app.py

import io
import torch
import streamlit as st
from PIL import Image
from transformers import AutoImageProcessor, SwinForImageClassification


class ImageClassifier(object):
    def __init__(self, model):
        self.model = model

    def get_top_5_predictions(self, image):       
        values, indices = torch.topk(self.get_output_probabilities(image), 5)
        return [
            {'label': self.model.config.id2label[i.item()], 'score': v.item()}
            for i, v in zip(indices, values)
        ]
    
    def get_output_probabilities(self, image):
        output = self.classify_image(image)
        return torch.nn.functional.softmax(output.logits[0], dim=0)

    def classify_image(self, image):
        image_processor = self.create_image_processor()
        inputs = image_processor(image, return_tensors='pt')
        with torch.no_grad():
            return self.model(**inputs)
    
    def create_image_processor(self):
        return AutoImageProcessor.from_pretrained(self.model.name_or_path)


class ImageClassificationApp(object):
    def __init__(self, title, classifier):
        self.title = title
        self.classifier = classifier

    def render(self):
        st.title(self.title)
        uploaded_image = self.get_uploaded_image()
        if uploaded_image is not None:
            self.show_image_and_results(uploaded_image)

    def get_uploaded_image(self):
        return st.file_uploader('Choose an image...', type=['jpg', 'png', 'jpeg'])
    
    def show_image_and_results(self, uploaded_image):
        self.show_uploaded_image(uploaded_image)
        self.show_classification_results(self.get_image(uploaded_image.read()))

    def show_uploaded_image(self, uploaded_image):
        st.image(uploaded_image, caption='Uploaded Image', use_column_width=True)

    def show_classification_results(self, image):
        st.subheader('Classification Results:')
        self.write_top_5_predictions(image)        

    def write_top_5_predictions(self, image):        
        for prediction in self.classifier.get_top_5_predictions(image):
            st.write(f"- {prediction['label']}: {prediction['score']:.4f}")

    def get_image(self, image_data):
        return Image.open(io.BytesIO(image_data))


if __name__ == '__main__':
    model = SwinForImageClassification.from_pretrained(
        'microsoft/swin-tiny-patch4-window7-224'
    )
    classifier = ImageClassifier(model)
    ImageClassificationApp(
        'Swin Image Classification App', 
        classifier
    ).render()