diff options
| author | Gustaf Rydholm <gustaf.rydholm@gmail.com> | 2021-10-25 22:32:10 +0200 |
|---|---|---|
| committer | Gustaf Rydholm <gustaf.rydholm@gmail.com> | 2021-10-25 22:32:10 +0200 |
| commit | 41c3e99fe57874ba1855c893bf47087d474ec6b8 (patch) | |
| tree | 2ba286209251f7ec8d7ab4cf061cffe5de86f398 /training/conf/experiment/cnn_htr_wp_lines.yaml | |
| parent | da29c1cf4d062087f1b29dc9402ee6384203b690 (diff) | |
Updates to configs
Diffstat (limited to 'training/conf/experiment/cnn_htr_wp_lines.yaml')
| -rw-r--r-- | training/conf/experiment/cnn_htr_wp_lines.yaml | 52 |
1 files changed, 28 insertions, 24 deletions
diff --git a/training/conf/experiment/cnn_htr_wp_lines.yaml b/training/conf/experiment/cnn_htr_wp_lines.yaml index f467b74..6cdd023 100644 --- a/training/conf/experiment/cnn_htr_wp_lines.yaml +++ b/training/conf/experiment/cnn_htr_wp_lines.yaml @@ -1,5 +1,8 @@ +# @package _global_ + defaults: - override /mapping: null + - override /callbacks: htr - override /criterion: null - override /datamodule: null - override /network: null @@ -7,28 +10,27 @@ defaults: - override /lr_schedulers: null - override /optimizers: null -epochs: &epochs 256 +epochs: &epochs 512 ignore_index: &ignore_index 1000 num_classes: &num_classes 1006 max_output_len: &max_output_len 72 criterion: - _target_: torch.nn.CrossEntropyLoss - ignore_index: *ignore_index - # _target_: text_recognizer.criterions.label_smoothing.LabelSmoothingLoss - # smoothing: 0.1 + # _target_: torch.nn.CrossEntropyLoss # ignore_index: *ignore_index + _target_: text_recognizer.criterions.label_smoothing.LabelSmoothingLoss + smoothing: 0.1 + ignore_index: *ignore_index -mapping: - _target_: text_recognizer.data.word_piece_mapping.WordPieceMapping - num_features: 1000 - tokens: iamdb_1kwp_tokens_1000.txt - lexicon: iamdb_1kwp_lex_1000.txt - data_dir: null - use_words: false - prepend_wordsep: false - special_tokens: [ <s>, <e>, <p> ] - # extra_symbols: [ "\n" ] +mapping: &mapping + mapping: + _target_: text_recognizer.data.mappings.word_piece_mapping.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> ] callbacks: stochastic_weight_averaging: @@ -42,7 +44,7 @@ callbacks: optimizers: madgrad: _target_: madgrad.MADGRAD - lr: 3.0e-4 + lr: 1.0e-4 momentum: 0.9 weight_decay: 0 eps: 1.0e-6 @@ -52,11 +54,11 @@ optimizers: lr_schedulers: network: _target_: torch.optim.lr_scheduler.OneCycleLR - max_lr: 3.0e-4 + max_lr: 1.0e-4 total_steps: null epochs: *epochs - steps_per_epoch: 90 - pct_start: 0.1 + steps_per_epoch: 179 + pct_start: 0.03 anneal_strategy: cos cycle_momentum: true base_momentum: 0.85 @@ -72,17 +74,19 @@ lr_schedulers: datamodule: _target_: text_recognizer.data.iam_lines.IAMLines - batch_size: 32 num_workers: 12 - train_fraction: 0.8 - augment: true + batch_size: 16 + train_fraction: 0.9 pin_memory: true - word_pieces: true + transform: transform/iam_lines.yaml + test_transform: transform/iam_lines.yaml + target_transform: target_transform/word_piece_iam.yaml + <<: *mapping network: _target_: text_recognizer.networks.conv_transformer.ConvTransformer input_dims: [1, 56, 1024] - hidden_dim: &hidden_dim 128 + hidden_dim: &hidden_dim 256 encoder_dim: 1280 dropout_rate: 0.2 num_classes: *num_classes |