpu')

weights_path = Path("ViT_Caltech101_five_epochs.pth")
demo_vit_model.load_state_dict(torch.load(f=weights_path,
                                          map_location='cpu'))


## Creating predict method => It returns prediction probability dictionary as well as time taken to do the prediction
def predict(img_path: str,
            model:torch.nn.Module=demo_vit_model,
            transform: torchvision.transforms=demo_vit_transforms,
            classes:List[str] = all_classes)->Tuple[Dict, int]:

  pred_prob_dict = dict()
  model = model.to('cpu')
  # img_path = Image.open(img_path)
  transformed_image = transform(img_path)

  start = timer()
  model.eval()
  with torch.inference_mode():

    batch_img = transformed_image.unsqueeze(dim=0).to(device='cpu')
    logit = model(batch_img)
    pred_probs = torch.softmax(input=logit,
                               dim=1)
    preds = torch.argmax(input=pred_probs,
                         dim=1).item()
  end = timer()

  total_time = round(end - start, 4)
  pred_probs = pred_probs[0].tolist()

  for idx in range(len(pred_probs)):
    class_name = classes[idx]
    pred_prob_dict[class_name] = pred_probs[idx]

  sorted_order = sorted(pred_prob_dict.items(), key=lambda kv: kv[1], reverse=True)

  return (pred_prob_dict, total_time)



title = "ObjectVision"
description = "ViT Feature Extractor trained for Image Classification based on Caltech101 dataset."
samples = [[path] for path in Path("examples").iterdir()]
demo = gr.Interface(fn=predict,
                    title=title,
                    description=description,
                    inputs=gr.Image(type="pil"),
                    examples=samples,
                    outputs=[
                        gr.Label(num_top_classes=5,
                                 label="Model thinks"),
                        gr.Number(label="Prediction time (in seconds)")
                    ])

if __name__ == "__main__":
  demo.launch(debug=True)