Update configs

1 files changed, 7 insertions, 92 deletions

diff --git a/training/conf/experiment/conv_transformer_lines.yaml b/training/conf/experiment/conv_transformer_lines.yaml
index 55dc896..8404cd1 100644
--- a/training/conf/experiment/conv_transformer_lines.yaml
+++ b/training/conf/experiment/conv_transformer_lines.yaml
@@ -1,17 +1,16 @@
----
 # @package _global_
 
 defaults:
-  - override /mapping: null
+  - override /mapping: characters
   - override /criterion: cross_entropy
   - override /callbacks: htr
   - override /datamodule: iam_lines
-  - override /network: null
-  - override /model: null
+  - override /network: conv_transformer
+  - override /model: lit_transformer
   - override /lr_schedulers: null
   - override /optimizers: null
 
-epochs: &epochs 200
+epochs: &epochs 512
 ignore_index: &ignore_index 3
 num_classes: &num_classes 57
 max_output_len: &max_output_len 89
@@ -21,10 +20,6 @@ criterion:
   ignore_index: *ignore_index
   # label_smoothing: 0.1
 
-mapping: &mapping
-  mapping:
-    _target_: text_recognizer.data.mappings.emnist.EmnistMapping
-
 callbacks:
   stochastic_weight_averaging:
     _target_: pytorch_lightning.callbacks.StochasticWeightAveraging
@@ -60,99 +55,19 @@ lr_schedulers:
 
 datamodule:
   batch_size: 16
-  num_workers: 12
-  train_fraction: 0.9
-  pin_memory: true
-  << : *mapping
-
-encoder: &encoder
-  _target_: text_recognizer.networks.efficientnet.efficientnet.EfficientNet
-  arch: b0
-  stochastic_dropout_rate: 0.2
-  bn_momentum: 0.99
-  bn_eps: 1.0e-3
-  depth: 5
-
-rotary_embedding: &rotary_embedding
-  rotary_embedding:
-    _target_: >
-      text_recognizer.networks.transformer.embeddings.rotary.RotaryEmbedding
-    dim: 64
-
-attn: &attn
-  dim: &hidden_dim 512
-  num_heads: 4
-  dim_head: 64
-  dropout_rate: &dropout_rate 0.4
-
-decoder: &decoder
-  _target_: text_recognizer.networks.transformer.decoder.Decoder
-  depth: 6
-  has_pos_emb: true
-  block:
-    _target_: text_recognizer.networks.transformer.decoder.DecoderBlock
-    self_attn:
-      _target_: text_recognizer.networks.transformer.attention.Attention
-      <<: *attn
-      causal: true
-      <<: *rotary_embedding
-    cross_attn:
-      _target_: text_recognizer.networks.transformer.attention.Attention
-      <<: *attn
-      causal: false
-    norm:
-      _target_: text_recognizer.networks.transformer.norm.RMSNorm
-      dim: *hidden_dim
-    ff:
-      _target_: text_recognizer.networks.transformer.mlp.FeedForward
-      dim: *hidden_dim
-      dim_out: null
-      expansion_factor: 2
-      glu: true
-      dropout_rate: *dropout_rate
-
-pixel_pos_embedding: &pixel_pos_embedding
-  _target_: >
-    text_recognizer.networks.transformer.embeddings.axial.AxialPositionalEmbedding
-  dim: *hidden_dim
-  shape: &shape [3, 64]
 
 network:
-  _target_: text_recognizer.networks.conv_transformer.ConvTransformer
   input_dims: [1, 1, 56, 1024]
-  hidden_dim: *hidden_dim
   num_classes: *num_classes
   pad_index: *ignore_index
-  encoder:
-    <<: *encoder
   decoder:
-    <<: *decoder
-  pixel_pos_embedding:
-    <<: *pixel_pos_embedding
+    depth: 10
+  pixel_embedding:
+    shape: [7, 128]
 
 model:
-  _target_: text_recognizer.models.transformer.LitTransformer
-  <<: *mapping
   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: 0.5
-  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