fibarrola / eval_gen

Evaluation code with models for the paper "On the Quantitative Analysis of Decoder-Based Generative Models"

#On the Quantitative Analysis of Decoder-Based Models

Dependencies: Theano, Lasagne

##Evaluate the models:

To evaluate a decoder-based generative model:

1. You need to provide your decoder gen in the form of a python function, which takes in a theano tensor variable Z (latent) and outputs a theano tensor variable X (the sample) (an example can be found in line 84-85 in ./lib/utils.py),

     def gen(Z):
       ... 
       return X

and modify the model loading procedure in load_model function at ./lib/utils.py.

2. Modify the data loading procedure at ./lib/utils.py (e.g. load_mnist), including validation/training split. Or for loading another dataset.

3. If you're evaluating your model on a dataset other than MNIST, you need to modify the data dimension in ./sampling/sampler.py, at line 19 and 20, and modify the procedure of calculating data likelihood at line 113 and 117. An example can be found at ./sampling/svhn_sampler.py

4. Various command instruction can be found at comment after the command.

5. Run function main(*args) in ./experiment/run.py

##Visualize posterior samples:

1. mkdir ./vis
2. Run function main(*args) in ./experiment/run.py, with plot_posterior set to 1.
3. For visualizing the posterior samples of a particular digit X, set exps to "postXtra" for digit in training set and "postXval" for digit in validation set.

See an tensorflow implementation: https://github.com/jiamings/ais