diff options
Diffstat (limited to 'training')
| -rw-r--r-- | training/conf/experiment/conv_transformer_paragraphs.yaml | 14 |
1 files changed, 7 insertions, 7 deletions
diff --git a/training/conf/experiment/conv_transformer_paragraphs.yaml b/training/conf/experiment/conv_transformer_paragraphs.yaml index 34eedab..60898da 100644 --- a/training/conf/experiment/conv_transformer_paragraphs.yaml +++ b/training/conf/experiment/conv_transformer_paragraphs.yaml @@ -14,7 +14,7 @@ epochs: &epochs 600 num_classes: &num_classes 58 ignore_index: &ignore_index 3 max_output_len: &max_output_len 682 -summary: [[1, 1, 576, 640], [1, 682]] +# summary: [[1, 1, 576, 640], [1, 682]] logger: wandb: @@ -55,7 +55,7 @@ lr_scheduler: monitor: val/cer datamodule: - batch_size: 8 + batch_size: 2 train_fraction: 0.95 network: @@ -67,9 +67,9 @@ network: encoder: _target_: text_recognizer.networks.convnext.ConvNext dim: 16 - dim_mults: [2, 4, 8, 8, 8] - depths: [3, 3, 3, 4, 6] - downsampling_factors: [[2, 2], [2, 2], [2, 2], [2, 1], [2, 1]] + dim_mults: [2, 4, 8, 8] + depths: [3, 3, 6, 6] + downsampling_factors: [[2, 2], [2, 2], [2, 2], [2, 1]] attn: _target_: text_recognizer.networks.convnext.TransformerBlock attn: @@ -118,7 +118,7 @@ network: _target_: "text_recognizer.networks.transformer.embeddings.axial.\ AxialPositionalEmbeddingImage" dim: *hidden_dim - axial_shape: [7, 128] + axial_shape: [36, 80] axial_dims: [64, 64] token_pos_embedding: _target_: "text_recognizer.networks.transformer.embeddings.fourier.\ @@ -130,4 +130,4 @@ network: trainer: gradient_clip_val: 1.0 max_epochs: *epochs - accumulate_grad_batches: 1 + accumulate_grad_batches: 6 |