diff options
Diffstat (limited to 'training/conf/experiment/conformer_lines.yaml')
| -rw-r--r-- | training/conf/experiment/conformer_lines.yaml | 43 |
1 files changed, 4 insertions, 39 deletions
diff --git a/training/conf/experiment/conformer_lines.yaml b/training/conf/experiment/conformer_lines.yaml index c3f4ea5..06e761e 100644 --- a/training/conf/experiment/conformer_lines.yaml +++ b/training/conf/experiment/conformer_lines.yaml @@ -5,15 +5,15 @@ defaults: - override /criterion: ctc - override /callbacks: htr - override /datamodule: iam_lines - - override /network: null + - override /network: conformer - override /model: null - override /lr_schedulers: null - override /optimizers: null -epochs: &epochs 200 +epochs: &epochs 999 num_classes: &num_classes 57 -max_output_len: &max_output_len 762 -summary: [[1, 57, 1024]] +max_output_len: &max_output_len 89 +summary: [[1, 56, 1024]] mapping: &mapping mapping: @@ -59,41 +59,6 @@ datamodule: pin_memory: true << : *mapping -network: - _target_: text_recognizer.networks.conformer.Conformer - depth: 16 - num_classes: *num_classes - dim: &dim 128 - block: - _target_: text_recognizer.networks.conformer.ConformerBlock - dim: *dim - attn: - _target_: text_recognizer.networks.conformer.Attention - dim: *dim - heads: 8 - dim_head: 64 - mult: 4 - ff: - _target_: text_recognizer.networks.conformer.Feedforward - dim: *dim - expansion_factor: 4 - dropout: 0.1 - conv: - _target_: text_recognizer.networks.conformer.ConformerConv - dim: *dim - expansion_factor: 2 - kernel_size: 31 - dropout: 0.1 - subsampler: - _target_: text_recognizer.networks.conformer.Subsampler - pixel_pos_embedding: - _target_: text_recognizer.networks.transformer.AxialPositionalEmbedding - dim: *dim - shape: [6, 127] - channels: *dim - depth: 3 - dropout: 0.1 - model: _target_: text_recognizer.models.conformer.LitConformer <<: *mapping |