Storypost | 2020.11.29

Neural style transfer Valentino Rossi Ducati Laguan Seca Ralph Steadman tiling
Thanksgiving

Thanksgiving lunch The Torrey Pines Lodge turkey wine

Since it was just the two-ish of us, Jes and I went to the Lodge for Thanksgiving lunch.

Pegboard backpack organizer

Clearing out the upstairs office meant there was some spare pegboard floating around. I replaced a crowded set of coat hooks with the more functional but less attractive paneling.

Olympus Has Fallen HBO dog weimaraner media room

After several instances of incorrect items (in a time crunch) and rather disappointing streaming availability, I kicked Prime to the curb. It's covid, so I replaced vanilla Prime HBO with HBO Max.
The wheel


Wheel trading (bouncing between cash-secured puts and covered calls) has been a fun hobby with better profit potential than GME YOLOs. The story of the above:
PUBG un-RIP(?)



A few months back I eulogized PUBG. Turns out, covid combined with a lack of attractive alternatives brought the squad back. Not much has changed in the game, we've learned to live with the bots and appreciate the game's marginally better stability.
Neural style transfer, but TensorFlow this time

I meandered back into neural style transfer over football. I'd last left it with my DL4J experimentation, which notably didn't have hardware acceleration. Since the Keras sample code was easy to hit 'go' on, I tried that.

Neural style transfer algorithm Valentino Rossi Ducati Laguna Seca Ralph Steadman content style

For this exercise I went with:

Same same


Naturally, a 1060 is much quicker than a Core i7. But I was treated to largely the same results as before.




The algorithm tends toward wavy lines in areas of low detail and seems to produce similar images regardless of style. Unlike the DL4J code that required VGG19-size (224x224) images, this one does scaling, for better or worse.

Scaling up

It didn't take long to modify the example to do tiles of a full res image. The tile boundaries are obvious and could be fixed inelegantly with photoshop and elegantly with feathering and staggering. I also found a blog post (whose url I have since lost) that recommended a few things:
Deep learning neural style transfer algorithm sample Maison Pour Erotomane example

As long as I was tiling the content/output, it made sense to apply another lesson and sample various portions of the style image. That is, the naive method is to take two 224x224 images and combine them, so you have crop/scale both content and style images to a small square. Scaling down means that whatever style you have quickly becomes lost - e.g. a 10x13 brushstroke make be condensed to 2x3. Cropping the style image means you're only looking at a portion of the art - so your Maison Pour Erotomane may be all car and no horse.

So my next revision applied a random 224x224 style square to each tile of the content.

Implementing an outer loop meant working out something that I hadn't yet done - loading incremental output. I think I accomplished this by initializing the initial guess field of fmin_l_bfgs_b() to the last value rather than the content. The algorithm still computes content/style loss from the original images, but can now be checkpointed.


Each progressive iteration shows more and more artistic stylization that quickly becomes pretty abstract. You also see the hard tile boundaries soften as I introduced a small translation to each input value.

You can also see the entire image bounce between styles as the random sample from the Steadman image changes from one section to the next. This, of course means progressive iterations move between 'substyles' of the style image. Numerically, it means trying to find the global minimum on a moving target. From a graphic art perspective, this creates a set of image variations that can be manually or automatically blended to a final product.

Style match

Neural style transfer Rossi Ducati Laguna Seca Ralph Steadman tiling style match

The thought occurred to me that there are tiles in the style image that might be more appropriate for a given content tile. Heuristics come to mind, i.e. selecting a tile from the style image based on its fitness for the content; matching color, matching contrast, etc. Ultimately though, it seemed like the easiest and best(?) approach would be to let loss make that determination. The optimization of each tile would sample a number of sources and only retain best results.

