app.py

import tensorflow as tf
from tensorflow.keras.applications.resnet_v2 import ResNet50V2
from tensorflow.keras.preprocessing import image
from tensorflow.keras.applications.resnet_v2 import preprocess_input, decode_predictions
import matplotlib.pyplot as plt
from alibi.explainers import IntegratedGradients
from alibi.datasets import load_cats
from alibi.utils.visualization import visualize_image_attr
import numpy as np
from PIL import Image
import io
import time
import os
import copy
import pickle
import datetime
import urllib.request
import gradio as gr


url = "https://upload.wikimedia.org/wikipedia/commons/3/38/Adorable-animal-cat-20787.jpg"
path_input = "/content/cat.jpg"
urllib.request.urlretrieve(url, filename=path_input)

url = "https://upload.wikimedia.org/wikipedia/commons/4/43/Cute_dog.jpg"
path_input = "/content/dog.jpg"
urllib.request.urlretrieve(url, filename=path_input)

model = ResNet50V2(weights='imagenet')

n_steps = 50
method = "gausslegendre"
internal_batch_size = 50
ig  = IntegratedGradients(model,
                          n_steps=n_steps,
                          method=method,
                          internal_batch_size=internal_batch_size)

# refs: 
# - fig2pil: https://stackoverflow.com/questions/57316491/how-to-convert-matplotlib-figure-to-pil-image-object-without-saving-image
def do_process(img, baseline):
  instance = image.img_to_array(img)
  instance = np.expand_dims(instance, axis=0)
  instance = preprocess_input(instance)
  preds = model.predict(instance)
  lstPreds = decode_predictions(preds, top=3)[0]
  dctPreds = {lstPreds[i][1]: round(float(lstPreds[i][2]),2) for i in range(len(lstPreds))}
  predictions = preds.argmax(axis=1)
  if baseline is 'white':
    baselines = bls = np.ones(instance.shape).astype(instance.dtype)
  elif baseline is 'black':
    baselines = bls = np.zeros(instance.shape).astype(instance.dtype)
  else:
    baselines = np.random.random_sample(instance.shape).astype(instance.dtype)
  explanation = ig.explain(instance,
                         baselines=baselines,
                         target=predictions)
  attrs = explanation.attributions[0]
  fig, ax = visualize_image_attr(attr=attrs.squeeze(), original_image=img, method='blended_heat_map',
                      sign='all', show_colorbar=True, title='Overlaid Attributions',
                      plt_fig_axis=None, use_pyplot=False)
  buf = io.BytesIO()
  fig.savefig(buf)
  buf.seek(0)
  img_res = Image.open(buf)
  return img_res, dctPreds
 
 input_im = gr.inputs.Image(shape=(224, 224), image_mode='RGB', 
                         invert_colors=False, source="upload",
                         type="pil")
input_drop = gr.inputs.Dropdown(label='Baseline (default: random)',
	                  choices=sorted(list(['black', 'white', 'random'])), default='random', type='value')

output_img = gr.outputs.Image(label='Output image', type='pil')
output_label = gr.outputs.Label(num_top_classes=3)

title = "XAI - Integrated gradients"
description = "Playground: Integrated gradients for a ResNet model trained on Imagenet dataset. Tools: Alibi, TF, Gradio."
examples = [['./cat.jpg'],['./dog.jpg']]
article="<p style='text-align: center'><a href='https://github.com/mawady/colab-recipes-cv' target='_blank'>Colab recipes for computer vision - Dr. Mohamed Elawady</a></p>"
iface = gr.Interface(
    fn=do_process, 
    inputs=[input_im, input_drop], 
    outputs=[output_img,output_label],
    live=False,
    interpretation=None,
    title=title,
    description=description,
    article=article,
    examples=examples
)

iface.test_launch()

iface.launch(share=True, debug=True)