Skip to content

Model Configuration

The model_config.yaml defines the architecture of the OCR model used in fast-plate-ocr. This configuration allows you to customize key components of the model, i.e., convolutional tokenizers, patching strategies, attention settings, etc. (all without modifying code).

All model configurations are validated using Pydantic, ensuring that every field, layer, and parameter is checked when building the model from the config.

Currently, the supported architectures are:

  • CCT (Compact Convolutional Transformer 1)

Supported Architectures

Compact Convolutional Transformer (CCT)

Inspired by the CCT architecture, this model structure:

  1. Uses a convolutional tokenizer to extract patch representations from the input image.
  2. Processes the resulting sequence with a Transformer Encoder.

Config Example

cct_model_config.yaml
 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
model: cct

rescaling:  # (1)!
  scale: 0.00392156862745098
  offset: 0.0

tokenizer:
  blocks:  # (2)!
    - { layer: Conv2D, filters: 32, kernel_size: 3, strides: 1 }  # (3)!
    - { layer: MaxBlurPooling2D, pool_size: 2, filter_size: 3 }
    - { layer: Conv2D, filters: 48, kernel_size: 3, strides: 1 }
    - { layer: Conv2D, filters: 64, kernel_size: 3, strides: 1 }
    - { layer: MaxBlurPooling2D, pool_size: 2, filter_size: 3 }
    - { layer: Conv2D, filters: 80, kernel_size: 3, strides: 1 }
    - { layer: Conv2D, filters: 96, kernel_size: 3, strides: 1 }

  positional_emb: true
  patch_size: 2
  patch_mlp:
    layer: MLP
    hidden_units: [64]
    activation: gelu
    dropout_rate: 0.05

transformer_encoder:
  layers: 4
  heads: 1
  projection_dim: 64
  units: [64, 64]
  activation: gelu
  stochastic_depth: 0.05
  attention_dropout: 0.05
  mlp_dropout: 0.1
  head_mlp_dropout: 0.05
  token_reducer_heads: 4
  normalization: dyt  # (4)!
  1. This scales values between range [0, 1]
  2. Supports a wide variety of layer types, such as Conv2D, MaxPooling2D, DepthwiseConv2D, SqueezeExcite, etc. See Model Config for all available option.
  3. Each layer supports the full set of corresponding Keras parameters. For example, Conv2D accepts filters, kernel_size, strides, etc.
  4. See Transformers without Normalization 2
Note on plate/model configs

The plate config is used throughout both inference and training scripts. In contrast, the model config (shown above) is only used for training, as it defines the architecture to be built.

Building Custom Tokenizers with Any Keras Layer

You can define your own tokenizer stacks by composing any supported layer like Conv2D, DepthwiseConv2D, SqueezeExcite, and many more directly in YAML, without writing any code.

Each layer accepts all its typical Keras parameters, and the model schema is validated with Pydantic, so typos or misconfigured fields are caught immediately.

Here's an example with a more diverse set of layers:

custom_model_config.yaml
tokenizer:
  blocks:
    - { layer: Conv2D, filters: 64, kernel_size: 3, activation: relu }
    - { layer: SqueezeExcite, ratio: 0.5 }
    - { layer: DepthwiseConv2D, kernel_size: 3, strides: 1 }
    - { layer: BatchNormalization }
    - { layer: MaxBlurPooling2D, pool_size: 2, filter_size: 3 }
    - { layer: Conv2D, filters: 128, kernel_size: 3 }
    - { layer: CoordConv2D, filters: 96, kernel_size: 3, with_r: true }
Tip

Each layer: value corresponds to a class in the Model Schema, check it out to see all the supported layers and options!

  1. Escaping the Big Data Paradigm with Compact Transformers. ↩

  2. Transformers without Normalization. ↩