summaryrefslogtreecommitdiff
path: root/training/conf/network/conv_transformer.yaml
blob: f618ba175156fde196542902bc2852250289583a (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
_target_: text_recognizer.networks.ConvTransformer
input_dims: [1, 1, 576, 640]
hidden_dim: &hidden_dim 144
num_classes: 58
pad_index: 3
encoder:
  _target_: text_recognizer.networks.EfficientNet
  arch: b0
  stochastic_dropout_rate: 0.2
  bn_momentum: 0.99
  bn_eps: 1.0e-3
  depth: 5
  out_channels: *hidden_dim
decoder:
  _target_: text_recognizer.networks.transformer.Decoder
  depth: 6
  block:
    _target_: text_recognizer.networks.transformer.DecoderBlock
    self_attn:
      _target_: text_recognizer.networks.transformer.Attention
      dim: *hidden_dim
      num_heads: 12
      dim_head: 64
      dropout_rate: &dropout_rate 0.2
      causal: true
      rotary_embedding:
        _target_: text_recognizer.networks.transformer.RotaryEmbedding
        dim: 64
    cross_attn:
      _target_: text_recognizer.networks.transformer.Attention
      dim: *hidden_dim
      num_heads: 8
      dim_head: 64
      dropout_rate: *dropout_rate
      causal: false
    norm:
      _target_: text_recognizer.networks.transformer.RMSNorm
      dim: *hidden_dim
    ff:
      _target_: text_recognizer.networks.transformer.FeedForward
      dim: *hidden_dim
      dim_out: null
      expansion_factor: 2
      glu: true
      dropout_rate: *dropout_rate
pixel_embedding:
  _target_: text_recognizer.networks.transformer.AxialPositionalEmbedding
  dim: *hidden_dim
  shape: [18, 79]