2 files changed, 68 insertions, 30 deletions
diff --git a/training/conf/experiment/mammut_lines.yaml b/training/conf/experiment/mammut_lines.yaml
new file mode 100644
index 0000000..e74e219
--- /dev/null
+++ b/training/conf/experiment/mammut_lines.yaml
@@ -0,0 +1,55 @@
+# @package _global_
+
+defaults:
+  - override /criterion: cross_entropy
+  - override /callbacks: htr
+  - override /datamodule: iam_lines
+  - override /network: mammut_lines
+  - override /model: lit_mammut
+  - override /lr_scheduler: cosine_annealing
+  - override /optimizer: adan
+
+tags: [lines, vit]
+epochs: &epochs 320
+ignore_index: &ignore_index 3
+# summary: [[1, 1, 56, 1024], [1, 89]]
+
+logger:
+  wandb:
+    tags: ${tags}
+
+criterion:
+  ignore_index: *ignore_index
+  # label_smoothing: 0.05
+
+
+decoder:
+  max_output_len: 89
+
+# callbacks:
+#   stochastic_weight_averaging:
+#     _target_: pytorch_lightning.callbacks.StochasticWeightAveraging
+#     swa_epoch_start: 0.75
+#     swa_lrs: 1.0e-5
+#     annealing_epochs: 10
+#     annealing_strategy: cos
+#     device: null
+
+lr_scheduler:
+  T_max: *epochs
+
+datamodule:
+  batch_size: 8
+  train_fraction: 0.95
+
+model:
+  max_output_len: 89
+
+trainer:
+  fast_dev_run: false
+  gradient_clip_val: 1.0
+  max_epochs: *epochs
+  accumulate_grad_batches: 1
+  limit_train_batches: 1.0
+  limit_val_batches: 1.0
+  limit_test_batches: 1.0
diff --git a/training/conf/experiment/vit_lines.yaml b/training/conf/experiment/vit_lines.yaml
index f3049ea..08ed481 100644
--- a/training/conf/experiment/vit_lines.yaml
+++ b/training/conf/experiment/vit_lines.yaml
@@ -6,11 +6,11 @@ defaults:
   - override /datamodule: iam_lines
   - override /network: vit_lines
   - override /model: lit_transformer
-  - override /lr_scheduler: null
-  - override /optimizer: null
+  - override /lr_scheduler: cosine_annealing
+  - override /optimizer: adan
 
 tags: [lines, vit]
-epochs: &epochs 128
+epochs: &epochs 320
 ignore_index: &ignore_index 3
 # summary: [[1, 1, 56, 1024], [1, 89]]
 
@@ -26,37 +26,20 @@ criterion:
 decoder:
   max_output_len: 89
 
-callbacks:
-  stochastic_weight_averaging:
-    _target_: pytorch_lightning.callbacks.StochasticWeightAveraging
-    swa_epoch_start: 0.75
-    swa_lrs: 1.0e-5
-    annealing_epochs: 10
-    annealing_strategy: cos
-    device: null
-
-optimizer:
-  _target_: adan_pytorch.Adan
-  lr: 3.0e-4
-  betas: [0.02, 0.08, 0.01]
-  weight_decay: 0.02
+# callbacks:
+#   stochastic_weight_averaging:
+#     _target_: pytorch_lightning.callbacks.StochasticWeightAveraging
+#     swa_epoch_start: 0.75
+#     swa_lrs: 1.0e-5
+#     annealing_epochs: 10
+#     annealing_strategy: cos
+#     device: null
 
 lr_scheduler:
-  _target_: torch.optim.lr_scheduler.ReduceLROnPlateau
-  mode: min
-  factor: 0.8
-  patience: 10
-  threshold: 1.0e-4
-  threshold_mode: rel
-  cooldown: 0
-  min_lr: 1.0e-5
-  eps: 1.0e-8
-  verbose: false
-  interval: epoch
-  monitor: val/cer
+  T_max: *epochs
 
 datamodule:
-  batch_size: 16
+  batch_size: 8
   train_fraction: 0.95
 
 model: