shayaf84

#Extract only the values for title (input) and class (output)
data = total[['title','class']]
data.head()

shayaf84

real = pd.read_csv("/content/drive/My Drive/News/True.csv")
fake = pd.read_csv("/content/drive/My Drive/News/Fake.csv")


#Shape of real news dataset
print("Real news: ",real.shape)
#Shape of fake news dataset
print("Fake News: ", fake.shape)

# Assigning a value of 0 for all real news data and placing it in the dataframe

shayaf84

def plotAttributions(baseline,attributions,
                          image,
                          cmap=None,
                          overlay_alpha=0.4):


  #Sum of the abs value of the attriutions determining feature importance
  mask = tf.reduce_sum(tf.math.abs(attributions), axis=-1)

shayaf84

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

integ = integralEstimates(gradients=computedGradients)

shayaf84

def gradient(images, targetClassindex):
  #Computes gradient of probabilities with respect to images
  with tf.GradientTape() as gradRecord:
    gradRecord.watch(images)
    logits = vgg(images)
    probs = tf.nn.softmax(logits, axis=-1)[:, targetClassindex]
    
  return gradRecord.gradient(probs, images)

shayaf84

interpolatedImages = interpolation(baseline=baseline,image=image,alpha=alpha)

fig = plt.figure(figsize=(20, 20))
i = 0
for alphas, image in zip(alpha[0::10], interpolatedImages[0::10]):
  i += 1
  plt.subplot(1, len(alpha[0::10]), i)
  plt.title(f'alpha: {alphas:.1f}')
  plt.imshow(image)
  plt.axis('off')

shayaf84

def interpolation(baseline, image, alpha):
  #Interpolates images by computing the difference between its pixel values 
  #and the baseline (all zeroes)
  #From there, construct a new tensor containing the baseline and interpolated
  #images, which are the picture values multiplied by the linear relationship.
  alphaTensor = alpha[:, tf.newaxis, tf.newaxis, tf.newaxis]
  baselineTensor = tf.expand_dims(baseline, axis=0)
  inputTensor = tf.expand_dims(image, axis=0)
  diff = inputTensor - baselineTensor
  images = baselineTensor +  alphaTensor * diff

shayaf84

from tensorflow import io

#Convert image to a tensor - 3 RGB Channels - resized to 224x224
image = tf.io.read_file("drive/MyDrive/oboe.png")
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)


#Generates 50 intervals for the interpolation of images

shayaf84

from tensorflow import io

#Convert image to a tensor - 3 RGB Channels - resized to 224x224
image = tf.io.read_file("drive/MyDrive/oboe.png")
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)


#Generates 50 intervals for the interpolation of images

shayaf84

#Implementing our Baseline
baseline = tf.zeros(shape=(224,224,3))

#Displaying it with matplotlib
plt.imshow(baseline)
plt.title("Baseline")
plt.axis('off')
plt.show()