Neural style transfer Ralph Steadman tiling Hunter Thompson Ducati

In the spirit of increasing the number of 'style matches' for a given square of content, I added a couple more Steadmans to the random sampling. The output looked like this:

   Processing tile: ( 0 ,  0 ) at ( 16 ,  0 )
      Style source  0
      Loss:  203719060.0
      Using result
      Style source  1
      Loss:  200921020.0
      Using result
      Style source  2
      Loss:  164412600.0
      Using result
      Style source  3
      Loss:  197183180.0
      Style source  4
      Loss:  155275980.0
      Using result
   Processing tile: ( 0 ,  1 ) at ( 16 ,  238 )
      Style source  0
      Loss:  249554060.0
      Using result
      Style source  1
      Loss:  214680510.0
      Using result
      Style source  2
      Loss:  179050400.0
      Using result
      Style source  3
      Loss:  198226380.0
      Style source  4
      Loss:  183831140.0

Neural style transfer Valentino Rossi Ducati Laguna Seca Ralph Steadman tiling

And like that, football was over. I ran the code enough to see a different-but-consistent style applied to my image set. The output is ultimately beholden to iteration count and content/style weight hyperparameters.

Deep learning neural style transfer algorithm Broncos content style

Deep learning neural style transfer algorithm Broncos Hinton handoff tiling

The code, as it is right now:

'''Neural style transfer with Keras.  Modified as an experiment.
# References
    - [A Neural Algorithm of Artistic Style](http://arxiv.org/abs/1508.
    06576)
'''

from __future__ import print_function
from keras.preprocessing.image import load_img, save_img, img_to_array,
array_to_img
import numpy as np
from scipy.optimize import fmin_l_bfgs_b
import time
import random
import glob
import argparse
import os.path
from os import path

from keras.applications import vgg19
from keras import backend as K


def preprocess_image(img):
    img = img_to_array(img)
    img = np.expand_dims(img, axis=0)
    img = vgg19.preprocess_input(img)
    return img    


def deprocess_image(x):
    if K.image_data_format() == 'channels_first':
        x = x.reshape((3, side_length, side_length))
        x = x.transpose((1, 2, 0))
    else:
        x = x.reshape((side_length, side_length, 3))
    # Remove zero-center by mean pixel
    x[:, :, 0] += 103.939
    x[:, :, 1] += 116.779
    x[:, :, 2] += 123.68
    # 'BGR'->'RGB'
    x = x[:, :, ::-1]
    x = np.clip(x, 0, 255).astype('uint8')
    return x


def gram_matrix(x):
    assert K.ndim(x) == 3
    if K.image_data_format() == 'channels_first':
        features = K.batch_flatten(x)
    else:
        features = K.batch_flatten(K.permute_dimensions(x, (2, 0, 1)))
    gram = K.dot(features, K.transpose(features))
    return gram

# the "style loss" is designed to maintain
# the style of the reference image in the generated image.
# It is based on the gram matrices (which capture style) of
# feature maps from the style reference image
# and from the generated image


def style_loss(style, combination):
    assert K.ndim(style) == 3
    assert K.ndim(combination) == 3
    S = gram_matrix(style)
    C = gram_matrix(combination)
    channels = 3
    size = side_length * side_length
    return K.sum(K.square(S - C)) / (4.0 * (channels ** 2) * (size ** 2))


# an auxiliary loss function
# designed to maintain the "content" of the
# base image in the generated image

def content_loss(base, combination):
    return K.sum(K.square(combination - base))


# edge detector - sum edges will be how busy it will look
def total_variation_loss(x):
    assert K.ndim(x) == 4
    if K.image_data_format() == 'channels_first':
        a = K.square(
            x[:, :, :side_length - 1, :side_length - 1] - x[:, :, 1:, :
            side_length - 1])
        b = K.square(
            x[:, :, :side_length - 1, :side_length - 1] - x[:, :, :
            side_length - 1, 1:])
    else:
        a = K.square(
            x[:, :side_length - 1, :side_length - 1, :] - x[:, 1:, :
            side_length - 1, :])
        b = K.square(
            x[:, :side_length - 1, :side_length - 1, :] - x[:, :
            side_length - 1, 1:, :])

    return K.sum(K.pow(a + b, 1.25))

def fidelity_loss(x, y):
    assert K.ndim(x) == 3
    assert K.ndim(y) == 3
    if K.image_data_format() == 'channels_first':
        x_g = K.sum(x[:3, :, :])
        y_g = K.sum(y[:3, :, :])
        return K.square(x_g - y_g)
    else:
        x_g = K.sum(x[:, :, :3])
        y_g = K.sum(y[:, :, :3])
        return K.square(x_g - y_g)

