src/Hockey_Breeds.py

import streamlit as st
from streamlit_image_select import image_select
import zip_files

import random
import logging
from huggingface_hub import from_pretrained_fastai

@st.cache_resource
def get_model():
    repo_id = "danbiagini/hockey_breeds"
    return from_pretrained_fastai(repo_id)

def classify_image(learn, img):
    categories = ('Hockey Goalie', 'Hockey Player', "Hockey Referee")
    pred,idx,prob = learn.predict(img)
    return dict(zip(categories, map(float, prob)))

def reroll_sample_images():
    # unzip the sample images
    players = zip_files.extract_files_from_zip("src/images/samples/player-samples.zip")
    goalies = zip_files.extract_files_from_zip("src/images/samples/goalie-samples.zip")
    referees = zip_files.extract_files_from_zip("src/images/samples/referee-samples.zip")

    #randomize a single file from players, goalies and referee for samples
    st.session_state.sample = dict()
    st.session_state.sample["player"] = players[list(players.keys())[random.randint(0, len(players) - 1)]]
    st.session_state.sample["goalie"] = goalies[list(goalies.keys())[random.randint(0, len(goalies) - 1)]]
    st.session_state.sample["referee"] = referees[list(referees.keys())[random.randint(0, len(referees) - 1)]]

if 'sample' not in st.session_state:
    reroll_sample_images()

st.set_page_config(page_title='Hockey Breeds', layout="wide",
                   page_icon=":frame_with_picture:")

st.title('Hockey Breeds - Hello Computer Vision')

st.subheader('Image Classification')

img_class = '''Image Classification in Computer Vision is the act of determining the most appropriate label for an entire image from a set of fixed labels.  
A popular topic of image classification in Computer Vision introductions and courses is to use an example problem of training a model to label images of various pet breeds.
*Hockey Breeds* is a hockey slant on this common theme in Computer Vision educational materials.'''

st.markdown(img_class)

st.subheader("Hockey Image Classification")
desc = '''This "Hockey Breeds" model was built using 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).  I fine tuned the model by labeling the hockey photos, then training using python (*Fast.ai* & *PyTorch* libraries).

The total training time for this was approximately 5 minutes running on a low-end GPU. It’s impressive how accurate this quick / small model can be!'''

st.markdown(desc)
st.image("src/images/samples/sampl_batch.png")
st.subheader("Validation Results")
st.markdown('Validation of the model\'s performance was done using 26 images not included in the training set.  The model performed fairly well against the validation dataset, with only 1 misclassified image.')
st.image("src/images/artifacts/confusion_matrix_v1.png", caption="Confusion Matrix for Hockey Breeds ")

st.subheader("Try It Out")

img = image_select(label="Select an image and hockey breeds will guess a label.  See if you can find some incorrect guesses...", images=list(st.session_state.sample.values()))

st.button("Re-roll Samples", on_click=reroll_sample_images)

model = get_model()

if img:
    res = classify_image(model, img)
    # Sort the dictionary items by value in descending order
    max = 0
    max_label = ""
    for k,v in res.items():
        prob = round(v*100, 2)
        if prob > max:
            max = prob
            max_label = k
    st.metric(label=max_label, value=max)