evolving-ai-lab / synthesizing Goto Github PK

I want to visualize my own models, but I don't know how to train the generator.caffemodel.
Can you help me? Thank you~

The provided links for downloading the Upconvolutional network is unavailable: http://lmb.informatik.uni-freiburg.de/resources/software.php appears to be down, meaning its not currently possible to fully install this repository.

I don't know if this unavailability is permanent - and have emailed the original papers' authors - but do you know of any other sources for this?

Thanks for posting the code; it looks really interesting! When I run any of the examples I get this sort of error:

Starting optimizing WARNING: Logging before InitGoogleLogging() is written to STDERR F1227 00:45:44.057626 2037047296 inner_product_layer.cpp:144] Cannot use GPU in CPU-only Caffe: check mode. *** Check failure stack trace: *** ./1_activate_output.sh: line 57: 14235 Abort trap: 6 python ./act_max.py --act_layer ${act_layer} --opt_layer ${opt_layer} --unit ${unit} --xy ${xy} --n_iters ${n_iters} --start_lr ${lr} --end_lr ${end_lr} --L2 ${L2} --seed ${seed} --clip ${clip} --bound ${bound_file} --debug ${debug} --output_dir ${output_dir} --init_file ${init_file} convert: unable to open image output/fc8_0643_200_0.99_8.0__0.jpg': No such file or directory @ error/blob.c/OpenBlob/2701.
convert: no images defined output/fc8_0643_200_0.99_8.0__0.jpg' @ error/convert.c/ConvertImageCommand/3258. convert: unable to open image output/fc8_0643_200_0.99_8.0__0.jpg': No such file or directory @ error/blob.c/OpenBlob/2701.

unit: 624 xy: 0
n_iters: 200
L2: 0.99
start learning rate: 8.0
end learning rate: 1e-10
seed: 0
opt_layer: fc6
act_layer: fc8
init_file: None
clip: 0
bound: act_range/3x/fc6.txt

debug: 0
output dir: output
net weights: nets/caffenet/bvlc_reference_caffenet.caffemodel
net definition: nets/caffenet/caffenet.prototxt

`
Any idea what could be causing it?
Thanks so much!

Hello ! I totally love this code, thank you so much for putting this together. But I just have a question, how to up the output size ? I just bought a very fast computer and I would like to have High rez results !

That would be amazing, thanks in advance !

This code looks cool. but is there an online demo so i can experiment it first

How to increase the resolution of the output images?

Looking at the code inside of act_max.py it appears to me that the way the neurons get activated in convolutional layers is a bit of a shortcut and will miss most of them. It looks like only the diagonal ones are accessed since xy is the same for x and y:

# Move in the direction of increasing activation of the given neuron
  if end in fc_layers:
    one_hot.flat[unit] = 1.
  elif end in conv_layers:
    one_hot[:, unit, xy, xy] = 1.
  else:
    raise Exception("Invalid layer type!")

I made a quick fix which seems to work, which treats the xy as an index and converts it to a column/row representation - I'm not sure if this makes sense, but it does return results that look reasonable:

# Move in the direction of increasing activation of the given neuron
  if end in fc_layers:
    one_hot.flat[unit] = 1.
  elif end in conv_layers:
    one_hot[:, unit, xy % one_hot.shape[2], int(xy/one_hot.shape[2])] = 1.
  else:
    raise Exception("Invalid layer type!")

further down in that function another fix is required:

# Check the activations
  if end in fc_layers:
    fc = acts[end][0]
    best_unit = fc.argmax()
    obj_act = fc[unit]

  elif end in conv_layers:
    fc = acts[end][0, :, xy % acts[end].shape[2] , int(xy/acts[end].shape[2])]
    best_unit = fc.argmax()
    obj_act = fc[unit]

Hello,

I am trying to run the code but I can't download the weights. The link seems to be deactivatex.

cd nets/upconv && ./download.sh or the direct link doesn't respond.

Can you help with this problem?

Thank you in advance.

This is not an issue, but a feature request. Is there code available for visualizing 2 (or more) neurons at the same time?

I'd like to explore the section of your paper about using 2, or more, neurons to generate a single image, for example volcano (n09472597 volcano) + Candle (n02948072 candle, taper, wax light).

Very cool project.

Hello,thanks a lot for sharing the code. I try to run the program following you description. A question to ask you: "1_activate_output.sh","2_activate_output_placesCNN.sh","4_activate_hidden.sh","5_activate_output_GoogLeNet.sh",they needn't to input image,right? My understanding is you train the model only classify the type(eg:mask,library,and so on) and collect features of these image, then input a type and the model will generate that type image,right?

Hello,

I'm trying to adapt your code for a different network (a regression), and I have trouble training the generator network. I'm using http://arxiv.org/abs/1602.02644 as a reference, but I have trouble finding a good combination of generator architecture, discriminant architecture, and hyper parameters.

Would you be open to shared some details of how you trained the generator?

Thanks,