diff options
| -rw-r--r-- | training/conf/experiment/conformer_lines.yaml | 43 | ||||
| -rw-r--r-- | training/conf/experiment/conv_transformer_lines.yaml | 16 | ||||
| -rw-r--r-- | training/conf/experiment/conv_transformer_paragraphs.yaml | 103 | ||||
| -rw-r--r-- | training/conf/mapping/characters.yaml | 2 | ||||
| -rw-r--r-- | training/conf/mapping/word_piece.yaml | 8 | ||||
| -rw-r--r-- | training/conf/model/lit_transformer.yaml | 5 | ||||
| -rw-r--r-- | training/conf/network/conv_transformer.yaml | 60 | ||||
| -rw-r--r-- | training/conf/trainer/default.yaml | 13 |
8 files changed, 76 insertions, 174 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 diff --git a/training/conf/experiment/conv_transformer_lines.yaml b/training/conf/experiment/conv_transformer_lines.yaml index 259e4ea..55dc896 100644 --- a/training/conf/experiment/conv_transformer_lines.yaml +++ b/training/conf/experiment/conv_transformer_lines.yaml @@ -93,12 +93,12 @@ decoder: &decoder _target_: text_recognizer.networks.transformer.decoder.DecoderBlock self_attn: _target_: text_recognizer.networks.transformer.attention.Attention - << : *attn + <<: *attn causal: true - << : *rotary_embedding + <<: *rotary_embedding cross_attn: _target_: text_recognizer.networks.transformer.attention.Attention - << : *attn + <<: *attn causal: false norm: _target_: text_recognizer.networks.transformer.norm.RMSNorm @@ -124,15 +124,15 @@ network: num_classes: *num_classes pad_index: *ignore_index encoder: - << : *encoder + <<: *encoder decoder: - << : *decoder + <<: *decoder pixel_pos_embedding: - << : *pixel_pos_embedding + <<: *pixel_pos_embedding model: - _target_: text_recognizer.models.transformer.TransformerLitModel - << : *mapping + _target_: text_recognizer.models.transformer.LitTransformer + <<: *mapping max_output_len: *max_output_len start_token: <s> end_token: <e> diff --git a/training/conf/experiment/conv_transformer_paragraphs.yaml b/training/conf/experiment/conv_transformer_paragraphs.yaml index afa1785..cc81645 100644 --- a/training/conf/experiment/conv_transformer_paragraphs.yaml +++ b/training/conf/experiment/conv_transformer_paragraphs.yaml @@ -1,31 +1,23 @@ ---- # @package _global_ defaults: - - override /mapping: null + - override /mapping: emnist - override /criterion: cross_entropy - override /callbacks: htr - override /datamodule: iam_extended_paragraphs - - override /network: null - - override /model: null + - override /network: conv_transformer + - override /model: lit_transformer - override /lr_schedulers: null - override /optimizers: null epochs: &epochs 200 ignore_index: &ignore_index 3 -num_classes: &num_classes 58 -max_output_len: &max_output_len 682 summary: [[1, 1, 576, 640], [1, 682]] criterion: ignore_index: *ignore_index # label_smoothing: 0.1 -mapping: &mapping - mapping: - _target_: text_recognizer.data.mappings.emnist.EmnistMapping - extra_symbols: ["\n"] - callbacks: stochastic_weight_averaging: _target_: pytorch_lightning.callbacks.StochasticWeightAveraging @@ -67,94 +59,5 @@ datamodule: 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: 7 - -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 [18, 20] - -network: - _target_: text_recognizer.networks.conv_transformer.ConvTransformer - input_dims: [1, 1, 576, 640] - hidden_dim: *hidden_dim - num_classes: *num_classes - pad_index: *ignore_index - encoder: - << : *encoder - decoder: - << : *decoder - pixel_pos_embedding: - << : *pixel_pos_embedding - -model: - _target_: text_recognizer.models.transformer.TransformerLitModel - << : *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: 2 - overfit_batches: 0 diff --git a/training/conf/mapping/characters.yaml b/training/conf/mapping/characters.yaml