summaryrefslogtreecommitdiff
path: root/training/conf/experiment/cnn_htr_char_lines.yaml
diff options
context:
space:
mode:
Diffstat (limited to 'training/conf/experiment/cnn_htr_char_lines.yaml')
-rw-r--r--training/conf/experiment/cnn_htr_char_lines.yaml19
1 files changed, 10 insertions, 9 deletions
diff --git a/training/conf/experiment/cnn_htr_char_lines.yaml b/training/conf/experiment/cnn_htr_char_lines.yaml
index 682f138..759161c 100644
--- a/training/conf/experiment/cnn_htr_char_lines.yaml
+++ b/training/conf/experiment/cnn_htr_char_lines.yaml
@@ -1,3 +1,5 @@
+# @package _glbal_
+
defaults:
- override /mapping: null
- override /criterion: null
@@ -34,7 +36,7 @@ callbacks:
optimizers:
madgrad:
_target_: madgrad.MADGRAD
- lr: 3.0e-4
+ lr: 1.0e-4
momentum: 0.9
weight_decay: 0
eps: 1.0e-6
@@ -63,21 +65,20 @@ lr_schedulers:
datamodule:
_target_: text_recognizer.data.iam_lines.IAMLines
- batch_size: 32
num_workers: 12
- train_fraction: 0.8
- augment: true
+ train_fraction: 0.9
pin_memory: true
- word_pieces: false
+ transform: transform/iam_lines.yaml
+ test_transform: transform/iam_lines.yaml
+ target_transform: target_transform/word_pieces
<<: *mapping
network:
_target_: text_recognizer.networks.conv_transformer.ConvTransformer
input_dims: [1, 56, 1024]
- hidden_dim: &hidden_dim 128
+ hidden_dim: &hidden_dim 256
encoder_dim: 1280
dropout_rate: 0.2
- <<: *mapping
num_classes: *num_classes
pad_index: *ignore_index
encoder:
@@ -111,8 +112,8 @@ network:
pixel_pos_embedding:
_target_: text_recognizer.networks.transformer.positional_encodings.PositionalEncoding2D
hidden_dim: *hidden_dim
- max_h: 1
- max_w: 32
+ max_h: 18
+ max_w: 20
token_pos_embedding:
_target_: text_recognizer.networks.transformer.positional_encodings.PositionalEncoding
hidden_dim: *hidden_dim