vdumoulin / discgen Goto Github PK

Hi,
First of all thank you for this very easy-to-follow code. One thing, I am getting an error while trying:
fuel-convert celeba 64

File "/usr/local/bin/fuel-convert", line 11, in <module> load_entry_point('fuel==0.2.0', 'console_scripts', 'fuel-convert')() File "/usr/local/lib/python2.7/dist-packages/fuel/bin/fuel_convert.py", line 69, in main output_paths = convert_function(**args_dict) File "/usr/local/lib/python2.7/dist-packages/fuel/converters/celeba.py", line 198, in convert_celeba directory, output_directory, output_filename) File "/usr/local/lib/python2.7/dist-packages/fuel/converters/base.py", line 45, in wrapped return f(directory, *args, **kwargs) File "/usr/local/lib/python2.7/dist-packages/fuel/converters/celeba.py", line 139, in convert_celeba_64 h5file = _initialize_conversion(directory, output_path, (64, 64)) File "/usr/local/lib/python2.7/dist-packages/fuel/converters/celeba.py", line 41, in _initialize_conversion skiprows=2, usecols=tuple(range(1, 41))) + File "/usr/local/lib/python2.7/dist-packages/numpy/lib/npyio.py", line 806, in loadtxt vals = [vals[i] for i in usecols] IndexError: list index out of range

My gut felling is that - this is a problem with fuel. But, may you please help me by mentioning a proper workaround?

Thanks,
Arghya

I wanna ask why you set log_sigma to be zero rather than learn it through the net?
And you set mu_theta to be the reconstructed or sampled image directly, why don't you sample from the gaussian distribution with mu_theta and log_sigma?

This code has been working great for me. Currently, the discriminative regularization in this codebase uses the loss from the batch norm layers of the ConvolutionalSequence, but the paper covers the more general case of taking loss from elsewhere in the classifier, as shown below:

I'd like to add losses at the higher layers - MLP and y-hat. I'm curious if anyone else is interested in working on this with me - the discriminative regularization parts are some of trickier parts of building the training computation graph, and y-hat currently doesn't have a named node in the classifier graph.