PyTorch implementation of DoppelGANger, work in progress. Does not produce good output at this time!
Based on https://github.com/fjxmlzn/DoppelGANger, but converted to PyTorch.
pip install -r requirements.txt
(Recommended to install within a virtual environment of some type.)
Very simple tests are currently in test_doppelganger.py. These mostly run 1
epoch on some random data to make sure tensor shapes align properly.
pytest test_doppelganger.py
Known differences between this code and the tf1 code:
- Variable length sequences are not supported. This means the gen_flag features are not included in training data or generated data.
- Softmax outputs used for later steps in generator instead of taking one-hot representation of the argmax.
- Parameter initialization differences between PyTorch and tf1. Notably, there is no equivalent in PyTorch to the forget_bias parameter in the tf1 LSTMCell that initializes the forget gate bias to 1.
- CPU only
- Additional attributes and attributes use the same input noise vector.