DeepDream

1. In InceptionNet , there are 11 layers where convolutions are concatenated. Exciting

these layers in different combinations results in different outputs.

So you can use layers "mixed0" to "mixed10" in the layers array.

In general, exciting as we go to deep layers, we get good extracted features. Hence

there will be a difference between images generated by deep layers vs shallow layers.

2. It looks like there is a bug in the tutorial

https://www.tensorflow.org/tutorials/generative/deepdream

in the method run_deep_dream_simple

The statement :

steps_remaining -= run_steps

should ideally be

steps_remaining -= 1

The current statement terminates the loop immediately after one run.

3. Another important point to be noted is we are using GRADIENT ASCENT , not DESCENT.

---------------------------------------------------------------------------------------

import tensorflow as tf 

from tensorflow import keras 

import PIL 

import IPython.display as display 

import numpy as np 

url = "https://storage.googleapis.com/download.tensorflow.org/example_images/YellowLabradorLooking_new.jpg"

def download(url , max_dim = None) : 

  name = url.split("/")[-1]

  image_file = tf.keras.utils.get_file(name, origin = url)

  

  img = PIL.Image.open(image_file)

  if max_dim : 

    img.thumbnail((max_dim , max_dim))

  return np.array(img)

def deprocess(img)  : 

  img =  255 *(img + 1.0 ) /2.0

  img = tf.cast(img , tf.uint8)

  return img 

def show(img) : 

  display.display(PIL.Image.fromarray(np.array(img)))  

original_image = download(url , max_dim = 500)

show(original_image)

display.display(display.HTML("image"))

base_model = tf.keras.applications.InceptionV3(include_top= False , weights = "imagenet")

names = ["mixed3" , "mixed5"]

layers = [base_model.get_layer(name).output for name in names]

dream_model = tf.keras.Model(inputs = base_model.input , outputs = layers)

def calc_loss(model , img) : 

  img_batch = tf.expand_dims(img, axis = 0) 

  activations = model(img_batch)

  if len(activations) ==1 : 

    activations = [activations]

  losses = [] 

  for act in activations: 

    loss = tf.reduce_mean(act)    

    losses.append(loss)

  return tf.reduce_sum(loss) 

class DeepDream(tf.Module) : 

  def __init__(self , model) :

    self.model = model 

  @tf.function(

      input_signature = (

                         tf.TensorSpec(shape = [None , None,3] , dtype = tf.float32) , 

                         tf.TensorSpec(shape = [] , dtype = tf.int32) , 

                         tf.TensorSpec(shape = [] , dtype = tf.float32) , 

      )

  )

  def __call__(self , img , steps , step_size)   : 

    loss = tf.constant(0.0)

    for  n in range(steps) : 

      with tf.GradientTape() as tape : 

        tape.watch(img)

        loss  = calc_loss(self.model, img)

      gradients = tape.gradient(loss , img)

      gradients = gradients / (tf.math.reduce_std(gradients)) + 1e-8

      img = img + gradients * step_size

      img = tf.clip_by_value(img , -1 , 1)

    return loss , img 

deepdream = DeepDream(dream_model)    

def run_deep_dream_simple(img , steps = 100 , step_size = 0.01) : 

  img = tf.keras.applications.inception_v3.preprocess_input(img) 

  img = tf.convert_to_tensor(img)

  step_size = tf.convert_to_tensor(step_size)

  steps_remaining = steps

  step = 0 

  while steps_remaining : 

    if  steps_remaining > 100 : 

      run_steps = tf.constant(100)

    else :

      run_steps = tf.constant(steps_remaining)      

      steps_remaining -= 1 

      step += run_steps

      

      loss , img =  deepdream(img , run_steps , tf.constant(step_size))     

      show(deprocess(img))

      print( "Step {} , loss {} ".format(step , loss))

  result = deprocess(img)

  show(result)

  return result

dream_img = run_deep_dream_simple(img = original_image , steps = 100 , step_size = 0.01)