index d91c9e5..8cbd55d 100644 --- a/training/conf/mapping/characters.yaml +++ b/training/conf/mapping/characters.yaml @@ -1,2 +1,2 @@ -_target_: text_recognizer.data.mappings.emnist.EmnistMapping +_target_: text_recognizer.data.mappings.EmnistMapping extra_symbols: [ "\n" ] diff --git a/training/conf/mapping/word_piece.yaml b/training/conf/mapping/word_piece.yaml deleted file mode 100644 index 6b4dc07..0000000 --- a/training/conf/mapping/word_piece.yaml +++ /dev/null @@ -1,8 +0,0 @@ -_target_: text_recognizer.data.mappings.word_piece.WordPieceMapping -num_features: 1000 -tokens: iamdb_1kwp_tokens_1000.txt -lexicon: iamdb_1kwp_lex_1000.txt -use_words: false -prepend_wordsep: false -special_tokens: [ <s>, <e>, <p> ] -extra_symbols: [ "\n" ] diff --git a/training/conf/model/lit_transformer.yaml b/training/conf/model/lit_transformer.yaml index 5172533..c1491ec 100644 --- a/training/conf/model/lit_transformer.yaml +++ b/training/conf/model/lit_transformer.yaml @@ -1,5 +1,8 @@ -_target_: text_recognizer.models.transformer.TransformerLitModel +_target_: text_recognizer.models.LitTransformer max_output_len: 682 start_token: <s> end_token: <e> pad_token: <p> +mapping: + _target_: text_recognizer.data.mappings.EmnistMapping + extra_symbols: ["\n"] diff --git a/training/conf/network/conv_transformer.yaml b/training/conf/network/conv_transformer.yaml index 1d61129..54eb028 100644 --- a/training/conf/network/conv_transformer.yaml +++ b/training/conf/network/conv_transformer.yaml @@ -1,11 +1,49 @@ -defaults: - - encoder: efficientnet - - decoder: transformer_decoder - -_target_: text_recognizer.networks.conv_transformer.ConvTransformer -input_dims: [1, 576, 640] -hidden_dim: 128 -encoder_dim: 1280 -dropout_rate: 0.2 -num_classes: 1006 -pad_index: 1002 +_target_: text_recognizer.networks.ConvTransformer +input_dims: [1, 1, 576, 640] +hidden_dim: &hidden_dim 144 +num_classes: 58 +pad_index: 3 +encoder: + _target_: text_recognizer.networks.EfficientNet + arch: b0 + stochastic_dropout_rate: 0.2 + bn_momentum: 0.99 + bn_eps: 1.0e-3 + depth: 3 + out_channels: 128 +decoder: + _target_: text_recognizer.networks.transformer.Decoder + depth: 6 + block: + _target_: text_recognizer.networks.transformer.DecoderBlock + self_attn: + _target_: text_recognizer.networks.transformer.Attention + dim: *hidden_dim + num_heads: 8 + dim_head: 64 + dropout_rate: &dropout_rate 0.4 + causal: true + rotary_embedding: + _target_: text_recognizer.networks.transformer.RotaryEmbedding + dim: 64 + cross_attn: + _target_: text_recognizer.networks.transformer.Attention + dim: *hidden_dim + num_heads: 8 + dim_head: 64 + dropout_rate: *dropout_rate + causal: false + norm: + _target_: text_recognizer.networks.transformer.RMSNorm + dim: *hidden_dim + ff: + _target_: text_recognizer.networks.transformer.FeedForward + dim: *hidden_dim + dim_out: null + expansion_factor: 2 + glu: true + dropout_rate: *dropout_rate +pixel_embedding: + _target_: text_recognizer.networks.transformer.AxialPositionalEmbedding + dim: *hidden_dim + shape: [72, 80] diff --git a/training/conf/trainer/default.yaml b/training/conf/trainer/default.yaml index ef5b018..d4ffcdc 100644 --- a/training/conf/trainer/default.yaml +++ b/training/conf/trainer/default.yaml @@ -1,16 +1,17 @@ _target_: pytorch_lightning.Trainer -stochastic_weight_avg: false +stochastic_weight_avg: true auto_scale_batch_size: binsearch auto_lr_find: false -gradient_clip_val: 0 +gradient_clip_val: 0.5 fast_dev_run: false gpus: 1 precision: 16 -max_epochs: 512 +max_epochs: 256 terminate_on_nan: true -weights_summary: top -limit_train_batches: 1.0 +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 +accumulate_grad_batches: 2 +overfit_batches: 0 |