    # Experiment with luminance
    #if K.image_data_format() == 'channels_first':
    #    x_g = np.dot(x[0, :3, :, :], [0.2989, 0.5870, 0.1140])
    #    y_g = np.dot(y[0, :3, :, :], [0.2989, 0.5870, 0.1140])
    #    return K.square(x_g - y_g)
    #else:
    #    x_g = np.dot(x[0, :, :, :3], [0.2989, 0.5870, 0.1140])
    #    y_g = np.dot(y[0, :, :, :3], [0.2989, 0.5870, 0.1140])
    #    return K.square(x_g - y_g)


# Returns style layers - this is the default (all), I experimented with
dropping random ones
def get_random_style_layers():
    return ['block1_conv1', 'block2_conv1', 'block3_conv1', 'block4_conv1',
    'block5_conv1']


def get_random_crop(image, width, height):
    '''Returns a random subimage with the given width and height.'''    
    
    # Crop to width and height, if specified.
    x = 0
    if width is not None and width != image.width:
        x = int(random.randrange(image.width - width))
    y = 0
    if height is not None and height != image.height:
        y = int(random.randrange(image.height - height))

    if x != 0 or y != 0:
        box = (x, y, x + width, y + width)
        image = image.crop(box)
    return image

def eval_loss_and_grads(x):
    if K.image_data_format() == 'channels_first':
        x = x.reshape((1, 3, side_length, side_length))
    else:
        x = x.reshape((1, side_length, side_length, 3))
    outs = f_outputs([x])
    loss_value = outs[0]
    if len(outs[1:]) == 1:
        grad_values = outs[1].flatten().astype('float64')
    else:
        grad_values = np.array(outs[1:]).flatten().astype('float64')
    return loss_value, grad_values

def random_style_tile():
    image = style_images[random.randrange(len(style_images))]
    return get_random_crop(image, side_length, side_length)


class Evaluator(object):

    def __init__(self):
        self.loss_value = None
        self.grads_values = None

    def loss(self, x):
        assert self.loss_value is None
        loss_value, grad_values = eval_loss_and_grads(x)
        self.loss_value = loss_value
        self.grad_values = grad_values
        return self.loss_value

    def grads(self, x):
        assert self.loss_value is not None
        grad_values = np.copy(self.grad_values)
        self.loss_value = None
        self.grad_values = None
        return grad_values

side_length = 224           # VGG19 is 224x224x3

# Iteration hyperparameters (modify these)
iterations_per_image = 10   # Number of image traversals
samples_per_tile = 3        # Number of style tiles to try per iteration
iterations_per_sample = 10  # Number of style transfer iterations per
sample

# Use all png files from the style subdirectory, random 224 squares will
be used
# from these files to perform style transfer - so image size should be
# approximately the dimensions of the content.
style_files = glob.glob('style/*.png')

file_id = random.randrange(696969)

# Load content.png, the size of this image will significantly impact run
time.
content_image = load_img('content.png')
content_width, content_height = content_image.size

# Load style images.
style_images = []
for style_name in style_files:
    style_image = load_img(style_name)
    style_images.append(style_image)

# If this setup was run previously, use the existing output image.
if (os.path.isfile('last.png')):
    output_image = load_img('last.png')
else:
    output_image = load_img('content.png')

# Compute the tile count/step size.  There will be overlap and it should
# be a good thing.
x_tiles = int(content_width / side_length)
if (content_width % side_length != 0):
    x_tiles += 1
x_step = int(content_width / x_tiles)

y_tiles = int(content_height / side_length)
if (content_height % side_length != 0):
    y_tiles += 1
y_step = int(content_height / y_tiles)

feature_layers = get_random_style_layers()
print('Style layers: ' + str(feature_layers))

# Number of times to cover the entire image (optimize each tile)
for image_iteration in range(iterations_per_image):
    print('Iteration: ', image_iteration)

    # Randomize hyperparameters because I don't know what good values are.
    # Modify these/make them fixed.
    total_variation_weight = random.uniform(0.001, 0.5)
    style_weight = random.uniform(0.001, 0.5)
    content_weight = random.uniform(0.001, 0.5)
    fidelity_weight = random.uniform(0.001, 0.5)

    # Bump the tile a random value to do a smoother stitch.
    x_jitter = int(random.randrange(-16, 16))
    y_jitter = int(random.randrange(-16, 16))

    # Iterate over each image tile.
    for x_tile in range(x_tiles):
        for y_tile in range (y_tiles):
            x_start = (x_tile * x_step) + x_jitter
            y_start = (y_tile * y_step) + y_jitter

