Reproducing 2d results

[taldatech]

Reproducing 2d results

[taldatech]

Hi,
We used the hyper-parameters that can be found in the appendix and the notebook here:
https://github.com/taldatech/soft-intro-vae-pytorch/blob/main/soft_intro_vae_tutorial/soft_intro_vae_2d_code_tutorial.ipynb

device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
print("device:", device)
seed = 92 # for reproducible results
datasets = ['8Gaussians', '2spirals', 'checkerboard', 'rings']
chosen_hyperparmas = {'8Gaussians': {'b_kl': 0.3, 'b_neg': 0.9, 'b_rec': 0.2},
'2spirals': {'b_kl': 0.5, 'b_neg': 1.0, 'b_rec': 0.2},
'checkerboard': {'b_kl': 0.1, 'b_neg': 0.2, 'b_rec': 0.2},
'rings': {'b_kl': 0.2, 'b_neg': 1.0, 'b_rec': 0.2}}

num_iter = 30_000
lr = 2e-4
batch_size = 512

Also, in: https://github.com/taldatech/soft-intro-vae-pytorch/blob/main/soft_intro_vae_2d/train_soft_intro_vae_2d.py

"""
Recommended hyper-parameters:
- 8Gaussians: beta_kl: 0.3, beta_rec: 0.2, beta_neg: 0.9, z_dim: 2, batch_size: 512
- 2spirals: beta_kl: 0.5, beta_rec: 0.2, beta_neg: 1.0, z_dim: 2, batch_size: 512
- checkerboard: beta_kl: 0.1, beta_rec: 0.2, beta_neg: 0.2, z_dim: 2, batch_size: 512
- rings: beta_kl: 0.2, beta_rec: 0.2, beta_neg: 1.0, z_dim: 2, batch_size: 512
"""

I hope this helps

[Atcold]

Thanks!