Gallerypost | 2020.12.06

surf surfing San Diego Blacks Beach focus
December swell

Surfing San Diego Blacks Beach 500mm Nikon telephoto crowds

The first week of December almost always brings one of the biggest swells of the year. This one wasn't epic, but good enough to take a vacation Friday for some 0730 shooting and 0930 surfing. The lead image is a pretty good depiction of the difference between pre-sunrise shots and early-sunrise shots. The Black's cliffs make for a pretty tight window for good lighting.

San Diego Blacks Beach surfing surfers big day overhead crowds

All those dots in the lens photo aren't misleading, Black's was a zoo from road to path. This - combined with the crazy closeouts - made the photography actually quite difficult; it was hard to tell who would make the drop and the waves didn't hold shape for long.

surf surfing San Diego Blacks Beach La Jolla Scripps Pier

I wanted to use Scripps Pier on an earlier shoot to get some background features. Turns out the field of view on the 500 is so tight I could get La Jolla as a backdrop if I shot down the line.

Blacks beach surfing telephoto San Diego surf photography Blacks beach surfing telephoto San Diego surf photography Blacks beach surfing telephoto San Diego surf photography
Blacks beach surfing telephoto San Diego surf photography Blacks beach surfing telephoto San Diego surf photography Blacks beach surfing telephoto San Diego surf photography Blacks beach surfing telephoto San Diego surf photography
Blacks beach surfing telephoto San Diego surf photography Blacks beach surfing telephoto San Diego surf photography Blacks beach surfing telephoto San Diego surf photography
Blacks beach surfing telephoto San Diego surf photography Blacks beach surfing telephoto San Diego surf photography

surf surfing San Diego Blacks Beach overhead takeoff
Neural style transfer

Last month I experimented some with the Tensorflow neural style transfer sample. Changes included tiling the content image and randomly sampling the style image(s).

I addressed the tile/seam issues by generating A/B outputs that have overlapped tiles that can be automatically or manually blended as a final step.

Compared to hard tile edges, even a straight overlay shows more style cohesion. With a little feathering, they stitch together nicely.

Last time I postulated that enough sampling would find an ideal tile from the style image. My implementation came up somewhat short as it didn't retain the last iteration's work - it would choose the best tile from a random sampling but then start over on the next full-image iteration. The previous work would be incorporated in the output, but the ideal(?) approach would be to continuously optimize the current best while still sampling other options.

Valentino Rossi Ducati Laguna Seca Hunter S Thompson Ralph Steadman neural style transfer tiling stitch style alignment

Rather than track coordinates and loss, I decided to save off style squares containing the best style sample for a given content tile. Each pass of a given tile would start with the saved image and optimize it further, then see if other samples can achieve a lower loss value. Conceptually, this isn't too far off of standard error descent methods; I'm optimizing my best guess while checking random samples in case I'm not at the global minimum.
Cautious theta gang

I'm hesitant to write any puts that are longer than a week out. Still, 20k to make 200 in a week is better than a muni. It's a weird time:
Bot league

PUBG stats report bots bot count

Between being terrible, having a new squadmate, and low player counts, MMR has put us in bot league. At least the final circles have a few human players, but we've hit a bit of a goldilocks situation between unfairly hard and unfairly easy. Maybe that will change when the season rolls over.

Kshot linked a few cool things. There's a site to watch yourself get killed by streamers or, less frequently, enjoy their disgust when you knock them. And PUBG Lookup has a lot of cool match metrics, including show how many real players were fighting for that chicken dinner.

Jon put together quick clips of a physics corner case (that really isn't a corner case).

Cat eyes

Cat photography catch lights

Koko sent me a cat photo. I'm not sure if it was an iphone effect or her lights or her camera case, but it had a neat catch light effect.
Updated style transfer code

'''Neural style transfer with Keras.  Modified as an experiment.
# References
    - [A Neural Algorithm of Artistic Style](

from __future__ import print_function
from keras.preprocessing.image import load_img, save_img, img_to_array,
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))
        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)
        features = K.batch_flatten(K.permute_dimensions(x, (2, 0, 1)))
    gram =, 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:])
        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)
        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 =[0, :3, :, :], [0.2989, 0.5870, 0.1140])
    #    y_g =[0, :3, :, :], [0.2989, 0.5870, 0.1140])
    #    return K.square(x_g - y_g)
    #    x_g =[0, :, :, :3], [0.2989, 0.5870, 0.1140])
    #    y_g =[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',

def get_random_crop(image, width, height):
    '''Returns a random subimage with the given width and height.'''    
    if (image is None or
        width is None or width > image.width or
        height is None or height > image.height):

    x_vals = image.width - width
    y_vals = image.height - height
    if (x_vals < 0 or y_vals < 0):
    # Crop to width and height, if specified.
    x = 0
    if (x_vals > 0):
        x = int(random.randrange(x_vals))
    y = 0
    if (y_vals > 0):
        y = int(random.randrange(y_vals))

    print('Rand: ',str(x),', ',str(y),' from ',str(x_vals),' by ',
    box = (x, y, x + width, y + width)
    crop = image.crop(box)

    return crop
def eval_loss_and_grads(x):
    if K.image_data_format() == 'channels_first':
        x = x.reshape((1, 3, side_length, side_length))
        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')
        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
step = 168

