diff options
Diffstat (limited to 'training/conf/experiment/barlow_twins.yaml')
| -rw-r--r-- | training/conf/experiment/barlow_twins.yaml | 12 |
1 files changed, 5 insertions, 7 deletions
diff --git a/training/conf/experiment/barlow_twins.yaml b/training/conf/experiment/barlow_twins.yaml index e3586bf..cc1295d 100644 --- a/training/conf/experiment/barlow_twins.yaml +++ b/training/conf/experiment/barlow_twins.yaml @@ -8,14 +8,12 @@ defaults: - override /lr_schedulers: null - override /optimizers: null - -print_config: true epochs: &epochs 1000 summary: [[1, 1, 56, 1024]] criterion: _target_: text_recognizer.criterions.barlow_twins.BarlowTwinsLoss - dim: 2048 + dim: 512 lambda_: 3.9e-3 callbacks: @@ -30,7 +28,7 @@ callbacks: optimizers: madgrad: _target_: madgrad.MADGRAD - lr: 3.0e-4 + lr: 1.0e-3 momentum: 0.9 weight_decay: 1.0e-6 eps: 1.0e-6 @@ -61,7 +59,7 @@ datamodule: _target_: text_recognizer.data.iam_lines.IAMLines batch_size: 16 num_workers: 12 - train_fraction: 0.8 + train_fraction: 0.9 pin_memory: false transform: transform/iam_lines_barlow.yaml test_transform: transform/iam_lines_barlow.yaml @@ -79,7 +77,7 @@ network: bn_eps: 1.0e-3 projector: _target_: text_recognizer.networks.barlow_twins.projector.Projector - dims: [1280, 2048, 2048] + dims: [1280, 512, 512, 512] model: _target_: text_recognizer.models.barlow_twins.BarlowTwinsLitModel @@ -100,5 +98,5 @@ trainer: limit_val_batches: 1.0 limit_test_batches: 1.0 resume_from_checkpoint: null - accumulate_grad_batches: 64 + accumulate_grad_batches: 32 overfit_batches: 0 |