Add experiment configs

1 files changed, 54 insertions, 36 deletions

diff --git a/training/conf/experiment/cnn_htr_char_lines.yaml b/training/conf/experiment/cnn_htr_char_lines.yaml
index 0f28ff9..0d62a73 100644
--- a/training/conf/experiment/cnn_htr_char_lines.yaml
+++ b/training/conf/experiment/cnn_htr_char_lines.yaml
@@ -10,28 +10,26 @@ defaults:
   - override /optimizers: null
 
 
+epochs: &epochs 200
+ignore_index: &ignore_index 3
+num_classes: &num_classes 58
+max_output_len: &max_output_len 89
+
 criterion:
   _target_: text_recognizer.criterions.label_smoothing.LabelSmoothingLoss
   smoothing: 0.1 
-  ignore_index: 1000
+  ignore_index: *ignore_index
+  # _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> ]
-  # _target_: text_recognizer.data.emnist_mapping.EmnistMapping
-  # extra_symbols: [ "\n" ]
+  _target_: text_recognizer.data.emnist_mapping.EmnistMapping
 
 callbacks:
   stochastic_weight_averaging:
     _target_: pytorch_lightning.callbacks.StochasticWeightAveraging
-    swa_epoch_start: 0.8
-    swa_lrs: 0.05
+    swa_epoch_start: 0.75
+    swa_lrs: 5.0e-5
     annealing_epochs: 10
     annealing_strategy: cos
     device: null
@@ -39,7 +37,7 @@ callbacks:
 optimizers:
   madgrad:
     _target_: madgrad.MADGRAD
-    lr: 1.0e-4
+    lr: 3.0e-4
     momentum: 0.9
     weight_decay: 0
     eps: 1.0e-6
@@ -48,34 +46,42 @@ optimizers:
 
 lr_schedulers:
   network:
-    _target_: torch.optim.lr_scheduler.ReduceLROnPlateau
-    mode: min
-    factor: 0.1
-    patience: 10
-    threshold: 1.0e-4
-    threshold_mode: rel
-    cooldown: 0
-    min_lr: 1.0e-7
-    eps: 1.0e-8
-    interval: epoch
+    _target_: torch.optim.lr_scheduler.OneCycleLR
+    max_lr: 3.0e-4
+    total_steps: null
+    epochs: *epochs
+    steps_per_epoch: 90
+    pct_start: 0.1
+    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_lines.IAMLines
-  batch_size: 16
+  batch_size: 32
   num_workers: 12
   train_fraction: 0.8
   augment: true
-  pin_memory: false
+  pin_memory: true
+  word_pieces: false
 
 network:
   _target_: text_recognizer.networks.conv_transformer.ConvTransformer
   input_dims: [1, 56, 1024]
-  hidden_dim: 128
+  hidden_dim: &hidden_dim 128
   encoder_dim: 1280
   dropout_rate: 0.2
-  num_classes: 1006
-  pad_index: 1000
+  num_classes: *num_classes
+  pad_index: *ignore_index
   encoder:
     _target_: text_recognizer.networks.encoders.efficientnet.EfficientNet
     arch: b0
@@ -85,14 +91,14 @@ network:
     bn_eps: 1.0e-3
   decoder:
     _target_: text_recognizer.networks.transformer.Decoder
-    dim: 128
+    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: torch.nn.LayerNorm
+    norm_fn: text_recognizer.networks.transformer.norm.ScaleNorm
     ff_fn: text_recognizer.networks.transformer.mlp.FeedForward
     ff_kwargs:
       dim_out: null
@@ -101,11 +107,23 @@ network:
       dropout_rate: 0.2
     cross_attend: true
     pre_norm: true
-    rotary_emb: null
+    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: 1
+    max_w: 32
+  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: 89
+  max_output_len: *max_output_len
   start_token: <s>
   end_token: <e>
   pad_token: <p>
@@ -115,11 +133,11 @@ trainer:
   stochastic_weight_avg: true
   auto_scale_batch_size: binsearch
   auto_lr_find: false
-  gradient_clip_val: 0
+  gradient_clip_val: 0.5
   fast_dev_run: false
   gpus: 1
   precision: 16
-  max_epochs: 1024
+  max_epochs: *epochs
   terminate_on_nan: true
   weights_summary: null
   limit_train_batches: 1.0 
@@ -127,6 +145,6 @@ trainer:
   limit_test_batches: 1.0
   resume_from_checkpoint: null
   accumulate_grad_batches: 4
-  overfit_batches: 0.0
+  overfit_batches: 0
 
 summary: [[1, 1, 56, 1024], [1, 89]]