summaryrefslogtreecommitdiff
path: root/training/conf
diff options
context:
space:
mode:
Diffstat (limited to 'training/conf')
-rw-r--r--training/conf/experiment/conformer_lines.yaml88
-rw-r--r--training/conf/network/conformer.yaml36
2 files changed, 0 insertions, 124 deletions
diff --git a/training/conf/experiment/conformer_lines.yaml b/training/conf/experiment/conformer_lines.yaml
deleted file mode 100644
index 06e761e..0000000
--- a/training/conf/experiment/conformer_lines.yaml
+++ /dev/null
@@ -1,88 +0,0 @@
-# @package _global_
-
-defaults:
- - override /mapping: null
- - override /criterion: ctc
- - override /callbacks: htr
- - override /datamodule: iam_lines
- - override /network: conformer
- - override /model: null
- - override /lr_schedulers: null
- - override /optimizers: null
-
-epochs: &epochs 999
-num_classes: &num_classes 57
-max_output_len: &max_output_len 89
-summary: [[1, 56, 1024]]
-
-mapping: &mapping
- mapping:
- _target_: text_recognizer.data.mappings.EmnistMapping
-
-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
-
-optimizers:
- radam:
- _target_: torch.optim.RAdam
- lr: 3.0e-4
- betas: [0.9, 0.999]
- weight_decay: 0
- eps: 1.0e-8
- parameters: network
-
-lr_schedulers:
- network:
- _target_: torch.optim.lr_scheduler.ReduceLROnPlateau
- mode: min
- factor: 0.5
- 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/loss
-
-datamodule:
- batch_size: 8
- num_workers: 12
- train_fraction: 0.9
- pin_memory: true
- << : *mapping
-
-model:
- _target_: text_recognizer.models.conformer.LitConformer
- <<: *mapping
- max_output_len: *max_output_len
- start_token: <s>
- end_token: <e>
- pad_token: <p>
- blank_token: <b>
-
-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: 1
- overfit_batches: 0
diff --git a/training/conf/network/conformer.yaml b/training/conf/network/conformer.yaml
deleted file mode 100644
index 1d72dd5..0000000
--- a/training/conf/network/conformer.yaml
+++ /dev/null
@@ -1,36 +0,0 @@
-_target_: text_recognizer.networks.conformer.Conformer
-depth: 8
-num_classes: 57
-dim: &dim 144
-dim_gru: 144
-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: 64
- shape: [6, 127]
- channels: 64
- height: 6
- dim: *dim
- depth: 3
- dropout: 0.1