apozas/ebm-torch

Energy-based models in PyTorch

Energy-based models in Pytorch

Author: Alejandro Pozas Kerstjens

Implementation of different generative models based on energy-based learning. Examples are provided with the MNIST dataset.

Libraries required:

• pytorch >= 0.4.0 as ML framework
• numpy for math operations
• matplotlib for image visualization
• tqdm for custom progress bar
• imageio for exporting outputs to .gif format

1. RBM

Restricted Boltzmann Machine with binary visible and hidden units. Although in this example it is used as a generative model, RBMs can also perform supervised tasks. The current implementation allows for both binary and continuous-valued visible units.

2. Deep Belief Network

Deep belief network with greedy pre-training plus global finetuning.

ebm-torch's Issues

Update for pytorch 1.3.1

Hello,

i saw this awesome library and wanted to test the examples, but right now both seem to have issues.

The RBM example runs through but produces a completely black image as the output.
The DBN example stops at the fine-tuning stage

  File ".../ebm-torch/dbn_example.py", line 84, in <module>
    batch_size=batch_size)
  File ".../ebm-torch/ebm/models.py", line 310, in finetune
    v_sample = self.sampler.get_v_sample(wakepos_samples[-2],
AttributeError: 'PersistentContrastiveDivergence' object has no attribute 'get_v_sample'

I hope this would be relatively easy to fix?