Skip to content

Defaults

The top-level defaults section specifies type-global:

  1. Preprocessing
  2. Encoder
  3. Decoder
  4. Loss

Any configurations set in the defaults section apply to all features of that particular feature type. Any default preprocessing and encoder configurations will be applied to all input features of that feature type, while decoder and loss configurations will be applied to all output features of that feature type.

These parameters can be set for individual features through the input feature configuration or output feature configuration.

Note

Feature-specific configurations override global defaults: When a parameter is defined for a specific feature and also modified via defaults, the feature specific configuration overrides the value set in the defaults section for that particular parameter.

input_features:
  - 
    name: title
    type: text
    preprocessing:
        most_common: 10
  - 
    name: summary
    type: text
....
defaults:
    text:
        preprocessing:
            most_common: 100

In the example config above, the most_common preprocessing value for title will be set to 10 instead of taking on the default value of 100, while summary will now have its most_common preprocessing value set to 100.

Defining Defaults

The defaults section of Ludwig has the following structure:

defaults:
    <feature_type_1>:
        preprocessing:
            parameter_1: value
            ...
        encoder:
            parameter_1: value
            ...
        decoder:
            parameter_1: value
            ...
        loss:
            parameter_1: value
            ...
    <feature_type_2>:
        preprocessing:
            parameter_1: value
            ...
        encoder:
            parameter_1: value
            ...
        decoder:
            parameter_1: value
            ...
        loss:
            parameter_1: value
            ...
    ...

Each of the sections preprocessing, encoder, decoder and loss are optional so you can define one or more as you need.

Type-Global Preprocessing

Specify preprocessing policies that apply globally across all input features of a certain data type. For example:

defaults:
    category:
        preprocessing:
            missing_value_strategy: fill_with_const
            fill_value: <UNK>

The preprocessing parameters that each data type accepts can be found in datatype-specific documentation.

Note that different features with the same datatype may require different preprocessing. Type-global preprocessing works in tandem with feature-specific preprocessing configuration parameters, however, feature-specific configurations override the global settings.

For example, a document classification model may have two text input features, one for the title of the document and one for the body.

As the length of the title is much shorter than the length of the body, the parameter max_sequence_length should be set to 10 for the title and 2000 for the body, but we want both features to share the same vocabulary, with most_common: 10000.

The way to do this is by adding a preprocessing key inside the title input_feature dictionary and one in the body input feature dictionary containing the desired parameter and value.

input_features:
  -   
    name: title
    type: text
    preprocessing:
        max_sequence_length: 20
  -   
    name: body
    type: text
    preprocessing:
        max_sequence_length: 2000
defaults:
    text:
        preprocessing:
            most_common: 10000
Tokenizers

Sequence, text, and set features tokenize features as part of preprocessing. There are several tokenization options that can be specified:

Basic tokenizers
  • characters: splits every character of the input string in a separate token.
  • space: splits on space characters using the regex \s+.
  • space_punct: splits on space characters and punctuation using the regex \w+|[^\w\s].
  • underscore: splits on the underscore character _.
  • comma: splits on the comma character ,.
  • untokenized: treats the whole string as a single token.
  • stripped: treats the whole string as a single token after removing spaces at the beginning and at the end of the string.
  • ngram: this will create a vocab that consists of unigrams and bigrams.
spaCy tokenizers

The spaCy based tokenizers are functions that use the powerful tokenization and NLP preprocessing models provided the library. Several languages are available: English (code en), Italian (code it), Spanish (code es), German (code de), French (code fr), Portuguese (code pt), Dutch (code nl), Greek (code el), Chinese (code zh), Danish (code da), Dutch (code el), Japanese (code ja), Lithuanian (code lt), Norwegian (code nb), Polish (code pl), Romanian (code ro) and Multi (code xx, useful in case you have a dataset containing different languages).

