app.py

import gradio as gr
import os
import torch
import torchvision

from model import create_model
from timeit import default_timer as timer


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

model, model_transforms = create_model(num_classes=len(class_names))
model.load_state_dict(torch.load("ConvNeXt_Tiny_101classes_20_10epochs.pth", 
                                 map_location=torch.device('cpu'))) # load -> load_state_dict


def predict(img):
    
    time_start = timer()
        
    weights = torchvision.models.ConvNeXt_Tiny_Weights.DEFAULT 
    transform_convnext_tiny = weights.transforms()
    img_tensor = transform_convnext_tiny(img).unsqueeze(dim=0) # [Channels, Height, Width] -> [Batch_size, Channels, Height, Width]

    model.eval()
    with torch.inference_mode():
        predicted_probs = model(img_tensor).softmax(dim=1)
    
    # Class name & predicted probability for each class (required by Gradio)
    pred_labels_probs = {}
    for i in range(len(class_names)):
        pred_labels_probs[class_names[i]] = float(predicted_probs[0][i])

    return pred_labels_probs, round(timer() - time_start, 5)


app = gr.Interface(fn=predict, # mapping function for [ input -> output ]
                   
                   inputs=gr.Image(type="pil"), # Input data
                   outputs=[gr.Label(num_top_classes=3, label="Predictions"), # Output data (fn function's return values)
                            gr.Number(label="Inference time (s)")], 
                   
                   examples=[["examples/" + example] for example in os.listdir("examples")],

                   title='ConvNeXt_Food101',
                   description='A ConvNext CV model to classify 101 foods',
                   article='Model trained on 150 images per class')

app.launch()