diff options
Diffstat (limited to 'training/conf/experiment')
-rw-r--r-- | training/conf/experiment/conv_perceiver_lines.yaml | 76 | ||||
-rw-r--r-- | training/conf/experiment/vq_transformer_lines.yaml | 81 |
2 files changed, 0 insertions, 157 deletions
diff --git a/training/conf/experiment/conv_perceiver_lines.yaml b/training/conf/experiment/conv_perceiver_lines.yaml deleted file mode 100644 index 26fe232..0000000 --- a/training/conf/experiment/conv_perceiver_lines.yaml +++ /dev/null @@ -1,76 +0,0 @@ -# @package _global_ - -defaults: - - override /criterion: cross_entropy - - override /callbacks: htr - - override /datamodule: iam_lines - - override /network: conv_perceiver - - override /model: lit_perceiver - - override /lr_scheduler: null - - override /optimizer: null - -tags: [lines, perceiver] -epochs: &epochs 260 -ignore_index: &ignore_index 3 -num_classes: &num_classes 57 -max_output_len: &max_output_len 89 -summary: [[1, 1, 56, 1024]] - -logger: - wandb: - tags: ${tags} - -criterion: - ignore_index: *ignore_index - # label_smoothing: 0.1 - -callbacks: - stochastic_weight_averaging: - _target_: pytorch_lightning.callbacks.StochasticWeightAveraging - swa_epoch_start: 0.75 - swa_lrs: 1.0e-5 - annealing_epochs: 10 - annealing_strategy: cos - device: null - -optimizer: - _target_: adan_pytorch.Adan - lr: 1.0e-4 - betas: [0.02, 0.08, 0.01] - weight_decay: 0.02 - -lr_scheduler: - _target_: torch.optim.lr_scheduler.ReduceLROnPlateau - mode: min - factor: 0.8 - patience: 10 - threshold: 1.0e-4 - threshold_mode: rel - cooldown: 0 - min_lr: 1.0e-5 - eps: 1.0e-8 - verbose: false - interval: epoch - monitor: val/cer - -datamodule: - batch_size: 8 - train_fraction: 0.95 - -network: - input_dims: [1, 1, 56, 1024] - num_classes: *num_classes - pad_index: *ignore_index - encoder: - depth: 5 - decoder: - depth: 6 - -model: - max_output_len: *max_output_len - -trainer: - gradient_clip_val: 1.0 - stochastic_weight_avg: true - max_epochs: *epochs - accumulate_grad_batches: 1 diff --git a/training/conf/experiment/vq_transformer_lines.yaml b/training/conf/experiment/vq_transformer_lines.yaml deleted file mode 100644 index dbd8a3b..0000000 --- a/training/conf/experiment/vq_transformer_lines.yaml +++ /dev/null @@ -1,81 +0,0 @@ -# @package _global_ - -defaults: - - override /criterion: cross_entropy - - override /callbacks: htr - - override /datamodule: iam_lines - - override /network: vq_transformer - - override /model: lit_vq_transformer - - override /lr_scheduler: null - - override /optimizer: null - -tags: [lines] -epochs: &epochs 200 -ignore_index: &ignore_index 3 -num_classes: &num_classes 57 -max_output_len: &max_output_len 89 -summary: [[1, 1, 56, 1024], [1, 89]] - -logger: - wandb: - tags: ${tags} - # id: 342qvr1p - -criterion: - ignore_index: *ignore_index - label_smoothing: 0.05 - -callbacks: - stochastic_weight_averaging: - _target_: pytorch_lightning.callbacks.StochasticWeightAveraging - swa_epoch_start: 0.75 - swa_lrs: 1.0e-5 - annealing_epochs: 10 - annealing_strategy: cos - device: null - -optimizer: - _target_: torch.optim.RAdam - lr: 3.0e-4 - betas: [0.9, 0.999] - weight_decay: 0 - eps: 1.0e-8 - -lr_scheduler: - _target_: torch.optim.lr_scheduler.ReduceLROnPlateau - mode: min - factor: 0.8 - patience: 10 - threshold: 1.0e-4 - threshold_mode: rel - cooldown: 0 - min_lr: 1.0e-5 - eps: 1.0e-8 - verbose: false - interval: epoch - monitor: val/cer - -datamodule: - batch_size: 8 - train_fraction: 0.95 - -network: - input_dims: [1, 1, 56, 1024] - num_classes: *num_classes - pad_index: *ignore_index - encoder: - depth: 5 - decoder: - depth: 6 - pixel_embedding: - shape: [1, 127] - -model: - max_output_len: *max_output_len - vq_loss_weight: 0.1 - -trainer: - gradient_clip_val: 1.0 - max_epochs: *epochs - accumulate_grad_batches: 1 - # resume_from_checkpoint: /home/aktersnurra/projects/text-recognizer/training/logs/runs/2022-06-27/00-37-40/checkpoints/last.ckpt |