diff options
Diffstat (limited to 'src/training/experiments/line_ctc_experiment.yml')
| -rw-r--r-- | src/training/experiments/line_ctc_experiment.yml | 103 |
1 files changed, 48 insertions, 55 deletions
diff --git a/src/training/experiments/line_ctc_experiment.yml b/src/training/experiments/line_ctc_experiment.yml index c21c6a2..337c830 100644 --- a/src/training/experiments/line_ctc_experiment.yml +++ b/src/training/experiments/line_ctc_experiment.yml @@ -1,55 +1,46 @@ -experiment_group: Sample Experiments +experiment_group: Lines Experiments experiments: - train_args: batch_size: 64 - max_epochs: 32 + max_epochs: &max_epochs 32 dataset: - type: EmnistLinesDataset + type: IamLinesDataset args: - subsample_fraction: 0.33 - max_length: 34 - min_overlap: 0 - max_overlap: 0.33 - num_samples: 10000 - seed: 4711 - blank: true + subsample_fraction: null + transform: null + target_transform: null train_args: - num_workers: 6 + num_workers: 8 train_fraction: 0.85 model: LineCTCModel metrics: [cer, wer] network: type: LineRecurrentNetwork args: - # encoder: ResidualNetworkEncoder - # encoder_args: + # backbone: ResidualNetwork + # backbone_args: # in_channels: 1 - # num_classes: 81 - # depths: [2, 2] - # block_sizes: [64, 128] - # activation: SELU + # num_classes: 64 # Embedding + # depths: [2,2] + # block_sizes: [32, 64] + # activation: selu # stn: false - encoder: WideResidualNetwork - encoder_args: - in_channels: 1 - num_classes: 81 - depth: 16 - num_layers: 4 - width_factor: 2 - dropout_rate: 0.2 - activation: selu - use_decoder: false - flatten: true - input_size: 256 - hidden_size: 128 + backbone: ResidualNetwork + backbone_args: + pretrained: training/experiments/CharacterModel_EmnistDataset_ResidualNetwork/0920_010806/model/best.pt + freeze: false + flatten: false + input_size: 64 + hidden_size: 64 + bidirectional: true num_layers: 2 - num_classes: 81 - patch_size: [28, 14] - stride: [1, 5] + num_classes: 80 + patch_size: [28, 18] + stride: [1, 4] criterion: type: CTCLoss args: - blank: 80 + blank: 79 optimizer: type: AdamW args: @@ -58,40 +49,42 @@ experiments: eps: 1.e-08 weight_decay: 5.e-4 amsgrad: false - # lr_scheduler: - # type: OneCycleLR - # args: - # max_lr: 1.e-03 - # epochs: null - # anneal_strategy: linear lr_scheduler: - type: CosineAnnealingLR + type: OneCycleLR args: - T_max: null + max_lr: 1.e-02 + epochs: *max_epochs + anneal_strategy: cos + pct_start: 0.475 + cycle_momentum: true + base_momentum: 0.85 + max_momentum: 0.9 + div_factor: 10 + final_div_factor: 10000 + interval: step + # lr_scheduler: + # type: CosineAnnealingLR + # args: + # T_max: *max_epochs swa_args: - start: 4 + start: 48 lr: 5.e-2 - callbacks: [Checkpoint, ProgressBar, WandbCallback, WandbImageLogger, SWA] # EarlyStopping, OneCycleLR] + callbacks: [Checkpoint, ProgressBar, WandbCallback, WandbImageLogger, EarlyStopping] callback_args: Checkpoint: monitor: val_loss mode: min ProgressBar: - epochs: null - log_batch_frequency: 100 - # EarlyStopping: - # monitor: val_loss - # min_delta: 0.0 - # patience: 5 - # mode: min + epochs: *max_epochs + EarlyStopping: + monitor: val_loss + min_delta: 0.0 + patience: 10 + mode: min WandbCallback: log_batch_frequency: 10 WandbImageLogger: num_examples: 6 - # OneCycleLR: - # null - SWA: - null verbosity: 1 # 0, 1, 2 resume_experiment: null test: true |