diff options
Diffstat (limited to 'training/conf')
-rw-r--r-- | training/conf/experiment/vqgan.yaml | 17 | ||||
-rw-r--r-- | training/conf/experiment/vqvae.yaml | 8 | ||||
-rw-r--r-- | training/conf/lr_schedulers/cosine_annealing.yaml (renamed from training/conf/lr_scheduler/cosine_annealing.yaml) | 0 | ||||
-rw-r--r-- | training/conf/lr_schedulers/one_cycle.yaml (renamed from training/conf/lr_scheduler/one_cycle.yaml) | 0 | ||||
-rw-r--r-- | training/conf/optimizers/madgrad.yaml (renamed from training/conf/optimizer/madgrad.yaml) | 0 |
5 files changed, 15 insertions, 10 deletions
diff --git a/training/conf/experiment/vqgan.yaml b/training/conf/experiment/vqgan.yaml index 570e7f9..554ec9e 100644 --- a/training/conf/experiment/vqgan.yaml +++ b/training/conf/experiment/vqgan.yaml @@ -8,6 +8,19 @@ defaults: - override /optimizers: null - override /lr_schedulers: null +criterion: + _target_: text_recognizer.criterions.vqgan_loss.VQGANLoss + reconstruction_loss: + _target_: torch.nn.L1Loss + reduction: mean + discriminator: + _target_: text_recognizer.criterions.n_layer_discriminator.NLayerDiscriminator + in_channels: 1 + num_channels: 32 + num_layers: 3 + vq_loss_weight: 0.8 + discriminator_weight: 0.6 + datamodule: batch_size: 8 @@ -33,7 +46,7 @@ lr_schedulers: optimizers: generator: _target_: madgrad.MADGRAD - lr: 2.0e-5 + lr: 4.5e-6 momentum: 0.5 weight_decay: 0 eps: 1.0e-6 @@ -42,7 +55,7 @@ optimizers: discriminator: _target_: madgrad.MADGRAD - lr: 2.0e-5 + lr: 4.5e-6 momentum: 0.5 weight_decay: 0 eps: 1.0e-6 diff --git a/training/conf/experiment/vqvae.yaml b/training/conf/experiment/vqvae.yaml index 397a039..8dbb257 100644 --- a/training/conf/experiment/vqvae.yaml +++ b/training/conf/experiment/vqvae.yaml @@ -10,16 +10,8 @@ defaults: trainer: max_epochs: 256 - # gradient_clip_val: 0.25 datamodule: batch_size: 8 -# lr_scheduler: - # epochs: 64 - # steps_per_epoch: 1245 - -# optimizer: - # lr: 1.0e-3 - summary: null diff --git a/training/conf/lr_scheduler/cosine_annealing.yaml b/training/conf/lr_schedulers/cosine_annealing.yaml index c53ee3a..c53ee3a 100644 --- a/training/conf/lr_scheduler/cosine_annealing.yaml +++ b/training/conf/lr_schedulers/cosine_annealing.yaml diff --git a/training/conf/lr_scheduler/one_cycle.yaml b/training/conf/lr_schedulers/one_cycle.yaml index c60577a..c60577a 100644 --- a/training/conf/lr_scheduler/one_cycle.yaml +++ b/training/conf/lr_schedulers/one_cycle.yaml diff --git a/training/conf/optimizer/madgrad.yaml b/training/conf/optimizers/madgrad.yaml index a6c059d..a6c059d 100644 --- a/training/conf/optimizer/madgrad.yaml +++ b/training/conf/optimizers/madgrad.yaml |