diff options
Diffstat (limited to 'training')
5 files changed, 46 insertions, 142 deletions
diff --git a/training/conf/datamodule/iam_extended_paragraphs.yaml b/training/conf/datamodule/iam_extended_paragraphs.yaml index f53e5b6..0921b25 100644 --- a/training/conf/datamodule/iam_extended_paragraphs.yaml +++ b/training/conf/datamodule/iam_extended_paragraphs.yaml @@ -2,7 +2,10 @@ _target_: text_recognizer.data.iam_extended_paragraphs.IAMExtendedParagraphs batch_size: 4 num_workers: 12 train_fraction: 0.8 -pin_memory: false +pin_memory: true transform: transform/paragraphs.yaml test_transform: test_transform/paragraphs.yaml target_transform: target_transform/pad.yaml +mapping: + _target_: text_recognizer.data.mappings.EmnistMapping + extra_symbols: ["\n"] diff --git a/training/conf/datamodule/iam_lines.yaml b/training/conf/datamodule/iam_lines.yaml index 19c72f8..85a023b 100644 --- a/training/conf/datamodule/iam_lines.yaml +++ b/training/conf/datamodule/iam_lines.yaml @@ -1,7 +1,9 @@ _target_: text_recognizer.data.iam_lines.IAMLines batch_size: 8 num_workers: 12 -train_fraction: 0.8 +train_fraction: 0.9 pin_memory: true transform: transform/lines.yaml test_transform: test_transform/lines.yaml +mapping: + _target_: text_recognizer.data.mappings.EmnistMapping diff --git a/training/conf/experiment/conv_transformer_lines.yaml b/training/conf/experiment/conv_transformer_lines.yaml index 55dc896..8404cd1 100644 --- a/training/conf/experiment/conv_transformer_lines.yaml +++ b/training/conf/experiment/conv_transformer_lines.yaml @@ -1,17 +1,16 @@ ---- # @package _global_ defaults: - - override /mapping: null + - override /mapping: characters - override /criterion: cross_entropy - override /callbacks: htr - override /datamodule: iam_lines - - override /network: null - - override /model: null + - override /network: conv_transformer + - override /model: lit_transformer - override /lr_schedulers: null - override /optimizers: null -epochs: &epochs 200 +epochs: &epochs 512 ignore_index: &ignore_index 3 num_classes: &num_classes 57 max_output_len: &max_output_len 89 @@ -21,10 +20,6 @@ criterion: ignore_index: *ignore_index # label_smoothing: 0.1 -mapping: &mapping - mapping: - _target_: text_recognizer.data.mappings.emnist.EmnistMapping - callbacks: stochastic_weight_averaging: _target_: pytorch_lightning.callbacks.StochasticWeightAveraging @@ -60,99 +55,19 @@ lr_schedulers: datamodule: batch_size: 16 - num_workers: 12 - train_fraction: 0.9 - pin_memory: true - << : *mapping - -encoder: &encoder - _target_: text_recognizer.networks.efficientnet.efficientnet.EfficientNet - arch: b0 - stochastic_dropout_rate: 0.2 - bn_momentum: 0.99 - bn_eps: 1.0e-3 - depth: 5 - -rotary_embedding: &rotary_embedding - rotary_embedding: - _target_: > - text_recognizer.networks.transformer.embeddings.rotary.RotaryEmbedding - dim: 64 - -attn: &attn - dim: &hidden_dim 512 - num_heads: 4 - dim_head: 64 - dropout_rate: &dropout_rate 0.4 - -decoder: &decoder - _target_: text_recognizer.networks.transformer.decoder.Decoder - depth: 6 - has_pos_emb: true - block: - _target_: text_recognizer.networks.transformer.decoder.DecoderBlock - self_attn: - _target_: text_recognizer.networks.transformer.attention.Attention - <<: *attn - causal: true - <<: *rotary_embedding - cross_attn: - _target_: text_recognizer.networks.transformer.attention.Attention - <<: *attn - causal: false - norm: - _target_: text_recognizer.networks.transformer.norm.RMSNorm - dim: *hidden_dim - ff: - _target_: text_recognizer.networks.transformer.mlp.FeedForward - dim: *hidden_dim - dim_out: null - expansion_factor: 2 - glu: true - dropout_rate: *dropout_rate - -pixel_pos_embedding: &pixel_pos_embedding - _target_: > - text_recognizer.networks.transformer.embeddings.axial.AxialPositionalEmbedding - dim: *hidden_dim - shape: &shape [3, 64] network: - _target_: text_recognizer.networks.conv_transformer.ConvTransformer input_dims: [1, 1, 56, 1024] - hidden_dim: *hidden_dim num_classes: *num_classes pad_index: *ignore_index - encoder: - <<: *encoder decoder: - <<: *decoder - pixel_pos_embedding: - <<: *pixel_pos_embedding + depth: 10 + pixel_embedding: + shape: [7, 128] model: - _target_: text_recognizer.models.transformer.LitTransformer - <<: *mapping max_output_len: *max_output_len - start_token: <s> - end_token: <e> - pad_token: <p> trainer: - _target_: pytorch_lightning.Trainer - stochastic_weight_avg: true - auto_scale_batch_size: binsearch - auto_lr_find: false gradient_clip_val: 0.5 - fast_dev_run: false - gpus: 1 - precision: 16 max_epochs: *epochs - terminate_on_nan: true - weights_summary: null - limit_train_batches: 1.0 - limit_val_batches: 1.0 - limit_test_batches: 1.0 - resume_from_checkpoint: null - accumulate_grad_batches: 1 - overfit_batches: 0 diff --git a/training/conf/experiment/conv_transformer_paragraphs.yaml b/training/conf/experiment/conv_transformer_paragraphs.yaml index cc81645..1465e62 100644 --- a/training/conf/experiment/conv_transformer_paragraphs.yaml +++ b/training/conf/experiment/conv_transformer_paragraphs.yaml @@ -1,7 +1,7 @@ # @package _global_ defaults: - - override /mapping: emnist + - override /mapping: characters - override /criterion: cross_entropy - override /callbacks: htr - override /datamodule: iam_extended_paragraphs @@ -10,7 +10,7 @@ defaults: - override /lr_schedulers: null - override /optimizers: null -epochs: &epochs 200 +epochs: &epochs 629 ignore_index: &ignore_index 3 summary: [[1, 1, 576, 640], [1, 682]] @@ -52,12 +52,8 @@ lr_schedulers: monitor: val/loss datamodule: - _target_: text_recognizer.data.iam_extended_paragraphs.IAMExtendedParagraphs batch_size: 6 - num_workers: 12 - train_fraction: 0.8 - pin_memory: true - << : *mapping + train_fraction: 0.9 trainer: max_epochs: *epochs diff --git a/training/conf/network/decoder/transformer_decoder.yaml b/training/conf/network/decoder/transformer_decoder.yaml index 7dced16..4588ee9 100644 --- a/training/conf/network/decoder/transformer_decoder.yaml +++ b/training/conf/network/decoder/transformer_decoder.yaml @@ -1,42 +1,30 @@ -_target_: text_recognizer.networks.transformer.Decoder +_target_: text_recognizer.networks.transformer.decoder.Decoder depth: 4 -pre_norm: true -local_depth: 2 -has_pos_emb: true -self_attn: - _target_: text_recognizer.networks.transformer.attention.Attention - dim: 64 - num_heads: 4 - dim_head: 64 - dropout_rate: 0.05 - causal: true - rotary_embedding: - _target_: text_recognizer.networks.transformer.embeddings.rotary.RotaryEmbedding - dim: 128 -local_self_attn: - _target_: text_recognizer.networks.transformer.local_attention.LocalAttention - dim: 64 - num_heads: 4 - dim_head: 64 - dropout_rate: 0.05 - window_size: 22 - look_back: 1 - rotary_embedding: - _target_: text_recognizer.networks.transformer.embeddings.rotary.RotaryEmbedding - dim: 128 -cross_attn: - _target_: text_recognizer.networks.transformer.attention.Attention - dim: 64 - num_heads: 4 - dim_head: 64 - dropout_rate: 0.05 - causal: false -norm: - _target_: text_recognizer.networks.transformer.norm.ScaleNorm - normalized_shape: 192 -ff: - _target_: text_recognizer.networks.transformer.mlp.FeedForward - dim_out: null - expansion_factor: 4 - glu: true - dropout_rate: 0.2 +block: + _target_: text_recognizer.networks.transformer.decoder.DecoderBlock + self_attn: + _target_: text_recognizer.networks.transformer.attention.Attention + dim: 64 + num_heads: 4 + dim_head: 64 + dropout_rate: 0.05 + causal: true + rotary_embedding: + _target_: text_recognizer.networks.transformer.embeddings.rotary.RotaryEmbedding + dim: 128 + cross_attn: + _target_: text_recognizer.networks.transformer.attention.Attention + dim: 64 + num_heads: 4 + dim_head: 64 + dropout_rate: 0.05 + causal: false + norm: + _target_: text_recognizer.networks.transformer.norm.RMSNorm + normalized_shape: 192 + ff: + _target_: text_recognizer.networks.transformer.mlp.FeedForward + dim_out: null + expansion_factor: 4 + glu: true + dropout_rate: 0.2 |