This notebook demonstrates a number of Convolutional Neural Network image generation techniques implemented with TensorFlow for fun and science:
The network under examination is the GoogLeNet architecture, trained to classify images into one of 1000 categories of the ImageNet dataset. It consists of a set of layers that apply a sequence of transformations to the input image. The parameters of these transformations were determined during the training process by a variant of gradient descent algorithm. The internal image representations may seem obscure, but it is possible to visualize and interpret them. In this notebook we are going to present a few tricks that allow to make these visualizations both efficient to generate and even beautiful. Impatient readers can start with exploring the full galleries of images generated by the method described here for GoogLeNet and VGG16 architectures.
In [ ]:
# boilerplate code
from __future__ import print_function
import os
from io import BytesIO
import numpy as np
from functools import partial
import PIL.Image
from IPython.display import clear_output, Image, display, HTML
import tensorflow as tf
#Added imports
import time
import uuid
The pretrained network can be downloaded here. Unpack the tensorflow_inception_graph.pb file from the archive and set its path to model_fn variable. Alternatively you can uncomment and run the following cell to download the network:
In [ ]:
#!wget https://storage.googleapis.com/download.tensorflow.org/models/inception5h.zip && unzip inception5h.zip
In [ ]:
model_fn = 'tensorflow_inception_graph.pb'
# creating TensorFlow session and loading the model
graph = tf.Graph()
sess = tf.InteractiveSession(graph=graph)
with tf.gfile.FastGFile(model_fn, 'rb') as f:
graph_def = tf.GraphDef()
graph_def.ParseFromString(f.read())
t_input = tf.placeholder(np.float32, name='input') # define the input tensor
imagenet_mean = 117.0
t_preprocessed = tf.expand_dims(t_input-imagenet_mean, 0)
tf.import_graph_def(graph_def, {'input':t_preprocessed})
To take a glimpse into the kinds of patterns that the network learned to recognize, we will try to generate images that maximize the sum of activations of particular channel of a particular convolutional layer of the neural network. The network we explore contains many convolutional layers, each of which outputs tens to hundreds of feature channels, so we have plenty of patterns to explore.
In [ ]:
layers = [op.name for op in graph.get_operations() if op.type=='Conv2D' and 'import/' in op.name]
feature_nums = [int(graph.get_tensor_by_name(name+':0').get_shape()[-1]) for name in layers]
print('Number of layers', len(layers))
print('Total number of feature channels:', sum(feature_nums))
# Helper functions for TF Graph visualization
def strip_consts(graph_def, max_const_size=32):
"""Strip large constant values from graph_def."""
strip_def = tf.GraphDef()
for n0 in graph_def.node:
n = strip_def.node.add()
n.MergeFrom(n0)
if n.op == 'Const':
tensor = n.attr['value'].tensor
size = len(tensor.tensor_content)
if size > max_const_size:
tensor.tensor_content = bytes("<stripped %d bytes>"%size, 'utf8')
return strip_def
def rename_nodes(graph_def, rename_func):
res_def = tf.GraphDef()
for n0 in graph_def.node:
n = res_def.node.add()
n.MergeFrom(n0)
n.name = rename_func(n.name)
for i, s in enumerate(n.input):
n.input[i] = rename_func(s) if s[0]!='^' else '^'+rename_func(s[1:])
return res_def
def show_graph(graph_def, max_const_size=32):
"""Visualize TensorFlow graph."""
if hasattr(graph_def, 'as_graph_def'):
graph_def = graph_def.as_graph_def()
strip_def = strip_consts(graph_def, max_const_size=max_const_size)
code = """
<script>
function load() {{
document.getElementById("{id}").pbtxt = {data};
}}
</script>
<link rel="import" href="https://tensorboard.appspot.com/tf-graph-basic.build.html" onload=load()>
<div style="height:600px">
<tf-graph-basic id="{id}"></tf-graph-basic>
</div>
""".format(data=repr(str(strip_def)), id='graph'+str(np.random.rand()))
iframe = """
<iframe seamless style="width:800px;height:620px;border:0" srcdoc="{}"></iframe>
""".format(code.replace('"', '"'))
display(HTML(iframe))
# Visualizing the network graph. Be sure expand the "mixed" nodes to see their
# internal structure. We are going to visualize "Conv2D" nodes.
tmp_def = rename_nodes(graph_def, lambda s:"/".join(s.split('_',1)))
show_graph(tmp_def)
Let's start with a naive way of visualizing these. Image-space gradient ascent!
