Training scripts

Dataset-independence

• train.py: train one model (eg. beta-vae, IWAE, bivae) on one specific hyperparamter config

  • E.g. Train BiVAE on osmnx_roads data of the following cities, with images of bgcolors

  nohup python train.py --model_name="bivae" \
  --latent_dim=10 --hidden_dims 32 64 128 256 --adv_dim 32 32 32 --adv_weight 1.0 \
  --data_root="/data/hayley-old/osmnx_data/images" \
  --data_name="osmnx_roads" \
  --cities 'la' 'charlotte' 'vegas' 'boston' 'paris' \
       'amsterdam' 'shanghai' 'seoul' 'chicago' 'manhattan' \
       'berlin' 'montreal' 'rome' \
  --bgcolors "k" "r" "g" "b" "y" --n_styles=5 \
  --zooms 14 \
  --gpu_id=2 --max_epochs=300   --terminate_on_nan=True  \
  -lr 3e-4 -bs 32 \
  --log_root="/data/hayley-old/Tenanbaum2000/lightning_logs/2021-05-18/" &

  • E.g.: Train BIVAE on Rotated MNIST of optionally specified subset (given as a filepath to .npy file containing the indices from the original Training MNIST data)

  ## Specify which indices to use among the MNIST -- comparable to DIVA's experiments
  ## change 0 to anything inbtw 0,...,9
  nohup python train.py --model_name="bivae" \
  --latent_dim=128 --hidden_dims 32 64 64 64 --adv_dim 32 32 32 \
  --data_name="multi_rotated_mnist" --angles -45 0 45 --n_styles=3 \
  --selected_inds_fp='/data/hayley-old/Tenanbaum2000/data/Rotated-MNIST/supervised_inds_0.npy' \
  --gpu_id=2

  • E.g.: Train Bivae on multi styles of maptiles from specified cities

  # Train BiVAE on Multi Maptiles MNIST
  nohup python train.py --model_name="bivae" \
  --latent_dim=10 --hidden_dims 32 64 128 256 --adv_dim 32 32 32 --adv_weight 15.0 \
  --data_name="multi_maptiles" \
  --cities la paris \
  --styles CartoVoyagerNoLabels StamenTonerBackground --n_styles=3 \
  --zooms 14 \
  --gpu_id=2 --max_epochs=400   --terminate_on_nan=True  \
  -lr 3e-4 -bs 32 \
  --log_root="/data/hayley-old/Tenanbaum2000/lightning_logs/2021-01-23/" &

Hyperparameter tuning using Ray Tune

• tune_asha.py: Use Tune's AsyncHyperBandScheduler to search hyperparameter space more efficiently. Use --tune_metric to specify the value of tune.run's metric argument, e.g. --tune_metric loss for
• tune_asha_with_beta_scheduler.py:
• `

Dataset-specific

Rotated MNIST

• tune_asha_mnists.py

osmnx_roads

• tune_asha_osmnx_roads.py