summaryrefslogtreecommitdiff
path: root/training/conf/experiment
diff options
context:
space:
mode:
Diffstat (limited to 'training/conf/experiment')
-rw-r--r--training/conf/experiment/convformer_lines.yaml58
-rw-r--r--training/conf/experiment/mammut_lines.yaml5
-rw-r--r--training/conf/experiment/vit_lines.yaml3
3 files changed, 65 insertions, 1 deletions
diff --git a/training/conf/experiment/convformer_lines.yaml b/training/conf/experiment/convformer_lines.yaml
new file mode 100644
index 0000000..f573433
--- /dev/null
+++ b/training/conf/experiment/convformer_lines.yaml
@@ -0,0 +1,58 @@
+# @package _global_
+
+defaults:
+ - override /criterion: cross_entropy
+ - override /callbacks: htr
+ - override /datamodule: iam_lines
+ - override /network: convformer_lines
+ - override /model: lit_transformer
+ - 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
+
+optimizer:
+ lr: 1.0e-3
+
+# 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/mammut_lines.yaml b/training/conf/experiment/mammut_lines.yaml
index e74e219..eb6f765 100644
--- a/training/conf/experiment/mammut_lines.yaml
+++ b/training/conf/experiment/mammut_lines.yaml
@@ -39,12 +39,15 @@ lr_scheduler:
T_max: *epochs
datamodule:
- batch_size: 8
+ batch_size: 16
train_fraction: 0.95
model:
max_output_len: 89
+optimizer:
+ lr: 1.0e-3
+
trainer:
fast_dev_run: false
gradient_clip_val: 1.0
diff --git a/training/conf/experiment/vit_lines.yaml b/training/conf/experiment/vit_lines.yaml
index 08ed481..2f7731e 100644
--- a/training/conf/experiment/vit_lines.yaml
+++ b/training/conf/experiment/vit_lines.yaml
@@ -26,6 +26,9 @@ criterion:
decoder:
max_output_len: 89
+optim4izer:
+ lr: 1.0e-3
+
# callbacks:
# stochastic_weight_averaging:
# _target_: pytorch_lightning.callbacks.StochasticWeightAveraging