            # Post-jitter boundary check.
            if (x_start < 0):
                x_start = 0
            if (y_start < 0):
                y_start = 0
            if (x_start + side_length > content_width):
                x_start = content_width - side_length - random.randrange(1,
                16)
            if (y_start + side_length > content_height):
                y_start = content_height - side_length - random.
                randrange(1, 16)

            print('   Processing tile: (', x_tile, ', ', str(y_tile), ')
            at (', x_start, ', ', y_start, ')')

            box = (x_start, y_start, x_start + side_length, y_start +
            side_length)
            tile_content = content_image.crop(box)
            base_image = K.variable(preprocess_image(tile_content))

            best_loss = -1

            # For each tile, sample the random portions of the style
            image(s) and choose
            # the best of the lot.
            for i in range(samples_per_tile):
                print('      Trying tile ', i)
                tile_style = random_style_tile()
                
                evaluator = Evaluator()
                tile_output = output_image.crop(box)

                # run scipy-based optimization (L-BFGS) over the pixels of
                the generated image
                # so as to minimize the neural style loss
                x = preprocess_image(tile_output)

                style_reference_image = K.
                variable(preprocess_image(tile_style))

                if K.image_data_format() == 'channels_first':
                    combination_image = K.placeholder((1, 3, side_length,
                    side_length))
                else:
                    combination_image = K.placeholder((1, side_length,
                    side_length, 3))
                
                # combine the 3 images into a single Keras tensor
                input_tensor = K.concatenate([base_image,
                                              style_reference_image,
                                              combination_image], axis=0)

                # Reinitialize VGG19.  There's probably a way to do this
                once and
                # improve performance.
                model = vgg19.VGG19(input_tensor=input_tensor, weights=
                'imagenet', include_top=False)

                outputs_dict = dict([(layer.name, layer.output) for layer
                in model.layers])

                # combine these loss functions into a single scalar
                loss = K.variable(0.0)
                layer_features = outputs_dict['block5_conv2']
                base_image_features = layer_features[0, :, :, :]
                combination_features = layer_features[2, :, :, :]
                loss = loss + content_weight *
                content_loss(base_image_features, combination_features)

                for layer_name in feature_layers:
                    layer_features = outputs_dict[layer_name]
                    style_reference_features = layer_features[1, :, :, :]
                    combination_features = layer_features[2, :, :, :]
                    loss = loss + (style_weight / len(feature_layers)) *
                    style_loss(style_reference_features,
                    combination_features)
                loss = loss + total_variation_weight *
                total_variation_loss(combination_image)

                loss = loss + fidelity_weight *
                fidelity_loss(combination_features, base_image_features)

                # get the gradients of the generated image wrt the loss
                grads = K.gradients(loss, combination_image)

                outputs = [loss]
                if isinstance(grads, (list, tuple)):
                    outputs += grads
                else:
                    outputs.append(grads)

                f_outputs = K.function([combination_image], outputs)

                # With this content/style combo, use the style transfer
                algorithm.
                for i in range(iterations_per_sample):
                    x, min_val, info = fmin_l_bfgs_b(evaluator.loss, x.
                    flatten(),
                                                     fprime=evaluator.
                                                     grads, maxfun=20)
                    print('        Loss[', i, ']: ', min_val)

                # If this is the first style sample or it's better than
                the last
                # use this output.
                if (best_loss == -1 or min_val < best_loss):
                    print('      Updating from prev: ', min_val, ' < ',
                    best_loss)
                    img = array_to_img(deprocess_image(x.copy()))
                    output_image.paste(img, (x_start, y_start))
                    best_loss = min_val

                # Reset the back end
                K.clear_session()
                        
    fname = 'output_' + str(file_id) + '_%d.png' % image_iteration
    print('   ->', fname)
    save_img(fname, output_image)
    save_img('last.png', output_image)
Fantasy

Deep learning neural style transfer algorithm Broncos Hinton handoff

25-8 on aggregate going into this wild weekend of Jacksonless Ravens, QBless Donkeys, and a Thursday-Tuesday weekend.