In [ ]:
# Picking some internal layer. Note that we use outputs before applying the ReLU nonlinearity
# to have non-zero gradients for features with negative initial activations.
layer = 'mixed4d_3x3_bottleneck_pre_relu'
channel = 139 # picking some feature channel to visualize
# start with a gray image with a little noise
img_noise = np.random.uniform(size=(224,224,3)) + 100.0
def showarray(a, fmt='jpeg'):
a = np.uint8(np.clip(a, 0, 1)*255)
f = BytesIO()
PIL.Image.fromarray(a).save(f, fmt)
display(Image(data=f.getvalue()))
def visstd(a, s=0.1):
'''Normalize the image range for visualization'''
return (a-a.mean())/max(a.std(), 1e-4)*s + 0.5
def T(layer):
'''Helper for getting layer output tensor'''
return graph.get_tensor_by_name("import/%s:0"%layer)
def render_naive(t_obj, img0=img_noise, iter_n=20, step=1.0):
t_score = tf.reduce_mean(t_obj) # defining the optimization objective
t_grad = tf.gradients(t_score, t_input)[0] # behold the power of automatic differentiation!
img = img0.copy()
for i in range(iter_n):
g, score = sess.run([t_grad, t_score], {t_input:img})
# normalizing the gradient, so the same step size should work
g /= g.std()+1e-8 # for different layers and networks
img += g*step
print(score, end = ' ')
clear_output()
showarray(visstd(img))
render_naive(T(layer)[:,:,:,channel])
Looks like the network wants to show us something interesting! Let's help it. We are going to apply gradient ascent on multiple scales. Details formed on smaller scale will be upscaled and augmented with additional details on the next scale.
With multiscale image generation it may be tempting to set the number of octaves to some high value to produce wallpaper-sized images. Storing network activations and backprop values will quickly run out of GPU memory in this case. There is a simple trick to avoid this: split the image into smaller tiles and compute each tile gradient independently. Applying random shifts to the image before every iteration helps avoid tile seams and improves the overall image quality.
In [ ]:
def tffunc(*argtypes):
'''Helper that transforms TF-graph generating function into a regular one.
See "resize" function below.
'''
placeholders = list(map(tf.placeholder, argtypes))
def wrap(f):
out = f(*placeholders)
def wrapper(*args, **kw):
return out.eval(dict(zip(placeholders, args)), session=kw.get('session'))
return wrapper
return wrap
# Helper function that uses TF to resize an image
def resize(img, size):
img = tf.expand_dims(img, 0)
return tf.image.resize_bilinear(img, size)[0,:,:,:]
resize = tffunc(np.float32, np.int32)(resize)
def calc_grad_tiled(img, t_grad, tile_size=512):
'''Compute the value of tensor t_grad over the image in a tiled way.
Random shifts are applied to the image to blur tile boundaries over
multiple iterations.'''
sz = tile_size
h, w = img.shape[:2]
sx, sy = np.random.randint(sz, size=2)
img_shift = np.roll(np.roll(img, sx, 1), sy, 0)
grad = np.zeros_like(img)
for y in range(0, max(h-sz//2, sz),sz):
for x in range(0, max(w-sz//2, sz),sz):
sub = img_shift[y:y+sz,x:x+sz]
g = sess.run(t_grad, {t_input:sub})
grad[y:y+sz,x:x+sz] = g
return np.roll(np.roll(grad, -sx, 1), -sy, 0)
In [ ]:
def render_multiscale(t_obj, img0=img_noise, iter_n=10, step=1.0, octave_n=3, octave_scale=1.4):
t_score = tf.reduce_mean(t_obj) # defining the optimization objective
t_grad = tf.gradients(t_score, t_input)[0] # behold the power of automatic differentiation!
img = img0.copy()
for octave in range(octave_n):
if octave>0:
hw = np.float32(img.shape[:2])*octave_scale
img = resize(img, np.int32(hw))
for i in range(iter_n):
g = calc_grad_tiled(img, t_grad)
# normalizing the gradient, so the same step size should work
g /= g.std()+1e-8 # for different layers and networks
img += g*step
print('.', end = ' ')
clear_output()
showarray(visstd(img))
render_multiscale(T(layer)[:,:,:,channel])
This looks better, but the resulting images mostly contain high frequencies. Can we improve it? One way is to add a smoothness prior into the optimization objective. This will effectively blur the image a little every iteration, suppressing the higher frequencies, so that the lower frequencies can catch up. This will require more iterations to produce a nice image. Why don't we just boost lower frequencies of the gradient instead? One way to achieve this is through the Laplacian pyramid decomposition. We call the resulting technique Laplacian Pyramid Gradient Normalization.
In [ ]:
k = np.float32([1,4,6,4,1])
k = np.outer(k, k)
k5x5 = k[:,:,None,None]/k.sum()*np.eye(3, dtype=np.float32)
def lap_split(img):
'''Split the image into lo and hi frequency components'''
with tf.name_scope('split'):
lo = tf.nn.conv2d(img, k5x5, [1,2,2,1], 'SAME')
lo2 = tf.nn.conv2d_transpose(lo, k5x5*4, tf.shape(img), [1,2,2,1])
hi = img-lo2
return lo, hi
def lap_split_n(img, n):
'''Build Laplacian pyramid with n splits'''
levels = []
for i in range(n):
img, hi = lap_split(img)
levels.append(hi)
levels.append(img)
return levels[::-1]
def lap_merge(levels):
'''Merge Laplacian pyramid'''
img = levels[0]
for hi in levels[1:]:
with tf.name_scope('merge'):
img = tf.nn.conv2d_transpose(img, k5x5*4, tf.shape(hi), [1,2,2,1]) + hi
return img
def normalize_std(img, eps=1e-10):
'''Normalize image by making its standard deviation = 1.0'''
with tf.name_scope('normalize'):
std = tf.sqrt(tf.reduce_mean(tf.square(img)))
return img/tf.maximum(std, eps)
def lap_normalize(img, scale_n=4):
'''Perform the Laplacian pyramid normalization.'''
img = tf.expand_dims(img,0)
tlevels = lap_split_n(img, scale_n)
tlevels = list(map(normalize_std, tlevels))
out = lap_merge(tlevels)
return out[0,:,:,:]
# Showing the lap_normalize graph with TensorBoard
lap_graph = tf.Graph()
with lap_graph.as_default():
lap_in = tf.placeholder(np.float32, name='lap_in')
lap_out = lap_normalize(lap_in)
show_graph(lap_graph)
In [ ]:
def render_lapnorm(t_obj, img0=img_noise, visfunc=visstd,
iter_n=10, step=1.0, octave_n=3, octave_scale=1.4, lap_n=4):
t_score = tf.reduce_mean(t_obj) # defining the optimization objective
t_grad = tf.gradients(t_score, t_input)[0] # behold the power of automatic differentiation!
# build the laplacian normalization graph
lap_norm_func = tffunc(np.float32)(partial(lap_normalize, scale_n=lap_n))
img = img0.copy()
for octave in range(octave_n):
if octave>0:
hw = np.float32(img.shape[:2])*octave_scale
img = resize(img, np.int32(hw))
for i in range(iter_n):
g = calc_grad_tiled(img, t_grad)
g = lap_norm_func(g)
img += g*step
print('.', end = ' ')
clear_output()
showarray(visfunc(img))
render_lapnorm(T(layer)[:,:,:,channel])
In [ ]:
render_lapnorm(T(layer)[:,:,:,65])
Lower layers produce features of lower complexity.
In [ ]:
render_lapnorm(T('mixed3b_1x1_pre_relu')[:,:,:,101])
There are many interesting things one may try. For example, optimizing a linear combination of features often gives a "mixture" pattern.
