app.py

import gradio as gr
from fastai.vision.all import *
from os import listdir
import random

title = "Who Am I? (Hockey Edition)"

desc = """<center>This app uses a 'neural network' (a type of machine learning model) to classify 
an image as containing a hockey player, a hockey goalie or a hockey referee.</center>"""

art = """I built this model using about 50 hockey related images found on the web and in my own collection. I started with a pretrained `resnet18` model (resnet18 is trained on `imagenet`, a very large dataset with millions of images)
and then performed `fine tuning` of this using python and the `fast.ai` library. 

The total training time for this was about 5 minutes on a basic GPU.  It's impressive how accurate this quick / small model can be!"""

learn = load_learner('hockey_model.pkl')

categories = ('Hockey Goalie', 'Hockey Player', "Hockey Referee")
image = gr.Image(shape=(192, 192))
label = gr.Label()
skater_example = 'assets/skaters/' + random.choice(listdir('assets/skaters')) 
ref_example = 'assets/referees/' + random.choice(listdir('assets/referees'))
goalie_example = 'assets/goalies/' + random.choice(listdir('assets/goalies'))

def classify_image(img):
    pred,idx,prob = learn.predict(img)
    return dict(zip(categories, map(float, prob)))

iface = gr.Interface(fn=classify_image, inputs=image, outputs=label, theme=gr.themes.Glass(), 
                    examples=[skater_example, ref_example, goalie_example], title=title, 
                    description=desc, article=art)
iface.launch(share=False, debug=True)