diff options
Diffstat (limited to 'training/conf/experiment/vqgan.yaml')
| -rw-r--r-- | training/conf/experiment/vqgan.yaml | 36 |
1 files changed, 21 insertions, 15 deletions
diff --git a/training/conf/experiment/vqgan.yaml b/training/conf/experiment/vqgan.yaml index 9224bc7..6c78deb 100644 --- a/training/conf/experiment/vqgan.yaml +++ b/training/conf/experiment/vqgan.yaml @@ -2,7 +2,7 @@ defaults: - override /network: vqvae - - override /criterion: vqgan_loss + - override /criterion: null - override /model: lit_vqgan - override /callbacks: wandb_vae - override /optimizers: null @@ -11,7 +11,7 @@ defaults: criterion: _target_: text_recognizer.criterions.vqgan_loss.VQGANLoss reconstruction_loss: - _target_: torch.nn.L1Loss + _target_: torch.nn.MSELoss reduction: mean discriminator: _target_: text_recognizer.criterions.n_layer_discriminator.NLayerDiscriminator @@ -21,35 +21,41 @@ criterion: vq_loss_weight: 0.25 discriminator_weight: 1.0 discriminator_factor: 1.0 - discriminator_iter_start: 2.0e4 + discriminator_iter_start: 5e2 datamodule: - batch_size: 6 + batch_size: 8 + resize: [288, 320] -lr_schedulers: null +lr_schedulers: + generator: + _target_: torch.optim.lr_scheduler.CosineAnnealingLR + T_max: 128 + eta_min: 4.5e-6 + last_epoch: -1 -# lr_schedulers: -# generator: + interval: epoch + monitor: val/loss # _target_: torch.optim.lr_scheduler.OneCycleLR # max_lr: 3.0e-4 # total_steps: null # epochs: 100 -# steps_per_epoch: 3369 +# steps_per_epoch: 2496 # pct_start: 0.1 # anneal_strategy: cos # cycle_momentum: true # base_momentum: 0.85 # max_momentum: 0.95 -# div_factor: 1.0e3 +# div_factor: 25 # final_div_factor: 1.0e4 # three_phase: true # last_epoch: -1 # verbose: false -# + # # Non-class arguments # interval: step # monitor: val/loss -# + # discriminator: # _target_: torch.optim.lr_scheduler.CosineAnnealingLR # T_max: 64 @@ -79,7 +85,7 @@ optimizers: parameters: loss_fn.discriminator trainer: - max_epochs: 64 - # gradient_clip_val: 1.0e1 - -summary: null + max_epochs: 128 + limit_train_batches: 0.1 + limit_val_batches: 0.1 + # gradient_clip_val: 100 |