app.py

import streamlit as st

from PIL import Image
import torch
from time import time

from transformers import CLIPProcessor, CLIPModel

def separator(chaine: str, sep: list[str] = [',', ';']) -> list[str]:
    """ ''.split() augmented method """
    elements = []
    for s in sep :
        chaine = chaine.replace(s, '\n')
    
    for ligne in chaine.split("\n"):
        elements.extend(ligne.split())
    
    return elements

st.set_page_config('Zero Shot Image', '🏷️', 'wide')
st.title('🏷️ Zero Shot Image Classification', help='Zero-shot image classification is the task of classifying previously unseen classes during training of a model')

info = st.empty()
col1, col2, col3 = st.columns(3, gap='large')

# Model
MODELS = {
    'Fast' : 'openai/clip-vit-base-patch32', 
    'Accurate' : 'openai/clip-vit-large-patch14'
}
model_type = col1.radio('Model', MODELS.keys(), horizontal=True)

# Input image
_uploaded_image = col1.file_uploader('Image', ['png', 'jpg', 'jpeg', 'webp'])
if _uploaded_image is not None :
    image = Image.open(_uploaded_image)
    col2.image(image)

# Classes
_help = 'Enter classes separated by commas, semicolons, tabs, or newlines'
_classes = col1.text_area('Classes', '', help=_help, placeholder=_help)
classes = [c.strip() for c in separator(_classes, [',', '\t', ';', '\n'])]

# Run
if col1.button('Classify', use_container_width=True, type='primary')     :
    if _uploaded_image is None :
        info.warning('Upload an image to start')

    elif len(classes) < 2 :
        info.warning('Insert fiew classes in text area')
        
    else :
        # Load model
        start = time()
        model = CLIPModel.from_pretrained(MODELS[model_type])
        processor = CLIPProcessor.from_pretrained(pretrained_model_name_or_path=MODELS[model_type])

        # Process
        inputs = processor(text=classes, images=image, return_tensors="pt", padding=True)

        # Outputs
        outputs = model(**inputs)
        logits_per_image = outputs.logits_per_image 
        probs = logits_per_image.softmax(dim=1)
        indice = torch.argmax(probs).item()
            
        with col3 : # Progress bars
            f'## {classes[indice]}'

            result = probs.tolist()[0]
            for i, c in enumerate(classes) :
                col3.progress(result[i], f'`{int(result[i]*100)}%` - {c}')
            
            f'{round(time()-start, 2)}s'