For each language different functions are available:

  • tokenize: uses spaCy tokenizer,
  • tokenize_filter: uses spaCy tokenizer and filters out punctuation, numbers, stopwords and words shorter than 3 characters,
  • tokenize_remove_stopwords: uses spaCy tokenizer and filters out stopwords,
  • lemmatize: uses spaCy lemmatizer,
  • lemmatize_filter: uses spaCy lemmatizer and filters out punctuation, numbers, stopwords and words shorter than 3 characters,
  • lemmatize_remove_stopwords: uses spaCy lemmatize and filters out stopwords.

In order to use these options, you must first download the the spaCy model:

python -m spacy download <language_code>

and provide <language>_<function> as tokenizer. Example: english_tokenizer, italian_lemmatize_filter, multi_tokenize_filter and so on. More details on the models can be found in the spaCy documentation. Note that spaCy tokenizers are not TorchScript-compatible.

TorchText and HuggingFace tokenizers
  • sentencepiece: XLM-RoBERTa sentencepeice tokenizer.
  • clip: CLIP Tokenizer. Based on Byte-Level BPE.
  • gpt2bpe: GPT-2 BPE Tokenizer.
  • bert: BERT tokenizer based on the WordPiece algorithm.
  • hf_tokenizer: uses the Hugging Face AutoTokenizer which uses a pretrained_model_name_or_path parameter to decide which tokenizer to load.

Note: when specifying one of these tokenizers in the config, preprocessing parameters most_common and most_common_percentile won't have an effect. The full vocab of these tokenizers will be used.

Type-Global Encoder

Specify the encoder type and encoder related parameters across all input features of a certain data type. This encoder will be shared across all features of this particular feature type. For example:

defaults:
    text:
        encoder:
            type: stacked_cnn
            embedding_size: 128
            num_filters: 512

Note

The encoder type is a required parameter when defining a default encoder for a feature type or changing the default value for a parameter for the encoder, since the parameters are tied to specific encoders. Only one default encoder can be defined for all features of that particular type.

The encoder types and parameters that each data type accepts can be found in datatype-specific documentation.

Type-Global Decoder

Specify the decoder type and decoder related parameters across all output features of a certain data type. For example:

defaults:
    text:
        decoder:
            type: generator
            output_size: 128
            bias_initializer: he_normal

Note

The decoder type is a required parameter when defining a default decoder for a feature type or changing the default value for a parameter for the decoder, since the parameters are tied to specific decoders. Only one default decoder can be defined for all features of that particular type.

The decoder types and parameters that each data type accepts can be found in datatype-specific documentation.

Type-Global Loss

Specify the loss type and loss related parameters across all output features of a certain data type. For example:

defaults:
    text:
        loss:
            type: softmax_cross_entropy
            confidence_penalty: 0.1

The loss types and parameters that each data type accepts can be found in datatype-specific documentation.

Defaults Example

Following is a full example of a Ludwig configuration with type-global defaults.

config.yaml
input_features:
  - 
    name: title
    type: text
  - 
    name: body
    type: text
  - 
    name: num_characters
    type: number
    preprocessing:
        normalization: zscore
combiner:
  type: concat
  num_fc_layers: 1
output_features:
  - 
    name: spam
    type: category
defaults:
    text:
        preprocessing:
            most_common: 10000
        encoder:
            type: rnn
            cell_type: lstm
            num_layers: 2
trainer:
  learning_rate: 0.001
  optimizer:
    type: adam

Example CLI command:

ludwig train --dataset spam.csv --config_str "{input_features: [{name: title, type: text}, {name: body, type: text}, {name: num_characters, type: number, preprocessing: {normalization: zscore}}], output_features: [{name: spam, type: category}], combiner: {type: concat, num_fc_layers: 1}, defaults: {text: {preprocessing: {word_vocab_size: 10000}, encoder: {type: rnn, cell_type: lstm, num_layers: 2}}}, trainer: {learning_rate: 0.001, optimizer: {type: adam}}"