This is pytorch implementation of paper "Glow: Generative Flow with Invertible 1x1 Convolutions" forked from chaiujin, adapted for stanford cars dataset(https://ai.stanford.edu/~jkrause/cars/car_dataset.html)
- Train a model with
train.py <hparams> <dataset> <dataset_root> - Generate interpolations and reconstructions with
infer_stanford.py <hparams> <dataset_root> <z_dir>
Currently, model is trained with hparams/cars.json using Stanford Cars dataset.
| HParam | Value |
|---|---|
| image_shape | (64, 64, 3) |
| hidden_channels | 512 |
| K | 32 |
| L | 3 |
| flow_permutation | invertible 1x1 conv |
| flow_coupling | affine |
| batch_size | 12 |
| learn_top | false |
| y_condition | false |