Add new config files

5 files changed, 372 insertions, 0 deletions

diff --git a/training/conf/datamodule/iam_lines.yaml b/training/conf/datamodule/iam_lines.yaml
new file mode 100644
index 0000000..d3c2af8
--- /dev/null
+++ b/training/conf/datamodule/iam_lines.yaml
@@ -0,0 +1,7 @@
+_target_: text_recognizer.data.iam_lines.IAMLines
+batch_size: 8
+num_workers: 12
+train_fraction: 0.8
+augment: true
+pin_memory: false
+# word_pieces: true
diff --git a/training/conf/experiment/cnn_htr_char_lines.yaml b/training/conf/experiment/cnn_htr_char_lines.yaml
new file mode 100644
index 0000000..08d9282
--- /dev/null
+++ b/training/conf/experiment/cnn_htr_char_lines.yaml
@@ -0,0 +1,109 @@
+# @package _global_
+
+defaults:
+  - override /mapping: null
+  - override /criterion: null
+  - override /datamodule: null
+  - override /network: null
+  - override /model: null
+  - override /lr_schedulers: null
+  - override /optimizers: null
+
+
+criterion:
+  _target_: torch.nn.CrossEntropyLoss
+  ignore_index: 3
+    
+mapping:
+  _target_: text_recognizer.data.emnist_mapping.EmnistMapping
+  # extra_symbols: [ "\n" ]
+
+optimizers:
+  madgrad:
+    _target_: madgrad.MADGRAD
+    lr: 1.0e-4
+    momentum: 0.9
+    weight_decay: 0
+    eps: 1.0e-6
+
+    parameters: network
+
+lr_schedulers:
+  network:
+    _target_: torch.optim.lr_scheduler.CosineAnnealingLR
+    T_max: 1024
+    eta_min: 4.5e-6
+    last_epoch: -1
+    interval: epoch
+    monitor: val/loss
+
+datamodule:
+  _target_: text_recognizer.data.iam_lines.IAMLines
+  batch_size: 24
+  num_workers: 12
+  train_fraction: 0.8
+  augment: true
+  pin_memory: false
+
+network:
+  _target_: text_recognizer.networks.conv_transformer.ConvTransformer
+  input_dims: [1, 56, 1024]
+  hidden_dim: 128
+  encoder_dim: 1280
+  dropout_rate: 0.2
+  num_classes: 58
+  pad_index: 3
+  encoder:
+    _target_: text_recognizer.networks.encoders.efficientnet.EfficientNet
+    arch: b0
+    out_channels: 1280
+    stochastic_dropout_rate: 0.2
+    bn_momentum: 0.99
+    bn_eps: 1.0e-3
+  decoder:
+    _target_: text_recognizer.networks.transformer.Decoder
+    dim: 128
+    depth: 3 
+    num_heads: 4
+    attn_fn: text_recognizer.networks.transformer.attention.Attention
+    attn_kwargs:
+      dim_head: 32
+      dropout_rate: 0.2
+    norm_fn: torch.nn.LayerNorm
+    ff_fn: text_recognizer.networks.transformer.mlp.FeedForward
+    ff_kwargs:
+      dim_out: null
+      expansion_factor: 4
+      glu: true
+      dropout_rate: 0.2
+    cross_attend: true
+    pre_norm: true
+    rotary_emb: null
+
+model:
+  _target_: text_recognizer.models.transformer.TransformerLitModel
+  max_output_len: 89
+  start_token: <s>
+  end_token: <e>
+  pad_token: <p>
+
+trainer:
+  _target_: pytorch_lightning.Trainer
+  stochastic_weight_avg: false
+  auto_scale_batch_size: binsearch
+  auto_lr_find: false
+  gradient_clip_val: 0
+  fast_dev_run: false
+  gpus: 1
+  precision: 16
+  max_epochs: 1024
+  terminate_on_nan: true
+  weights_summary: top
+  limit_train_batches: 1.0 
+  limit_val_batches: 1.0
+  limit_test_batches: 1.0
+  resume_from_checkpoint: null
+  accumulate_grad_batches: 4
+  overfit_batches: 0.0
+
+# summary: [[1, 1, 56, 1024], [1, 89]]
diff --git a/training/conf/experiment/vqgan_htr_char.yaml b/training/conf/experiment/vqgan_htr_char.yaml
