htoyryla

-- modified version of @jcjohnson's neural-style
-- by @htoyryla 13 Feb 2018
-- allows giving emphasis to nc best channel(s) in each style layer
-- use -style_layers to select layer as usual, using a single layer is recommended
-- -nc to set how many of the best channels are used per layer
-- during target capture, tests the model using the style image
-- and selects nc channels with strongest activations to be given emphasis during iterations
-- not tested with multiple style images

require 'torch'

htoyryla

# git clone from https://github.com/tkarras/progressive_growing_of_gans
# download the snapshot from their Google drive
# use the following code in the same directory to generate random faces
import os
import sys
import time
import glob
import shutil
import operator
import theano

htoyryla

-- modified version of @jcjohnson's neural-style
-- by @htoyryla 14 Feb 2018
-- allows exploring the effect of equalizing the activations of different channels in a style layer


require 'torch'
require 'nn'
require 'image'
require 'optim'

htoyryla

-- this program takes in an image
-- and finds nc channels on a given layer
-- having the strongest activations 

require 'torch'
require 'nn'
require 'image'
require 'loadcaffe' 

function preprocess(img)

htoyryla

-- neural-style modified by @htoyryla 27 Oct 2017
-- to use both gram matrix (as a statistical evaluation) and a mean of n featuremaps (to retain spatial sturcture of the style image)
-- to evaluate style
-- StyleLoss and GramMatrix have been copied from fast-neural-style
-- but modified to calculate a spatial mean of the feature maps 
-- new param mean_weight to control the amount of mean for style 
-- loss_type to select between L2 and SmoothL1 (as in fast-neural-style) 
-- **** style_scale no longer works, the mean calculation requires that style image is resized as the content image ****

htoyryla

-- this is a modified version of neural_style.lua originally by @jcjohnson
-- @htoyryla 24 Oct 2017
--
-- original neural-style includes scaling of gram matrix output and gradient in the StyleLoss
-- using division by the number of elements in input
-- this is done to make the style representations from different layers (of different size) better comparable
-- but it also has the effect of decreasing gradients as the image size is increased
--
-- this version attempts to keep the gradients from style loss modules in the same range 
-- (i.e. that what is obtained at image size 512px)

htoyryla

-- modified version of @jcjohnson's neural-style
-- by @htoyryla 14 May 2017
-- allows exploring th effect of giving emphasis to specific channel(s) in a style layer
-- use -style_layers to select layer as usual, using a single layer is recommended
-- -style_channels to specify one or more channels to be given emphasis, other channels will be attenuated
-- if a channel does not exist in a layer, it is ignored and given a warning
-- note: if multiple style layers are used, style_channels setting affects both

require 'torch'
require 'nn'

htoyryla

-- neural-style by jcjohnson modified by @htoyryla
-- 13 May 2017
-- generate style variants by modifying style target Gram matrix with randomized eigenvalues
-- see function randomizeEigenvalues() for details

require 'torch'
require 'nn'
require 'image'
require 'optim'

htoyryla

-- neural-style by jcjohnson modified by @htoyryla
-- 13 May 2017
-- generate style variants by modifying style target Gram matrix with randomized eigenvalues
-- see function randomizeEigenvalues() for details
-- this version produces slight variants of a given style 
-- by multiplying gram matrix eigenvalues by random multiplies 0.1 to 1.9 


require 'torch'
require 'nn'

htoyryla

--
-- Neural-style modified by Hannu Töyrylä 13 Sep 2016
-- according to the delta gram matrix method 
-- as described in http://arxiv.org/abs/1606.01286v1
--
-- Install and use just like http://github.com/jcjohnson/neural-style/
-- except for the additional parameter delta (use values 0 to 16, too large values can result in error)
--
require 'torch'
require 'nn'