app.py

import utils
from huggingface_hub.keras_mixin import from_pretrained_keras
from PIL import Image
import streamlit as st
import tensorflow as tf

# Inputs
st.title("Input your image")
image_url = st.text_input(
	label="URL of image",
	value="https://dl.fbaipublicfiles.com/dino/img.png",
	placeholder="https://your-favourite-image.png"
)

# Outputs
st.title("Original Image from URL")

# Preprocess the same image but with normlization.
image, preprocessed_image = utils.load_image_from_url(
	image_url,
	model_type="dino"
)
st.image(image, caption="Original Image")

st.title("Attention Heat Maps")

# Load the DINO model
with st.spinner('Processing the heat maps. Wait for it 🔥...'):
	dino = from_pretrained_keras("probing-vits/vit-dino-base16")

	# Get the attention scores
	_, attention_score_dict = dino.predict(preprocessed_image)

	# De-normalize the image for visual clarity. 
	in1k_mean = tf.constant([0.485 * 255, 0.456 * 255, 0.406 * 255])  
	in1k_std = tf.constant([0.229 * 255, 0.224 * 255, 0.225 * 255])  
	preprocessed_img_orig = (preprocessed_image * in1k_std) + in1k_mean
	preprocessed_img_orig = preprocessed_img_orig / 255.
	preprocessed_img_orig = tf.clip_by_value(preprocessed_img_orig, 0.0, 1.0).numpy()

	attentions = utils.attention_heatmap(
		attention_score_dict=attention_score_dict,
		image=preprocessed_img_orig
	)

	utils.plot(attentions=attentions, image=preprocessed_img_orig)

# Show the attention maps
image = Image.open("heat_map.png")
st.image(image, caption="Attention Heat Maps")