Updates to cfgs

2 files changed, 164 insertions, 9 deletions

diff --git a/training/conf/experiment/cnn_transformer_paragraphs.yaml b/training/conf/experiment/cnn_transformer_paragraphs.yaml
index b415c29..e9cd254 100644
--- a/training/conf/experiment/cnn_transformer_paragraphs.yaml
+++ b/training/conf/experiment/cnn_transformer_paragraphs.yaml
@@ -10,7 +10,7 @@ defaults:
   - override /optimizers: null
 
 
-epochs: &epochs 512
+epochs: &epochs 1000
 ignore_index: &ignore_index 3
 num_classes: &num_classes 58
 max_output_len: &max_output_len 682
@@ -36,7 +36,7 @@ callbacks:
 optimizers:
   madgrad:
     _target_: madgrad.MADGRAD
-    lr: 3.0e-4
+    lr: 2.0e-4
     momentum: 0.9
     weight_decay: 0
     eps: 1.0e-6
@@ -46,11 +46,11 @@ optimizers:
 lr_schedulers:
   network:
     _target_: torch.optim.lr_scheduler.OneCycleLR
-    max_lr: 3.0e-4
+    max_lr: 2.0e-4
     total_steps: null
     epochs: *epochs
-    steps_per_epoch: 52
-    pct_start: 0.1
+    steps_per_epoch: 79
+    pct_start: 0.3
     anneal_strategy: cos
     cycle_momentum: true
     base_momentum: 0.85
@@ -70,13 +70,13 @@ datamodule:
   num_workers: 12
   train_fraction: 0.8
   augment: true
-  pin_memory: false
+  pin_memory: true
   word_pieces: false
   resize: null
 
 network:
   _target_: text_recognizer.networks.conv_transformer.ConvTransformer
-  input_dims: [1, 56, 1024]
+  input_dims: [1, 576, 640]
   hidden_dim: &hidden_dim 128
   encoder_dim: 1280
   dropout_rate: 0.2
@@ -133,7 +133,7 @@ trainer:
   stochastic_weight_avg: true
   auto_scale_batch_size: binsearch
   auto_lr_find: false
-  gradient_clip_val: 0.5
+  gradient_clip_val: 0.0
   fast_dev_run: false
   gpus: 1
   precision: 16
@@ -144,5 +144,5 @@ trainer:
   limit_val_batches: 1.0
   limit_test_batches: 1.0
   resume_from_checkpoint: null
-  accumulate_grad_batches: 32
+  accumulate_grad_batches: 16
   overfit_batches: 0
diff --git a/training/conf/experiment/cnn_transformer_paragraphs_wp.yaml b/training/conf/experiment/cnn_transformer_paragraphs_wp.yaml
new file mode 100644
index 0000000..6d2bbdb
--- /dev/null
+++ b/training/conf/experiment/cnn_transformer_paragraphs_wp.yaml
@@ -0,0 +1,155 @@
+# @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
+
+
+epochs: &epochs 1000
+ignore_index: &ignore_index 1000
+num_classes: &num_classes 1006
+max_output_len: &max_output_len 451
+summary: [[1, 1, 576, 640], [1, 451]]
+
+criterion:
+  _target_: torch.nn.CrossEntropyLoss
+  ignore_index: *ignore_index
+    
+mapping:
+  _target_: text_recognizer.data.word_piece_mapping.WordPieceMapping
+  num_features: 1000
+  tokens: iamdb_1kwp_tokens_1000.txt
+  lexicon: iamdb_1kwp_lex_1000.txt
+  data_dir: null
+  use_words: false
+  prepend_wordsep: false
+  special_tokens: [ <s>, <e>, <p> ]
+  extra_symbols: [ "\n" ]
+
+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:
+  madgrad:
+    _target_: madgrad.MADGRAD
+    lr: 3.0e-4
+    momentum: 0.9
+    weight_decay: 0
+    eps: 1.0e-6
+
+    parameters: network
+
+lr_schedulers:
+  network:
+    _target_: torch.optim.lr_scheduler.OneCycleLR
+    max_lr: 3.0e-4
+    total_steps: null
+    epochs: *epochs
+    steps_per_epoch: 79
+    pct_start: 0.3
+    anneal_strategy: cos
+    cycle_momentum: true
+    base_momentum: 0.85
+    max_momentum: 0.95
+    div_factor: 25
+    final_div_factor: 1.0e4
+    three_phase: false
+    last_epoch: -1
+    verbose: false
+    # Non-class arguments
+    interval: step
+    monitor: val/loss
+
+datamodule:
+  _target_: text_recognizer.data.iam_extended_paragraphs.IAMExtendedParagraphs
+  batch_size: 4
+  num_workers: 12
+  train_fraction: 0.8
+  augment: true
+  pin_memory: true
+  word_pieces: true
+  resize: null
+
+network:
+  _target_: text_recognizer.networks.conv_transformer.ConvTransformer
+  input_dims: [1, 576, 640]
+  hidden_dim: &hidden_dim 128
+  encoder_dim: 1280
+  dropout_rate: 0.2
+  num_classes: *num_classes
+  pad_index: *ignore_index
+  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: *hidden_dim
+    depth: 3 
+    num_heads: 4
+    attn_fn: text_recognizer.networks.transformer.attention.Attention
+    attn_kwargs:
+      dim_head: 32
+      dropout_rate: 0.2
+    norm_fn: text_recognizer.networks.transformer.norm.ScaleNorm
+    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:
+      _target_: text_recognizer.networks.transformer.positional_encodings.rotary_embedding.RotaryEmbedding
+      dim: 32
+  pixel_pos_embedding:
+    _target_: text_recognizer.networks.transformer.positional_encodings.PositionalEncoding2D
+    hidden_dim: *hidden_dim 
+    max_h: 18
+    max_w: 20
+  token_pos_embedding:
+    _target_: text_recognizer.networks.transformer.positional_encodings.PositionalEncoding
+    hidden_dim: *hidden_dim 
+    dropout_rate: 0.2
+    max_len: *max_output_len
+
+model:
+  _target_: text_recognizer.models.transformer.TransformerLitModel
+  max_output_len: *max_output_len
+  start_token: <s>
+  end_token: <e>
+  pad_token: <p>
+
+trainer:
+  _target_: pytorch_lightning.Trainer
+  stochastic_weight_avg: true
+  auto_scale_batch_size: binsearch
+  auto_lr_find: false
+  gradient_clip_val: 0.0
+  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: 16
+  overfit_batches: 0