app.py

from timeit import default_timer as timer
import gradio as gr
import torch
from model import create_effnetb2_model
from classes import class_names

# Setup DataLoaders
effnetb2, effnetb2_transforms = create_effnetb2_model(num_classes=525,
                                                      seed=42)

# Load the Best Model (effnetb2 with 10 epochs)

def load_best_effnetb2_model(model_path):

    # Create an instance of the EfficientNet-B2 model using your function
    model = effnetb2

    # Load the saved best model state_dict() into the model
    model.load_state_dict(torch.load(model_path, map_location=torch.device('cpu')))

    # Set the model in evaluation mode
    model.eval()

    return model

# Specify the path to the saved model state dictionary
best_model_path = "best_effnetb2_10_epochs.pth"

# Load the model
best_model = load_best_effnetb2_model(best_model_path)

# Put EffNetB2 on CPU
best_model.to("cpu")

# Check the device
next(iter(best_model.parameters()))


# Prediction Function for the Gradio App
def predict(input_image, model=best_model, transform=effnetb2_transforms, class_names=class_names):

    # Create a dictionary to store prediction information
    pred_dict = {}

    # Start the prediction timer
    start_time = timer()

    # Preprocess the input image
    input_image = transform(input_image).unsqueeze(0)  # Add batch dimension

    # Prepare the model for inference
    model.eval()

    # Get prediction probability, prediction label, and prediction class
    with torch.inference_mode():
        pred_logit = model(input_image)               # perform inference on the input image
        pred_prob = torch.softmax(pred_logit, dim=1)  # convert logits into prediction probabilities
        pred_label = torch.argmax(pred_prob, dim=1)   # convert prediction probabilities into prediction label
        pred_class = class_names[pred_label.cpu()]    # get the predicted class (on CPU)

        # Store prediction information in the dictionary
        pred_dict["pred_prob"] = round(pred_prob.max().cpu().item(), 4)
        pred_dict["pred_class"] = pred_class

        # End the timer and calculate the time for prediction
        end_time = timer()
        pred_dict["time_for_pred"] = round(end_time - start_time, 4)

    return pred_class, pred_dict["pred_prob"], pred_dict["time_for_pred"]

## Deploy Gradio App

# Define sample images as PIL.Image objects
sample_image1 = "sample_images/JABIRU.jpg"
sample_image2 = "sample_images/APAPANE.jpg"
sample_image3 = "sample_images/LITTLE_AUK.jpg"

# Create title, description and article strings
title = "Birds Prediction 🕊️🦅🦆"
description = "An EfficientNetB2 <b>Computer Vision</b> Model to Classify Birds. \
               <br>The model achieved <b>93% accuracy</b> on the validation set. \
               <br>To upload your own photo you can check first in the classes.py file \
               the different <b>525 classes</b> trained with this model."
article = "Created by Benito Martin"

# Create the Gradio demo
demo = gr.Interface(fn=predict,
                    inputs=[
                        gr.Image(type="pil", label="Upload Your Image")
                            ],
                    outputs=[
                        gr.Label(num_top_classes=525, label="Predicted Bird"),
                        gr.Number(label="Prediction Probability"),
                        gr.Number(label="Prediction Time (s)")
                    ],
                    examples=[sample_image1, sample_image2, sample_image3],
                    title=title,
                    description=description,
                    article=article)

# Launch the demo!
demo.launch(debug=True, share=True)