积分梯度

import matplotlib.pylab as plt
import numpy as np
import tensorflow as tf
import tensorflow_hub as hub

2023-11-08 00:00:52.002097: E external/local_xla/xla/stream_executor/cuda/cuda_dnn.cc:9261] Unable to register cuDNN factory: Attempting to register factory for plugin cuDNN when one has already been registered
2023-11-08 00:00:52.002146: E external/local_xla/xla/stream_executor/cuda/cuda_fft.cc:607] Unable to register cuFFT factory: Attempting to register factory for plugin cuFFT when one has already been registered
2023-11-08 00:00:52.003852: E external/local_xla/xla/stream_executor/cuda/cuda_blas.cc:1515] Unable to register cuBLAS factory: Attempting to register factory for plugin cuBLAS when one has already been registered

model = tf.keras.Sequential([
    hub.KerasLayer(
        name='inception_v1',
        handle='https://tfhub.dev/google/imagenet/inception_v1/classification/4',
        trainable=False),
])
model.build([None, 224, 224, 3])
model.summary()

Model: "sequential"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 inception_v1 (KerasLayer)   (None, 1001)              6633209   
                                                                 
=================================================================
Total params: 6633209 (25.30 MB)
Trainable params: 0 (0.00 Byte)
Non-trainable params: 6633209 (25.30 MB)
_________________________________________________________________

def load_imagenet_labels(file_path):
  labels_file = tf.keras.utils.get_file('ImageNetLabels.txt', file_path)
  with open(labels_file) as reader:
    f = reader.read()
    labels = f.splitlines()
  return np.array(labels)

imagenet_labels = load_imagenet_labels('https://storage.googleapis.com/download.tensorflow.org/data/ImageNetLabels.txt')

def read_image(file_name):
  image = tf.io.read_file(file_name)
  image = tf.io.decode_jpeg(image, channels=3)
  image = tf.image.convert_image_dtype(image, tf.float32)
  image = tf.image.resize_with_pad(image, target_height=224, target_width=224)
  return image

img_url = {
    'Fireboat': 'http://storage.googleapis.com/download.tensorflow.org/example_images/San_Francisco_fireboat_showing_off.jpg',
    'Giant Panda': 'http://storage.googleapis.com/download.tensorflow.org/example_images/Giant_Panda_2.jpeg',
}

img_paths = {name: tf.keras.utils.get_file(name, url) for (name, url) in img_url.items()}
img_name_tensors = {name: read_image(img_path) for (name, img_path) in img_paths.items()}

Downloading data from http://storage.googleapis.com/download.tensorflow.org/example_images/San_Francisco_fireboat_showing_off.jpg
3954129/3954129 [==============================] - 0s 0us/step
Downloading data from http://storage.googleapis.com/download.tensorflow.org/example_images/Giant_Panda_2.jpeg
802859/802859 [==============================] - 0s 0us/step

plt.figure(figsize=(8, 8))
for n, (name, img_tensors) in enumerate(img_name_tensors.items()):
  ax = plt.subplot(1, 2, n+1)
  ax.imshow(img_tensors)
  ax.set_title(name)
  ax.axis('off')
plt.tight_layout()

def top_k_predictions(img, k=3):
  image_batch = tf.expand_dims(img, 0)
  predictions = model(image_batch)
  probs = tf.nn.softmax(predictions, axis=-1)
  top_probs, top_idxs = tf.math.top_k(input=probs, k=k)
  top_labels = imagenet_labels[tuple(top_idxs)]
  return top_labels, top_probs[0]

for (name, img_tensor) in img_name_tensors.items():
  plt.imshow(img_tensor)
  plt.title(name, fontweight='bold')
  plt.axis('off')
  plt.show()

  pred_label, pred_prob = top_k_predictions(img_tensor)
  for label, prob in zip(pred_label, pred_prob):
    print(f'{label}: {prob:0.1%}')

fireboat: 32.6%
pier: 12.7%
suspension bridge: 5.7%

giant panda: 89.4%
teddy: 0.3%
gibbon: 0.3%

def f(x):
  """A simplified model function."""
  return tf.where(x < 0.8, x, 0.8)

def interpolated_path(x):
  """A straight line path."""
  return tf.zeros_like(x)

x = tf.linspace(start=0.0, stop=1.0, num=6)
y = f(x)

baseline = tf.zeros(shape=(224,224,3))

plt.imshow(baseline)
plt.title("Baseline")
plt.axis('off')
plt.show()

m_steps=50
alphas = tf.linspace(start=0.0, stop=1.0, num=m_steps+1) # Generate m_steps intervals for integral_approximation() below.

