how to generate deep dream data from own Images ?
2 vues (au cours des 30 derniers jours)
Afficher commentaires plus anciens
Selva Karna
le 9 Sep 2017
Réponse apportée : voxey
le 4 Avr 2018
how to generate deep dream data from own Images ?
0 commentaires
Réponse acceptée
Image Analyst
le 9 Sep 2017
Find an article on it and code it up, because there is no built in function for that.
Plus de réponses (9)
Delprat Sebastien
le 4 Jan 2018
Modifié(e) : Star Strider
le 4 Jan 2018
clear all;
close all;
clc;
% ===============================
% You will need to download the net using the Matlab app finder tool
% ===============================
No=1;
switch No
case 1
net=alexnet;
netname='AlexNet';
layer = 23; % Ouput layer before classification
case 2
net=vgg16;
netname='VGG16';
layer = 39; % Ouput layer before classification
end
InputSize=net.Layers(1).InputSize;
fprintf('==============================\n');
fprintf('Network layers\n');
net.Layers
fprintf('==============================\n');
fprintf('Network classes\n');
for i=1:length(net.Layers(end).ClassNames)
fprintf('%i - %s\n',i,net.Layers(end).ClassNames{i});
end
% ==============================================
% Your image to be used as a starting dream
% ==============================================
FileName='Poisson';
initImg=imread([FileName '.jpg']);
initImg=imresize(initImg,InputSize(1:2));
[label,scores] = classify(net,initImg);
% Display network output, just for info (display the 5 more probable
% classes)
[~,idx] = sort(scores,'descend');
idx = idx(5:-1:1);
classNames = net.Layers(end).ClassNames;
classNamesTop = classNames(idx);
scoreTop = scores(idx);
fprintf('Most probable classes:\n');
for i=1:length(idx)
fprintf('%i - %s - %f\n',i,classNamesTop{i},scoreTop(i));
end
% Display initial image
figure;
imshow(initImg)
channels = [208]; % Golden retriever
channels = 53; % Thunder snake
channels = 481; % Cash machine
channels = 158; % Papillon
channels = 162; % Basset
channels = 328; % Starfish
channels = 818; % sports car
fprintf('Targeted class: %s\n',net.Layers(end).ClassNames{channels});
levels = 1; % 1 or 2 is good
iterations = 20; % higher better, but also longer
I = deepDreamImage(net,layer,channels, ...
'Verbose',true, ...
'NumIterations',iterations, ...
'PyramidLevels',levels,...
'InitialImage',initImg);% 'PyramidScale',1.2,...
% Save result and built file name using algorithm config
Classname=net.Layers(end).ClassNames{channels};
name=sprintf('dreem_%s_%s_chl_%i_iter_%i_lvl_%i_%s',netname,FileName,channels,iterations,levels,Classname);
imwrite(I,[name '.jpg']);
% Display result
figure
imshow(I)
Original image:
Dreamed image (fish interpreted as a snake)
2 commentaires
Selva Karna
le 29 Mar 2018
Modifié(e) : Selva Karna
le 29 Mar 2018
6 commentaires
Image Analyst
le 31 Mar 2018
Alexnet is built assuming that size. If you want different size you'd have to use a different network. However, you will probably find it does just fine with resizing your images to that. It will most likely still work with your images.
Delprat Sebastien
le 1 Avr 2018
Please look at the following code https://fr.mathworks.com/matlabcentral/fileexchange/66710-deep-dreaming-animation-for-matlab
There is : 1) a required modification to add clipping to the original matlab code 2) a script that generate an animation from image (with resizing) using recursive deepdream
NB : please read carefully the patch.m file... Enjoy
0 commentaires
Selva Karna
le 2 Avr 2018
Modifié(e) : Image Analyst
le 2 Avr 2018
2 commentaires
Delprat Sebastien
le 2 Avr 2018
Modifié(e) : Delprat Sebastien
le 2 Avr 2018
If you do not copy the error in detail, I cannot help. Did you copy the deepDreamImageLaplacianNorm as deepDreamImageLaplacianNorm2 in the specified matlab toolbox folder?
Delprat Sebastien
le 3 Avr 2018
In the deepDreamImageLaplacianNorm2.m you must not comment the gradient ascend % X = X + gradient * stepSize; X=max(XMin2,min(X,XMax2));
=> should be (mind the % removal) X = X + gradient * stepSize; X=max(XMin2,min(X,XMax2));
Delprat Sebastien
le 3 Avr 2018
Unfortunately, due to Mathworks copyright, I cannot disclose the source code of their functions. If you follow the steps detailed in the patch. M files, this should be OK.
In your previous code, you did add a comment on the original line x = x + g*stepsize. Remove this comment and it should work.
2 commentaires
Voir également
Catégories
En savoir plus sur Get Started with MATLAB dans Help Center et File Exchange
Community Treasure Hunt
Find the treasures in MATLAB Central and discover how the community can help you!
Start Hunting!