new file mode 100644
index 0000000..426524f
--- /dev/null
+++ b/training/conf/experiment/vqgan_htr_char.yaml
@@ -0,0 +1,61 @@
+# @package _global_
+
+defaults:
+  - override /mapping: null
+  - override /network: null
+  - override /model: null
+
+mapping:
+  _target_: text_recognizer.data.emnist_mapping.EmnistMapping
+  extra_symbols: [ "\n" ]
+
+datamodule:
+  word_pieces: false
+  batch_size: 8
+  augment: false
+
+criterion:
+  ignore_index: 3
+
+network:
+  _target_: text_recognizer.networks.vq_transformer.VqTransformer
+  input_dims: [1, 576, 640]
+  encoder_dim: 32
+  hidden_dim: 256
+  dropout_rate: 0.1
+  num_classes: 58
+  pad_index: 3
+  no_grad: true
+  decoder:
+    _target_: text_recognizer.networks.transformer.Decoder
+    dim: 256
+    depth: 2
+    num_heads: 8
+    attn_fn: text_recognizer.networks.transformer.attention.Attention
+    attn_kwargs:
+      dim_head: 32
+      dropout_rate: 0.2
+    norm_fn: torch.nn.LayerNorm
+    ff_fn: text_recognizer.networks.transformer.mlp.FeedForward
+    ff_kwargs:
+      dim_out: null
+      expansion_factor: 4
+      glu: true
+      dropout_rate: 0.2
+    cross_attend: true
+    pre_norm: true
+    rotary_emb: null
+  pretrained_encoder_path: "training/logs/runs/2021-09-25/23-07-28"
+
+model:
+  _target_: text_recognizer.models.vq_transformer.VqTransformerLitModel
+  start_token: <s>
+  end_token: <e>
+  pad_token: <p>
+  max_output_len: 682  # 451
+  alpha: 1.0
+
+trainer:
+  max_epochs: 64
+  limit_train_batches: 0.1
+  limit_val_batches: 0.1
diff --git a/training/conf/experiment/vqgan_htr_char_iam_lines.yaml b/training/conf/experiment/vqgan_htr_char_iam_lines.yaml
new file mode 100644
index 0000000..9f4791f
--- /dev/null
+++ b/training/conf/experiment/vqgan_htr_char_iam_lines.yaml
@@ -0,0 +1,90 @@
+# @package _global_
+
+defaults:
+  - override /mapping: null
+  - override /criterion: null
+  - override /datamodule: null
+  - override /network: null
+  - override /model: null
+  - override /lr_schedulers: null
+  # - override /optimizers: null
+
+
+criterion:
+  _target_: text_recognizer.criterions.label_smoothing.LabelSmoothingLoss
+  smoothing: 0.1 
+  ignore_index: 3
+    
+mapping:
+  _target_: text_recognizer.data.emnist_mapping.EmnistMapping
+  # extra_symbols: [ "\n" ]
+
+lr_schedulers:
+  network:
+    _target_: torch.optim.lr_scheduler.CosineAnnealingLR
+    T_max: 512
+    eta_min: 4.5e-6
+    last_epoch: -1
+    interval: epoch
+    monitor: val/loss
+
+datamodule:
+  _target_: text_recognizer.data.iam_lines.IAMLines
+  batch_size: 4
+  num_workers: 12
+  train_fraction: 0.8
+  augment: false
+  pin_memory: false
+
+
+# optimizers:
+#   - _target_: madgrad.MADGRAD
+#     lr: 2.0e-4
+#     momentum: 0.9
+#     weight_decay: 0
+#     eps: 1.0e-7
+#     parameters: network
+
+network:
+  _target_: text_recognizer.networks.vq_transformer.VqTransformer
+  input_dims: [1, 56, 1024]
+  encoder_dim: 32
+  hidden_dim: 32
+  dropout_rate: 0.1
+  num_classes: 58
+  pad_index: 3
+  no_grad: true
+  decoder:
+    _target_: text_recognizer.networks.transformer.Decoder
+    dim: 32
+    depth: 4
+    num_heads: 8
+    attn_fn: text_recognizer.networks.transformer.attention.Attention
+    attn_kwargs:
+      dim_head: 32
+      dropout_rate: 0.2