# Iteration hyperparameters (modify these)
iterations_per_image = 32   # Number of image traversals
samples_per_tile = 5        # Number of style tiles to try per iteration
iterations_per_sample = 16  # Number of style transfer iterations per

# 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
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)

# If this setup was run previously, use the existing output image.
if (os.path.isfile('last.png')):
    output_image_a = load_img('last_a.png')
    output_image_b = load_img('last_b.png')
    output_image_a = load_img('content.png')
    output_image_b = 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 / step)
y_tiles = int(content_height / step)

# Add iterleaved tiles
tiles = []
x_start = 0
for i in range(x_tiles):
    y_start = 0
    for j in range(y_tiles):
        if (i + j) % 2 == 0:
            tiles += [(x_start, y_start, True)]
            tiles += [(x_start, y_start, False)]
        y_start = y_start + step       
    x_start = x_start + step

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

total_variation_weight = random.uniform(0.001, 0.1)
style_weight = random.uniform(0.001, 0.1)
content_weight = random.uniform(0.0001, 0.1)
fidelity_weight = random.uniform(0.001, 0.1)

# 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.
    fidelity_weight *= 2
    content_weight *= 2

    # Bump the tile a random value to do a smoother stitch.
    fname_a = 'output_' + str(file_id) + '_%d_a.png' % image_iteration
    fname_b = 'output_' + str(file_id) + '_%d_b.png' % image_iteration

    # Iterate over each image tile.
    for tile in tiles:
        x_start = tile[0]
        y_start = tile[1]
        is_a = tile[2]
        print('   Processing tile: (', x_start, ', ', y_start, ')')
        style_file = 'style_' + str(x_start) + '_' + str(y_start) + '.png'
        box = (x_start, y_start, x_start + side_length, y_start +
        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 sample_index in range(samples_per_tile):
            # Set baseline error with current best style sample
            if (sample_index == 0 and os.path.isfile(style_file)):
                print('      Using existing style: ',style_file)
                tile_style = load_img(style_file)
                using_best = True
                print('      Using random tile')
                tile_style = random_style_tile()
                using_best = False
            evaluator = Evaluator()

            if (is_a):
                tile_output = output_image_a.crop(box)
                tile_output = output_image_b.crop(box)
            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,
                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,
                                          combination_image], axis=0)

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

            outputs_dict = dict([(, layer.output) for layer in

            # 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 *

            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

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

            # Use optimization iterations hyperparameter, for the current
            # just do half as many.
            iterations = iterations_per_sample
            if (using_best == True):
                iterations = int(iterations / 2)
            # With this content/style combo, use the style transfer
            for iteration in range(iterations):
                x, min_val, info = fmin_l_bfgs_b(evaluator.loss, x.
                if (min_val > 1000000):
                    min_str = str(int(min_val / 1000000)) + 'M'
                elif (min_val > 1000):
                    min_str = str(int(min_val / 1000)) + 'k'
                    min_str = str(int(min_val))
                print('        Loss[', iteration, ']: ', min_str)

            # If this is the first style sample or it's better than the
            # use this output.
            if (best_loss == -1 or min_val < best_loss):
                print('      Updating from prev: ', min_val / 1000000, 'M
                < ', best_loss / 1000000, 'M')
                img = array_to_img(deprocess_image(x.copy()))
                best_loss = min_val

                if (is_a):
                    output_image_a.paste(img, (x_start, y_start))
                    output_image_b.paste(img, (x_start, y_start))
                save_img(style_file, tile_style)
                print('      Updated style file: ', style_file)
                print('      Skipping update: ', min_val / 1000000, 'M > ',
                best_loss / 1000000, 'M')

            # Reset the back end

        # Save per-tile progress for long runs
        if (samples_per_tile * iterations_per_sample > 64):
            save_img(fname_a, output_image_a)
            save_img(fname_b, output_image_b)
            save_img('last_a.png', output_image_a)
            save_img('last_b.png', output_image_b)
    save_img(fname_a, output_image_a)
    save_img(fname_b, output_image_b)
    save_img('last_a.png', output_image_a)
    save_img('last_b.png', output_image_b)

Related - internal

Some posts from this site with similar content.



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

The next break

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


After some interesting reads, I implemented a convolution+pooling block inspired by ResNet. It looks like this:

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.


Using JAX, numpy, and optimization techniques to improve separable image filters | Bart Wronski

By the end of this blog post, we aim to improve the results of my previous blog post of approximating circular bokeh with a sum of 4 separable filters and reducing its artifacts (left) through an offline optimization process to get the results on the right. In today's blog post I will look at two...

Time Series for scikit-learn People (Part III): Horizon Optimization | Ethan Rosenthal

In my previous posts in the "time series for scikit-learn people" series, I discussed how one can train a machine learning model to predict the next element in a time series. Often, one may want to predict the value of the time series further in the future. In those posts, I gave two methods to accomplish this. One method is to train the machine learning model to specifically predict that point in the future.

A Visual Exploration of Random Neural Networks

Created 2024.06 from an index of 272,339 pages.