Week d'san andreas da bears
- Medieval Gridiron - 		Covid-20
- Password is Taco - 		Dominicas
- Siren -
1 Danville Isotopes
110.8 - 72.5 W (1-0) 		Black Cat Cowboys
155.66 - 78.36 W (1-0) 		TeamNeverSkipLegDay
136.24 - 107.50 W (1-0)
2 Screaming Goat Battering Rams
119.9 - 105.9 W (2-0) 		[Random UTF characters resembling an EQ]
115.50 - 115.74 L (1-1) 		Dem' Arby's Boyz
94.28 - 102.02 L (1-1)
3 Nogales Chicle
106.5 - 117.8 L (2-1) 		Circle the Wagons
100.42 - 90.02 W (2-1) 		JoeExotic'sPrisonOil
127.90 - 69.70 W (2-1)
4 Britons Longbowmen
122.9 - 105.1 W (3-1) 		Staying at Mahomes
123.28 - 72.90 W (3-1) 		Daaaaaaaang
138.10 - 108.00 W (3-1)
5 Toronto Tanto
105.0 - 108.2 L (3-2) 		Robocop's Posse
111.32 - 134.26 L (3-2) 		Alpha Males
86.20 - 76.12 W (4-1)
6 Only Those Who Stand
108.2 - 66.7 W (4-2) 		KickAssGreenNinja
65.10 - 84.02 L (3-3) 		SlideCode #Jab
71.60 - 53.32 W (5-1)
7 San Francisco Seduction
121.7 - 126.4 L (4-3) 		Ma ma ma my Corona
118.22 - 84.20 W (4-3) 		G's Unit
109.20 - 92.46 W (6-1)
8 LA Boiling Hot Tar
116.2 - 59.4 W (5-3) 		Kamaravirus
118.34 - 109.94 W (5-3) 		WeaponX
113.14 - 85.40 W (7-1)
9 SD The Rapier
135.0 - 90.8 W (6-3) 		C. UNONEUVE
117.80 - 90.16 W (6-3) 		Chu Fast Chu Furious
128.28 - 59.06 W (8-1)
10 West Grove Wankers
72.9 - 122.8 L (6-4) 		Pug Runners
98.90 - 77.46 W (7-3) 		NY Giants LARP
75.24 - 75.06 W (9-1)
11 SF Lokovirus
127.9 - 87.1 W (7-4) 		Bravo Zulus
116.34 - 45.50 W (8-3) 		HitMeBradyOneMoTime
107.42 - 89.22 W (10-1)
12 Danville Isotopes
154.7 - 64.1* 		Forget the Titans
57.04 - 99.74* 		TeamNeverSkipLegDay
122.68- 91.12*

Dog weimaraner label maker labeling indignant


This one time, my buddy Keith, on a DARE, got a tattoo: "I'm a moron" right across his forehead, man. 'Course, he made two hundred bucks off that, so...you ask yourself: Who's the REAL moron, huh?

?




◄
2020.11.15

Decisionvirus

 Election day, commentary, and fallout. The Nioh journey comes to an end. A sprinkling of BL3, Among Us, Risk of Rain.
2020.12.06

Edges and corners

 Taking the 500mm out for some surf shots. Tweaking neural style transfer. 		►


Related / internal

Some posts from this site with similar content.

Post
2020.12.06

Edges and corners

 Taking the 500mm out for some surf shots. Tweaking neural style transfer.
Post
2020.12.12

On lock

 The covid surge that everyone expected after Thanksgiving has hit. Jes is busy at work. I can get by with games, streaming, jogging, and taking the dog out.
Post
2021.01.02

The next break

 Reflecting on investments and Warren B. Some sunset surf shots, video games, and AI image stylization.

Related / external

Risky click advisory: these links are produced algorithmically from a crawl of the subsurface web (and some select mainstream web). I haven't personally looked at them or checked them for quality, decency, or sanity. None of these links are promoted, sponsored, or affiliated with this site. For more information, see this post.

Has a preview image link and yet 404 :/
pyimagesearch.com

PyTorch: Training your first Convolutional Neural Network (CNN) - PyImageSearch

 In this tutorial, you will receive a gentle introduction to training your first Convolutional Neural Network (CNN) using the PyTorch deep learning library.
machinethink.net

One-stage object detection

 An in-depth look at how fast object detection models are trained
thenumb.at

Differentiable Programming from Scratch

Created 2024.10 from an index of 433,370 pages.