Solid gray image as a result

[guillefar]

So, after many hours running the software, i just got 10 images like this:

The command I run was:

./Software/Git/neural-style/neural_style.lua -style_image ./Imágenes/style.jpg -content_image ./Imágenes/foto.png -output_image ./Imágenes/NeuralStyle_test -model_file ./Software/Git/neural-style/models/vgg_normalised.caffemodel -proto_file ./Software/Git/neural-style/models/VGG_ILSVRC_19_layers_deploy.prototxt -gpu -1 -num_iterations 1000 -seed 123 -content_layers relu0,relu3,relu7,relu12 -style_layers relu0,relu3,relu7,relu12 -content_weight 10 -style_weight 1000 -image_size 512 -optimizer adam

Any ideas?

[jcjohnson]

Try turning off TV regularization: -tv_weight 0

[guillefar]

Ok, will try, thanks.

[Gordakan]

Hi.

I got the same result. Input image is the same size as the style image (468x600). I tried with only 3 iterations (running on cpu) just to give it a try.

Here is the command with arguments used : th neural_style.lua -style_image examples/inputs/picasso_selfport1907.jpg -content_image photo.jpg -output_image photo_styled.jpg -model_file models/vgg_normalised.caffemodel -proto_file models/train_val.prototxt -gpu -1 -backend clnn -num_iterations 3 -seed 123 -content_layers relu0,relu3,relu7,relu12 -style_layers relu0,relu3,relu7,relu12 -content_weight 10 -style_weight 100 -image_size 256 -optimizer adam -print_iter 1 -tv_weight 0

[Flyingvette]

Not even close to enough iterations to do anything. You can flag it to use the source image as the starting point rather than a blank field.

On March 5, 2016 3:59:19 AM PST, Gordakan notifications@github.com wrote:

Hi.

I got the same result. Input image is the same size as the style image (468x600). I tried with only 3 iterations (running on cpu) just to give it a try.

Here is the command with arguments used : th neural_style.lua -style_image examples/inputs/picasso_selfport1907.jpg -content_image photo.jpg -output_image photo_styled.jpg -model_file models/vgg_normalised.caffemodel -proto_file models/train_val.prototxt -gpu -1 -backend clnn -num_iterations 3 -seed 123 -content_layers relu0,relu3,relu7,relu12 -style_layers relu0,relu3,relu7,relu12 -content_weight 10 -style_weight 100 -image_size 256 -optimizer adam -print_iter 1 -tv_weight 0

---

Reply to this email directly or view it on GitHub: https://github.com/jcjohnson/neural-style/issues/149#issuecomment-192626774

[shessenauer]

@Flyingvette How do I tag it as a source image? would I do -init /directoryToImage/image.png?

[htoyryla]

Sam Hesssenauer notifications@github.com kirjoitti 26.4.2016 kello 10.34:

@Flyingvette How do I tag it as a source image?

-init image

As described in the Readme.md:

-init: Method for generating the generated image; one of random or image. Default is random which uses a noise initialization as in the paper; image initializes with the content image.

[StryxZilla]

Has anyone ever gotten this working? Even with tv_weight at 0 I get images that are largely grey versions of the content image @ 1k iterations.

[shaharz]

For the record, I noticed you're using vgg_normalised.caffemodel, which produces a blank image for me as well. On the other hand VGG_ILSVRC_19_layers.caffemodel works just fine!

[winnyec]

It appears to me that the issue is that the size of the layer deck is not #cnn, but #cnn.modules when using the backend nn (in particular, when -gpu -1).

[htoyryla]

I used neural-style extensively with only nn on CPU for almost a year, and never experienced any problem with #cnn. I also experimented with many modifications to the code.

So while loadcaffe appears sometimes to return a model object where #cnn does not give the number of layers, it surely does not depend merely on nn and cpu.

The #cnn question has also been discussed in https://github.com/jcjohnson/neural-style/pull/338

PS. You may still have a point. When I try

require 'loadcaffe'
cnn = loadcaffe.load("models/VGG_ILSVRC_19_layers_deploy.prototxt", "models/VGG_ILSVRC_19_layers.caffemodel", "nn"):float()
print(#cnn)
c = cnn.modules
print(#c)

both give the number of layers correctly, but otherwise cnn and cnn.modules are not identical. The model structure can be printed from cnn

th> print(cnn)
nn.Sequential {
  [input -> (1) -> (2) -> (3) -> (4) -> (5) -> (6) -> (7) -> (8) -> (9) -> (10) -> (11) -> (12) -> (13) -> (14) -> (15) -> (16) -> (17) -> (18) -> (19) -> (20) -> (21) -> (22) -> (23) -> (24) -> (25) -> (26) -> (27) -> (28) -> (29) -> (30) -> (31) -> (32) -> (33) -> (34) -> (35) -> (36) -> (37) -> (38) -> (39) -> (40) -> (41) -> (42) -> (43) -> (44) -> (45) -> (46) -> output]
  (1): nn.SpatialConvolution(3 -> 64, 3x3, 1,1, 1,1)
  (2): nn.ReLU
  (3): nn.SpatialConvolution(64 -> 64, 3x3, 1,1, 1,1)
  (4): nn.ReLU
 ...

but the actual layer stack really is under cnn.modules:

th> print(cnn.modules)
{
  1 :
    {
      padW : 1
      nInputPlane : 3
      output : FloatTensor - empty
      name : "conv1_1"
      _type : "torch.FloatTensor"
      dH : 1
      dW : 1
      nOutputPlane : 64
      padH : 1
      kH : 3
      weight : FloatTensor - size: 64x3x3x3
      gradWeight : FloatTensor - size: 64x3x3x3
      gradInput : FloatTensor - empty
      kW : 3
      bias : FloatTensor - size: 64
      gradBias : FloatTensor - size: 64
    }
  2 :
    {
      inplace : true
      threshold : 0
      _type : "torch.FloatTensor"
      output : FloatTensor - empty
      gradInput : FloatTensor - empty
      name : "relu1_1"
      val : 0
    }
...

So while #cnn appears to work for many, using #cnn.modules is probably safer.

[winnyec]

Sorry I am not an expert at all: I was just looking for this solid grey error, and since I found no help here that worked for me, I started debugging. I have found the key to this fix in the nn package, hence I was assuming that there might be the difference… given also that things are probably working in most of the cases for everybody else.

Anyway: probably it is an Arch thing, since I have Torch and the packages from AUR too like in #338.

[syz825211943]

@winnyec How to fix the nn package?

[winnyec]

@syz825211943: The computer on which I did this is packed up for the time being, so cannot check what exactly I did. But based on my comment above try editing #cnn into #cnn.modules. Apparently #cnn should also work just the same way, but with the versions Arch has (had?) only the explicit #cnn.modules works.

[ballerburg9005]

I am on Archlinux and just used the AUR packages.

It works if you just change #cnn into #cnn.modules in neural_style.lua .