diff options
Diffstat (limited to 'training/conf/experiment/vqvae.yaml')
| -rw-r--r-- | training/conf/experiment/vqvae.yaml | 38 |
1 files changed, 32 insertions, 6 deletions
diff --git a/training/conf/experiment/vqvae.yaml b/training/conf/experiment/vqvae.yaml index d3db471..d9fa2c4 100644 --- a/training/conf/experiment/vqvae.yaml +++ b/training/conf/experiment/vqvae.yaml @@ -2,26 +2,52 @@ defaults: - override /network: vqvae - - override /criterion: mae + - override /criterion: mse - override /model: lit_vqvae - override /callbacks: wandb_vae - - override /lr_schedulers: - - cosine_annealing + - override /optimizers: null + # - override /lr_schedulers: + # - cosine_annealing + +# lr_schedulers: null +# network: +# _target_: torch.optim.lr_scheduler.OneCycleLR +# max_lr: 1.0e-2 +# total_steps: null +# epochs: 100 +# steps_per_epoch: 200 +# pct_start: 0.1 +# anneal_strategy: cos +# cycle_momentum: true +# base_momentum: 0.85 +# max_momentum: 0.95 +# div_factor: 25 +# final_div_factor: 1.0e4 +# three_phase: true +# last_epoch: -1 +# verbose: false + +# # Non-class arguments +# interval: step +# monitor: val/loss optimizers: network: _target_: madgrad.MADGRAD - lr: 3.0e-4 + lr: 1.0e-4 momentum: 0.9 weight_decay: 0 - eps: 1.0e-6 + eps: 1.0e-7 parameters: network trainer: - max_epochs: 256 + max_epochs: 128 + limit_train_batches: 0.01 + limit_val_batches: 0.1 datamodule: batch_size: 8 + # resize: [288, 320] summary: null |