def interpolate_images(baseline,
                       image,
                       alphas):
  alphas_x = alphas[:, tf.newaxis, tf.newaxis, tf.newaxis]
  baseline_x = tf.expand_dims(baseline, axis=0)
  input_x = tf.expand_dims(image, axis=0)
  delta = input_x - baseline_x
  images = baseline_x +  alphas_x * delta
  return images

interpolated_images = interpolate_images(
    baseline=baseline,
    image=img_name_tensors['Fireboat'],
    alphas=alphas)

fig = plt.figure(figsize=(20, 20))

i = 0
for alpha, image in zip(alphas[0::10], interpolated_images[0::10]):
  i += 1
  plt.subplot(1, len(alphas[0::10]), i)
  plt.title(f'alpha: {alpha:.1f}')
  plt.imshow(image)
  plt.axis('off')

plt.tight_layout();

def compute_gradients(images, target_class_idx):
  with tf.GradientTape() as tape:
    tape.watch(images)
    logits = model(images)
    probs = tf.nn.softmax(logits, axis=-1)[:, target_class_idx]
  return tape.gradient(probs, images)

path_gradients = compute_gradients(
    images=interpolated_images,
    target_class_idx=555)

print(path_gradients.shape)

(51, 224, 224, 3)

pred = model(interpolated_images)
pred_proba = tf.nn.softmax(pred, axis=-1)[:, 555]

def integral_approximation(gradients):
  # riemann_trapezoidal
  grads = (gradients[:-1] + gradients[1:]) / tf.constant(2.0)
  integrated_gradients = tf.math.reduce_mean(grads, axis=0)
  return integrated_gradients

ig = integral_approximation(
    gradients=path_gradients)

print(ig.shape)

(224, 224, 3)

def integrated_gradients(baseline,
                         image,
                         target_class_idx,
                         m_steps=50,
                         batch_size=32):
  # Generate alphas.
  alphas = tf.linspace(start=0.0, stop=1.0, num=m_steps+1)

  # Collect gradients.    
  gradient_batches = []

  # Iterate alphas range and batch computation for speed, memory efficiency, and scaling to larger m_steps.
  for alpha in tf.range(0, len(alphas), batch_size):
    from_ = alpha
    to = tf.minimum(from_ + batch_size, len(alphas))
    alpha_batch = alphas[from_:to]

    gradient_batch = one_batch(baseline, image, alpha_batch, target_class_idx)
    gradient_batches.append(gradient_batch)

  # Concatenate path gradients together row-wise into single tensor.
  total_gradients = tf.concat(gradient_batches, axis=0)

  # Integral approximation through averaging gradients.
  avg_gradients = integral_approximation(gradients=total_gradients)

  # Scale integrated gradients with respect to input.
  integrated_gradients = (image - baseline) * avg_gradients

  return integrated_gradients

@tf.function
def one_batch(baseline, image, alpha_batch, target_class_idx):
    # Generate interpolated inputs between baseline and input.
    interpolated_path_input_batch = interpolate_images(baseline=baseline,
                                                       image=image,
                                                       alphas=alpha_batch)

    # Compute gradients between model outputs and interpolated inputs.
    gradient_batch = compute_gradients(images=interpolated_path_input_batch,
                                       target_class_idx=target_class_idx)
    return gradient_batch

ig_attributions = integrated_gradients(baseline=baseline,
                                       image=img_name_tensors['Fireboat'],
                                       target_class_idx=555,
                                       m_steps=240)

print(ig_attributions.shape)

(224, 224, 3)

_ = plot_img_attributions(image=img_name_tensors['Fireboat'],
                          baseline=baseline,
                          target_class_idx=555,
                          m_steps=240,
                          cmap=plt.cm.inferno,
                          overlay_alpha=0.4)

_ = plot_img_attributions(image=img_name_tensors['Giant Panda'],
                          baseline=baseline,
                          target_class_idx=389,
                          m_steps=55,
                          cmap=plt.cm.viridis,
                          overlay_alpha=0.5)