In [ ]:
render_lapnorm(T(layer)[:,:,:,65]+T(layer)[:,:,:,139], octave_n=4)
In [ ]:
def render_deepdream(t_obj, img0=img_noise,
iter_n=10, step=1.5, octave_n=4, octave_scale=1.4,
show_image=False):
t_score = tf.reduce_mean(t_obj) # defining the optimization objective
t_grad = tf.gradients(t_score, t_input)[0] # behold the power of automatic differentiation!
# split the image into a number of octaves
img = img0
octaves = []
for i in range(octave_n-1):
hw = img.shape[:2]
lo = resize(img, np.int32(np.float32(hw)/octave_scale))
hi = img-resize(lo, hw)
img = lo
octaves.append(hi)
# generate details octave by octave
for octave in range(octave_n):
if octave>0:
hi = octaves[-octave]
img = resize(img, hi.shape[:2])+hi
for i in range(iter_n):
g = calc_grad_tiled(img, t_grad)
img += g*(step / (np.abs(g).mean()+1e-7))
print('.',end = ' ')
clear_output()
#Added as a flag to not have to show the image every iteration
if show_image == True:
showarray(img/255.0)
return img/255.0
In [ ]:
img0 = PIL.Image.open('pilatus800.jpg')
img0 = np.float32(img0)
showarray(img0/255.0)
In [ ]:
parameters = [
{'name': uuid.uuid4(), 't_obj_filter': 139, 'img0': img0, 'iter_n':10, 'step':1.5, 'octave_n':4, 'octave_scale':1.4},
{'name': uuid.uuid4(), 't_obj_filter': 139, 'img0': img0, 'iter_n':10, 'step':1.5, 'octave_n':4, 'octave_scale':1.0},
#{'name': uuid.uuid4(), 't_obj_filter': 139, 'img0': img0, 'iter_n':10, 'step':1.5, 'octave_n':4, 'octave_scale':2.0},
#{'name': uuid.uuid4(), 't_obj_filter': 139, 'img0': img0, 'iter_n':10, 'step':1.5, 'octave_n':10, 'octave_scale':1.4},
#{'name': uuid.uuid4(), 't_obj_filter': 139, 'img0': img0, 'iter_n':10, 'step':1.5, 'octave_n':4, 'octave_scale':1},
#{'name': uuid.uuid4(), 't_obj_filter': 139, 'img0': img0, 'iter_n':10, 'step':1.5, 'octave_n':4, 'octave_scale':4},
#{'name': uuid.uuid4(), 't_obj_filter': 139, 'img0': img0, 'iter_n':10, 'step':1.5, 'octave_n':4, 'octave_scale':1}
]
In [ ]:
output_data={}
for i in range(0,len(parameters)):
start_time=time.time()
# Perform the deep dream rendering and store output array, which is called output_image
output_image=render_deepdream(T(layer)[:,:,:,parameters[i]['t_obj_filter']],
img0=parameters[i]['img0'],
iter_n=parameters[i]['iter_n'],
step=parameters[i]['step'],
octave_n=parameters[i]['octave_n'],
octave_scale=parameters[i]['octave_scale'],
show_image=False)
#log the timing as a dictionary and save the image
end_time=time.time()
output_data[str(parameters[i]['name'])]=(output_image.copy(), end_time-start_time, parameters[i])
In [ ]:
# Output images for comparison
for key in output_data.keys():
showarray(output_data[key][0])
print ('Test Name:',key)
print ('Timing:',output_data[key][1])
print(("Filter: %s\niter_n: %s\nstep: %s\noctave_n: %s, octave_scale:%s") %
(output_data[key][2]['t_obj_filter'],
output_data[key][2]['iter_n'],
output_data[key][2]['step'],
output_data[key][2]['octave_n'],
output_data[key][2]['octave_scale'], ) )
print ('\n')
# Compare the average, best and worst times
processing_times=[]
for key in output_data.keys():
processing_times.append(output_data[key][1])
processing_times=np.array(processing_times)
print(("Average Processing Time: %s\nSTDEVA: %s\nNumber of Samples: %s\n") %
(processing_times.mean(),processing_times.std(),len(processing_times)) )
In [ ]:
In [ ]: