diff options
Diffstat (limited to 'training/conf/experiment')
3 files changed, 8 insertions, 7 deletions
diff --git a/training/conf/experiment/conv_transformer_lines.yaml b/training/conf/experiment/conv_transformer_lines.yaml index 6c266b8..2918317 100644 --- a/training/conf/experiment/conv_transformer_lines.yaml +++ b/training/conf/experiment/conv_transformer_lines.yaml @@ -64,10 +64,10 @@ rotary_embedding: &rotary_embedding dim: 64 attn: &attn - dim: &hidden_dim 256 + dim: &hidden_dim 128 num_heads: 4 dim_head: 64 - dropout_rate: &dropout_rate 0.5 + dropout_rate: &dropout_rate 0.2 network: _target_: text_recognizer.networks.conv_transformer.ConvTransformer @@ -76,8 +76,9 @@ network: num_classes: *num_classes pad_index: *ignore_index encoder: - _target_: text_recognizer.networks.encoders.efficientnet.EfficientNet - arch: b1 + _target_: text_recognizer.networks.efficientnet.EfficientNet + arch: b0 + depth: 5 stochastic_dropout_rate: 0.2 bn_momentum: 0.99 bn_eps: 1.0e-3 @@ -113,7 +114,7 @@ network: dim: *hidden_dim heads: 4 shape: *shape - depth: 1 + depth: 2 dim_head: 64 dim_index: 1 diff --git a/training/conf/experiment/conv_transformer_paragraphs.yaml b/training/conf/experiment/conv_transformer_paragraphs.yaml index 4f15ef2..32f5763 100644 --- a/training/conf/experiment/conv_transformer_paragraphs.yaml +++ b/training/conf/experiment/conv_transformer_paragraphs.yaml @@ -78,7 +78,7 @@ network: num_classes: *num_classes pad_index: *ignore_index encoder: - _target_: text_recognizer.networks.encoders.efficientnet.EfficientNet + _target_: text_recognizer.networks.efficientnet.EfficientNet arch: b1 stochastic_dropout_rate: 0.2 bn_momentum: 0.99 diff --git a/training/conf/experiment/conv_transformer_paragraphs_wp.yaml b/training/conf/experiment/conv_transformer_paragraphs_wp.yaml index 91fba9a..bf192ec 100644 --- a/training/conf/experiment/conv_transformer_paragraphs_wp.yaml +++ b/training/conf/experiment/conv_transformer_paragraphs_wp.yaml @@ -89,7 +89,7 @@ network: num_classes: *num_classes pad_index: *ignore_index encoder: - _target_: text_recognizer.networks.encoders.efficientnet.EfficientNet + _target_: text_recognizer.networks.efficientnet.EfficientNet arch: b0 out_channels: 1280 stochastic_dropout_rate: 0.2 |