+    norm_fn: torch.nn.LayerNorm
+    ff_fn: text_recognizer.networks.transformer.mlp.FeedForward
+    ff_kwargs:
+      dim_out: null
+      expansion_factor: 4
+      glu: true
+      dropout_rate: 0.2
+    cross_attend: true
+    pre_norm: true
+    rotary_emb: null
+  pretrained_encoder_path: "training/logs/runs/2021-09-26/23-27-57"
+
+model:
+  _target_: text_recognizer.models.vq_transformer.VqTransformerLitModel
+  start_token: <s>
+  end_token: <e>
+  pad_token: <p>
+  max_output_len: 89  # 451
+  alpha: 0.0
+
+trainer:
+  max_epochs: 512
+  # limit_train_batches: 0.1
+  # limit_val_batches: 0.1
+  # gradient_clip_val: 0.5
diff --git a/training/conf/experiment/vqgan_iam_lines.yaml b/training/conf/experiment/vqgan_iam_lines.yaml
new file mode 100644
index 0000000..8bdf415
--- /dev/null
+++ b/training/conf/experiment/vqgan_iam_lines.yaml
@@ -0,0 +1,105 @@
+# @package _global_
+
+defaults:
+  - override /network: null
+  - override /criterion: null
+  - override /datamodule: null
+  - override /model: lit_vqgan
+  - override /callbacks: wandb_vae
+  - override /optimizers: null
+  - override /lr_schedulers: null
+
+criterion:
+  _target_: text_recognizer.criterions.vqgan_loss.VQGANLoss
+  reconstruction_loss:
+    _target_: torch.nn.BCEWithLogitsLoss
+    reduction: mean
+  discriminator:
+    _target_: text_recognizer.criterions.n_layer_discriminator.NLayerDiscriminator
+    in_channels: 1
+    num_channels: 64
+    num_layers: 3
+  commitment_weight: 0.25
+  discriminator_weight: 0.8
+  discriminator_factor: 1.0
+  discriminator_iter_start: 1.5e4
+
+datamodule:
+  _target_: text_recognizer.data.iam_lines.IAMLines
+  batch_size: 24
+  num_workers: 12
+  train_fraction: 0.8
+  augment: true
+  pin_memory: false
+
+lr_schedulers:
+  generator:
+    _target_: torch.optim.lr_scheduler.CosineAnnealingLR
+    T_max: 64
+    eta_min: 4.5e-6
+    last_epoch: -1
+    interval: epoch
+    monitor: val/loss
+#   discriminator:
+#     _target_: torch.optim.lr_scheduler.CosineAnnealingLR
+#     T_max: 64
+#     eta_min: 0.0
+#     last_epoch: -1
+# 
+#     interval: epoch
+#     monitor: val/loss
+
+optimizers:
+  generator:
+    _target_: madgrad.MADGRAD
+    lr: 1.0e-4
+    momentum: 0.5
+    weight_decay: 0
+    eps: 1.0e-7
+    parameters: network
+
+  discriminator:
+    _target_: madgrad.MADGRAD
+    lr: 4.5e-6
+    momentum: 0.5
+    weight_decay: 0
+    eps: 1.0e-6
+    parameters: loss_fn.discriminator
+
+network:
+  _target_: text_recognizer.networks.vqvae.vqvae.VQVAE
+  hidden_dim: 256
+  embedding_dim: 32
+  num_embeddings: 512
+  decay: 0.99
+  encoder:
+    _target_: text_recognizer.networks.vqvae.encoder.Encoder
+    in_channels: 1 
+    hidden_dim: 32
+    channels_multipliers: [1, 4, 8]
+    dropout_rate: 0.0
+    activation: mish
+    use_norm: true
+    num_residuals: 2
+    residual_channels: 32
+  decoder:
+    _target_: text_recognizer.networks.vqvae.decoder.Decoder
+    out_channels: 1 
+    hidden_dim: 32
+    channels_multipliers: [8, 4, 1]
+    dropout_rate: 0.0
+    activation: mish
+    use_norm: true
+    num_residuals: 2 
+    residual_channels: 32
+
+trainer:
+  max_epochs: 64
+  # limit_train_batches: 0.1
+  # limit_val_batches: 0.1
+  # gradient_clip_val: 100
+
+# tune: false
+# train: true
+# test: false
+summary: [2, 1, 56, 1024]