1 files changed, 13 insertions, 13 deletions
diff --git a/training/conf/experiment/conv_transformer_lines.yaml b/training/conf/experiment/conv_transformer_lines.yaml
index 11646ca..20e369e 100644
--- a/training/conf/experiment/conv_transformer_lines.yaml
+++ b/training/conf/experiment/conv_transformer_lines.yaml
@@ -10,7 +10,7 @@ defaults:
   - override /lr_schedulers: null
   - override /optimizers: null
 
-epochs: &epochs 620
+epochs: &epochs 300
 ignore_index: &ignore_index 3
 num_classes: &num_classes 57
 max_output_len: &max_output_len 89
@@ -27,7 +27,7 @@ callbacks:
   stochastic_weight_averaging:
     _target_: pytorch_lightning.callbacks.StochasticWeightAveraging
     swa_epoch_start: 0.75
-    swa_lrs: 1.0e-5
+    swa_lrs: 1.0e-4
     annealing_epochs: 10
     annealing_strategy: cos
     device: null
@@ -43,15 +43,15 @@ optimizers:
 
 lr_schedulers:
   network:
-      _target_: torch.optim.lr_scheduler.CosineAnnealingLR
-      T_max: *epochs
-      eta_min: 1.0e-5
-      last_epoch: -1
-      interval: epoch
-      monitor: val/loss
+    _target_: torch.optim.lr_scheduler.CosineAnnealingLR
+    T_max: *epochs
+    eta_min: 1.0e-4
+    last_epoch: -1
+    interval: epoch
+    monitor: val/loss
 
 datamodule:
-  batch_size: 32
+  batch_size: 16
   num_workers: 12
   train_fraction: 0.9
   pin_memory: true
@@ -64,7 +64,7 @@ rotary_embedding: &rotary_embedding
 
 attn: &attn
   dim: &hidden_dim 256
-  num_heads: 4
+  num_heads: 6
   dim_head: 64
   dropout_rate: &dropout_rate 0.5
 
@@ -76,12 +76,12 @@ network:
   pad_index: *ignore_index
   encoder:
     _target_: text_recognizer.networks.encoders.efficientnet.EfficientNet
-    arch: b3
+    arch: b0
     stochastic_dropout_rate: 0.2
     bn_momentum: 0.99
     bn_eps: 1.0e-3
   decoder:
-    depth: 6
+    depth: 3
     _target_: text_recognizer.networks.transformer.layers.Decoder
     self_attn:
       _target_: text_recognizer.networks.transformer.attention.Attention
@@ -106,7 +106,7 @@ network:
   pixel_pos_embedding:
     _target_: text_recognizer.networks.transformer.embeddings.axial.AxialPositionalEmbedding
     dim: *hidden_dim
-    shape: [1, 32]
+    shape: [3, 64]
 
 model:
   _target_: text_recognizer.models.transformer.TransformerLitModel