diff options
Diffstat (limited to 'training/conf/experiment')
-rw-r--r-- | training/conf/experiment/conv_transformer_lines.yaml | 8 | ||||
-rw-r--r-- | training/conf/experiment/conv_transformer_paragraphs.yaml | 8 |
2 files changed, 8 insertions, 8 deletions
diff --git a/training/conf/experiment/conv_transformer_lines.yaml b/training/conf/experiment/conv_transformer_lines.yaml index 48df78d..eb9bc9e 100644 --- a/training/conf/experiment/conv_transformer_lines.yaml +++ b/training/conf/experiment/conv_transformer_lines.yaml @@ -6,8 +6,8 @@ defaults: - override /datamodule: iam_lines - override /network: conv_transformer - override /model: lit_transformer - - override /lr_schedulers: null - - override /optimizers: null + - override /lr_scheduler: null + - override /optimizer: null epochs: &epochs 512 ignore_index: &ignore_index 3 @@ -28,7 +28,7 @@ callbacks: annealing_strategy: cos device: null -optimizers: +optimizer: _target_: torch.optim.RAdam lr: 3.0e-4 betas: [0.9, 0.999] @@ -36,7 +36,7 @@ optimizers: eps: 1.0e-8 parameters: network -lr_schedulers: +lr_scheduler: _target_: torch.optim.lr_scheduler.OneCycleLR max_lr: 3.0e-4 total_steps: null diff --git a/training/conf/experiment/conv_transformer_paragraphs.yaml b/training/conf/experiment/conv_transformer_paragraphs.yaml index d0d0314..7c6e231 100644 --- a/training/conf/experiment/conv_transformer_paragraphs.yaml +++ b/training/conf/experiment/conv_transformer_paragraphs.yaml @@ -6,8 +6,8 @@ defaults: - override /datamodule: iam_extended_paragraphs - override /network: conv_transformer - override /model: lit_transformer - - override /lr_schedulers: null - - override /optimizers: null + - override /lr_scheduler: null + - override /optimizer: null epochs: &epochs 600 num_classes: &num_classes 58 @@ -28,7 +28,7 @@ callbacks: annealing_strategy: cos device: null -optimizers: +optimizer: _target_: torch.optim.RAdam lr: 3.0e-4 betas: [0.9, 0.999] @@ -36,7 +36,7 @@ optimizers: eps: 1.0e-8 parameters: network -lr_schedulers: +lr_scheduler: _target_: torch.optim.lr_scheduler.OneCycleLR max_lr: 3.0e-4 total_steps: null |