summaryrefslogtreecommitdiff
path: root/training/conf/experiment/vit_lines.yaml
blob: f57eead609671bff3100bcc3607f0ac1940732fc (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
# @package _global_

defaults:
  - override /criterion: cross_entropy
  - override /callbacks: htr
  - override /datamodule: iam_lines
  - override /network: null
  - override /model: lit_transformer
  - override /lr_scheduler: null
  - override /optimizer: null

tags: [lines, vit]
epochs: &epochs 256
ignore_index: &ignore_index 3
# summary: [[1, 1, 56, 1024], [1, 89]]

logger:
  wandb:
    tags: ${tags}

criterion:
  ignore_index: *ignore_index
  # label_smoothing: 0.05


decoder:
  max_output_len: 89

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_: adan_pytorch.Adan
  lr: 3.0e-4
  betas: [0.02, 0.08, 0.01]
  weight_decay: 0.02

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: 16
  train_fraction: 0.95

network:
  _target_: text_recognizer.network.vit.VisionTransformer
  image_height: 56
  image_width: 1024
  patch_height: 28
  patch_width: 32
  dim: &dim 1024
  num_classes: &num_classes 58
  encoder:
    _target_: text_recognizer.network.transformer.encoder.Encoder
    dim: *dim
    inner_dim: 2048
    heads: 16
    dim_head: 64
    depth: 4
    dropout_rate: 0.0
  decoder:
    _target_: text_recognizer.network.transformer.decoder.Decoder
    dim: *dim
    inner_dim: 2048
    heads: 16
    dim_head: 64
    depth: 4
    dropout_rate: 0.0
  token_embedding:
    _target_: "text_recognizer.network.transformer.embedding.token.\
      TokenEmbedding"
    num_tokens: *num_classes
    dim: *dim
    use_l2: true
  pos_embedding:
    _target_: "text_recognizer.network.transformer.embedding.absolute.\
      AbsolutePositionalEmbedding"
    dim: *dim
    max_length: 89
    use_l2: true
  tie_embeddings: true
  pad_index: 3

model:
  max_output_len: 89

trainer:
  fast_dev_run: false
  gradient_clip_val: 1.0
  max_epochs: *epochs
  accumulate_grad_batches: 4
  limit_train_batches: 1.0
  limit_val_batches: 1.0
  limit_test_batches: 1.0