app.py

### Imports and class names setup ---------------------------------------------------- ###
import os
import torch
import torchvision
import gradio as gr

from model import create_vit
from timeit import default_timer as timer
from typing import Tuple, Dict

# Setup class names
with open("class_names.txt", "r") as f:
  class_names = [food.strip() for food in f.readlines()]

# Device agnostic code
if torch.backends.mps.is_available():
  device = 'mps'
elif torch.cuda.is_available():
  device = 'cuda'
else:
  device = 'cpu'

### Model and transforms preparation ---------------------------------------------------- ###
vit_model, vit_transforms = create_vit(pretrained_weights=torchvision.models.ViT_B_16_Weights.DEFAULT,
                                       model=torchvision.models.vit_b_16,
                                       in_features=768,
                                       out_features=101,
                                       device='cpu')

# Load save weights
vit_model.load_state_dict(torch.load(f="pretrained_vit_feature_extractor_food101.pth",
                          map_location=torch.device("cpu"))) # load the model to the CPU

### Predict function ---------------------------------------------------- ###
def predict(img) -> Tuple[Dict, float]:
  # Start a timer
  start_time = timer()
  # Transform the input image for use with ViT Model
  img = vit_transforms(img).unsqueeze(0) # unsqueeze = add batch dimension on 0th index (3, 224, 224) into (1, 3, 224, 224)
  # Put model into eval mode, make prediction
  vit_model.eval()
  with torch.inference_mode():
    # Pass transformed image through the model and turn the prediction logits into probabilities
    pred_logits = vit_model(img)
    pred_probs = torch.softmax(pred_logits, dim=1)
  # Create a prediction label and prediction probability dictionary
  pred_labels_and_probs = {class_names[i]: float(pred_probs[0][i]) for i in range(len(class_names))}
  
  # Calculate pred time
  end_timer = timer()
  pred_time = round(end_timer - start_time, 4)
  
  # Return pred dict and pred time
  return pred_labels_and_probs, pred_time

### Gradio interface and launch ------------------------------------------------------------------ ###

# Create title and description
title = "FoodVision: ViT Model"
description = "A ViT model trained on 20% of the Food101 dataset to classify Food images"

# Create example list
example_list = [["examples/" + example] for example in os.listdir("examples")]

# Create the Gradio demo
demo = gr.Interface(fn=predict, inputs=gr.Image(type="pil"), outputs=[gr.Label(num_top_classes=5, label="Predictions"),
                                                         gr.Number(label="Prediction time(s)")], title=title, description=description, examples=example_list)
demo.launch()