app.py

# Imports
import torch
import gradio as gr
from typing import Tuple, Dict
from timeit import default_timer as timer
import model
import os

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

# Create instance of model
effnetb2, effnetb2_transforms = model.create_effnetb2_model(num_classes=len(class_names))
# Load Weights
effnetb2.load_state_dict(state_dict=torch.load("effnetb2_food101_20pct.pth",
                                               map_location=torch.device("cpu") # hard-coded load to cpu
                                               ))

# Predict function
def predict(img) -> Tuple[Dict, float]:
  # Start timer
  start = timer()
  # Transform input image for use
  img = effnetb2_transforms(img).unsqueeze(dim=0)
  # Put model in eval mode
  effnetb2.eval()
  with torch.inference_mode():
    logits = effnetb2(img)
    pred_probs = torch.softmax(logits, dim=1).squeeze()
    prediction = logits.argmax(dim=1).item()
    prediction_label = class_names[prediction]
  end = timer()
  pred_dict = {class_names[i]: pred_probs[i].item() for i in range(len(class_names))}
  delta_time = round(end-start, 4)
  return pred_dict, delta_time

# Gradio app

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

title = "FoodVision Big"
description = "EfficientNetB2 feature extractor CV model to classify images of 101 types of food from the Food101 dataset."
article = "Created for PyTorch ZTM course"

demo = gr.Interface(fn=predict, # maps inputs to outputs
                    inputs=gr.Image(type="pil"),
                    outputs=[gr.Label(num_top_classes=5, label="Predictions"),  # for the prediction dictionary
                             gr.Number(label="Prediction time (s)")
                             ],
                    examples=example_list,
                    title=title,
                    description=description,
                    article=article,
                    )

demo.launch()