Add vq configs

3 files changed, 151 insertions, 0 deletions

diff --git a/training/conf/experiment/vq_transformer_lines.yaml b/training/conf/experiment/vq_transformer_lines.yaml
new file mode 100644
index 0000000..dbd8a3b
--- /dev/null
+++ b/training/conf/experiment/vq_transformer_lines.yaml
@@ -0,0 +1,81 @@
+# @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
diff --git a/training/conf/model/lit_vq_transformer.yaml b/training/conf/model/lit_vq_transformer.yaml
new file mode 100644
index 0000000..4173151
--- /dev/null
+++ b/training/conf/model/lit_vq_transformer.yaml
@@ -0,0 +1,5 @@
+_target_: text_recognizer.models.LitVqTransformer
+max_output_len: 682
+start_token: <s>
+end_token: <e>
+pad_token: <p>
diff --git a/training/conf/network/vq_transformer.yaml b/training/conf/network/vq_transformer.yaml
new file mode 100644
index 0000000..d62a4b7
--- /dev/null
+++ b/training/conf/network/vq_transformer.yaml
@@ -0,0 +1,65 @@
+_target_: text_recognizer.networks.VqTransformer
+input_dims: [1, 1, 576, 640]
+hidden_dim: &hidden_dim 144
+num_classes: 58
+pad_index: 3
+encoder:
+  _target_: text_recognizer.networks.EfficientNet
+  arch: b0
+  stochastic_dropout_rate: 0.2
+  bn_momentum: 0.99
+  bn_eps: 1.0e-3
+  depth: 5
+  out_channels: *hidden_dim
+decoder:
+  _target_: text_recognizer.networks.transformer.Decoder
+  depth: 6
+  block:
+    _target_: text_recognizer.networks.transformer.DecoderBlock
+    self_attn:
+      _target_: text_recognizer.networks.transformer.Attention
+      dim: *hidden_dim
+      num_heads: 8
+      dim_head: 64
+      dropout_rate: &dropout_rate 0.4
+      causal: true
+      rotary_embedding:
+        _target_: text_recognizer.networks.transformer.RotaryEmbedding
+        dim: 64
+    cross_attn:
+      _target_: text_recognizer.networks.transformer.Attention
+      dim: *hidden_dim
+      num_heads: 8
+      dim_head: 64
+      dropout_rate: *dropout_rate
+      causal: false
+    norm:
+      _target_: text_recognizer.networks.transformer.RMSNorm
+      dim: *hidden_dim
+    ff:
+      _target_: text_recognizer.networks.transformer.FeedForward
+      dim: *hidden_dim
+      dim_out: null
+      expansion_factor: 2
+      glu: true
+      dropout_rate: *dropout_rate
+pixel_embedding:
+  _target_: text_recognizer.networks.transformer.AxialPositionalEmbedding
+  dim: *hidden_dim
+  shape: [18, 79]
+quantizer:
+  _target_: text_recognizer.networks.quantizer.VectorQuantizer
+  input_dim: *hidden_dim
+  codebook:
+    _target_: text_recognizer.networks.quantizer.CosineSimilarityCodebook
+    dim: 16
+    codebook_size: 64
+    kmeans_init: true
+    kmeans_iters: 10
+    decay: 0.8
+    eps: 1.0e-5
+    threshold_dead: 2
+    temperature: 0.0
+  commitment: 0.25
+  ort_reg_weight: 10
+  ort_reg_max_codes: 64