svalleco/3dgan

3dgan's Issues

neon 3D conv and pooling layers do not instantiate properly

The link to the code with our 3D GAN:
https://github.com/svalleco/3Dgan/blob/master/neon/gan3D.py

Concerning the errors: for example,
https://github.com/svalleco/3Dgan/blob/master/neon/gan3D.py#L57
also next 2 lines are wrong too. It doesn't work in 3D case, but as you said GAN model with 3D Conv/Deconv layers has not been put up yet.

The error messege:

Traceback (most recent call last):
  File "GAN.py", line 51, in <module>
    Deconv((6, 6, 8, 64), **conv4),
  File "//anaconda/envs/neon/lib/python3.6/site-packages/neon-2.0.0-py3.6.egg/neon/layers/layer.py", line 1655, in __init__
    dilation=dilation, init=init, bsum=batch_norm))
  File "//anaconda/envs/neon/lib/python3.6/site-packages/neon-2.0.0-py3.6.egg/neon/layers/layer.py", line 950, in __init__
    self.deconvparams.update(d)
TypeError: cannot convert dictionary update sequence element #0 to a sequence

And some problems with types. If I remove dilation, padding parameters or put only one number, it works, but it isn't what we need. In the discriminator case, it seems to work, but in fact, it works in 2d not 3d.

Besides, for example, the Poling layer isn't implemented for 3D too
https://github.com/NervanaSystems/neon/blob/master/neon/layers/layer.py#L493
We need the polling layer for the discriminator :)

Thank you for your attention and participation!

Best,
Elena.

How to properly run the 3Dgan Keras model?

Hello,

I'd like to know how to properly run the Keras based 3Dgan model? Which of the branches is recommended? It seems that master branch was not developed for a long time.
I've tried to run the model on Nvidia V100 with TF2 and Keras2 but I'v got some syntax errors.
Please, let me know if there are any guidelines within repository about how to run the model in proper way.

Thanks and regards,
Adam

What's the license of this repo

Intel's Tensorflow team has worked in this repo for several months, we hope to fork it to Intel's Tensorflow model zoo and display it to users. But we didn't know that we can directly use the code or not because lack of license.

Will author license the code to Intel under the Apache. 2.0 license? If so, we could add a license.

Regards,
Teng