dpkingma/nips14-ssl

Code for reproducing results of NIPS 2014 paper "Semi-Supervised Learning with Deep Generative Models"

nips14-ssl's Issues

PCA Parameters are absent in /models /svhn_z_x_pca_300-500-500

I notice that there are some pre-trained models in models/ directory. I want to use the svhn_z_x_pca_300-500-500 model, but can't find pca_params.ndict.tar.gz in directory models/svhn_z_x_pca_300-500-500/. Could you be kind to provide it?

M2 with fully observed labels: prior p(y) commented out?

Hi, I've been going through the code for the M2 model, and noticed the prior p(y) (logpy in the code) has been commented out and missing from your final reconstruction loss? (See line 126.) Is there a reason logpy was commented out? How does the model still work without this?

python run_gpulearn_z_x.py mnist got some problem

Hi, I follow your steps and got error when calling

python run_gpulearn_z_x.py mnist

python run_gpulearn_z_x.py mnist
logdir: results/gpulearn_z_x_mnist_50-(500, 500)__1432466595/
gpulearn_z_x 50 (500, 500) mnist 0
AdaM 0.0003 0.1 0.001 0.02
Variational Auto-Encoder 1000
Traceback (most recent call last):
  File "run_gpulearn_z_x.py", line 10, in <module>
    gpulearn_z_x.main(dataset='mnist', n_z=50, n_hidden=n_hidden, seed=0, comment='', gfx=True)
  File "/home/steven/nips14-ssl/gpulearn_z_x.py", line 369, in main
    loop_va(dostep, hook)
  File "/home/steven/nips14-ssl/gpulearn_z_x.py", line 378, in loop_va
    L = doEpoch()
  File "/home/steven/nips14-ssl/gpulearn_z_x.py", line 402, in doEpoch
    L += model.evalAndUpdate(x_minibatch, {}).sum()
  File "/home/steven/nips14-ssl/anglepy/GPUVAEModel.py", line 110, in evalAndUpdate
    L = self.f_evalAndUpdate(*allvars)
  File "/home/steven/nips14-ssl/anglepy/misc.py", line 81, in func
    f[0] = theano.function(*args, **kwargs)
  File "/home/steven/src/theano/theano/compile/function.py", line 265, in function
    profile=profile)
  File "/home/steven/src/theano/theano/compile/pfunc.py", line 489, in pfunc
    no_default_updates=no_default_updates)
  File "/home/steven/src/theano/theano/compile/pfunc.py", line 217, in rebuild_collect_shared
    raise TypeError(err_msg, err_sug)
TypeError: ('An update must have the same type as the original shared variable (shared_var=<TensorType(float32, matrix)>, shared_var.type=TensorType(float32, matrix), update_val=Elemwise{add,no_inplace}.0, update_val.type=TensorType(float64, matrix)).', 'If the difference is related to the broadcast pattern, you can call the tensor.unbroadcast(var, axis_to_unbroadcast[, ...]) function to remove broadcastable dimensions.')

f_dec referenced before assignment when running gpulearn_z_x.py mnist

Hi, I tried running the unsupervised M1 model on the mnist dataset, using

sh python run gpulearn_z_x.py mnist,

and got

NameError: free variable 'f_dec' referenced before assignment in enclosing scope

Looking at the code for the supervised M1+M2 model, I think the line,

f_enc, f_dec = lambda x:x, lambda x:x

might have been deleted around line 41, of gpulearn_z_x.py ? Might also be needed for some of the other cases, with no preprocessing, mnist_binarized, freyface, freyface_bernoulli, norb, etc, but I have'nt downloaded the data for these yet?

Downloaded the 3 datasets for svhn, from http://ufldl.stanford.edu/housenumbers/, and that's fine.

Thanks 👍

Could you provide a description about the network structure used in Model2?

Sorry but I'm not very familiar with Theano. Could please provide a word/image description of the network structure used in M2 for supervised learning?

I'm trying to reproduce the result in tensorflow.

adam.py

adam.py:55 , Is it 'alpha_t' instead of 'self.alpha' , because 'alpha_t' is never used.

Remove rng_curand import from anglepy.models

Given the lines using rng_curand have been commented out, you may want to remove the imports from line 11 of GPUVAE_YZ_X.py and GPUVAE_Z_X.py in anglepy.models. For me at least, import theano.sandbox.cuda.rng_curand as rng_curand was causing a segfault.