Keras 3 API documentation / KerasNLP / Models / DebertaV3 / DebertaV3Backbone model

DebertaV3Backbone model

[source]

DebertaV3Backbone class

keras_nlp.models.DebertaV3Backbone(
    vocabulary_size,
    num_layers,
    num_heads,
    hidden_dim,
    intermediate_dim,
    dropout=0.1,
    max_sequence_length=512,
    bucket_size=256,
    dtype=None,
    **kwargs
)

DeBERTa encoder network.

This network implements a bi-directional Transformer-based encoder as described in "DeBERTaV3: Improving DeBERTa using ELECTRA-Style Pre-Training with Gradient-Disentangled Embedding Sharing". It includes the embedding lookups and transformer layers, but does not include the enhanced masked decoding head used during pretraining.

The default constructor gives a fully customizable, randomly initialized DeBERTa encoder with any number of layers, heads, and embedding dimensions. To load preset architectures and weights, use the from_preset constructor.

Note: DebertaV3Backbone has a performance issue on TPUs, and we recommend other models for TPU training and inference.

Disclaimer: Pre-trained models are provided on an "as is" basis, without warranties or conditions of any kind. The underlying model is provided by a third party and subject to a separate license, available here.

Arguments

  • vocabulary_size: int. The size of the token vocabulary.
  • num_layers: int. The number of transformer layers.
  • num_heads: int. The number of attention heads for each transformer. The hidden size must be divisible by the number of attention heads.
  • hidden_dim: int. The size of the transformer encoding layer.
  • intermediate_dim: int. The output dimension of the first Dense layer in a two-layer feedforward network for each transformer.
  • dropout: float. Dropout probability for the DeBERTa model.
  • max_sequence_length: int. The maximum sequence length this encoder can consume. The sequence length of the input must be less than max_sequence_length.
  • bucket_size: int. The size of the relative position buckets. Generally equal to max_sequence_length // 2.
  • dtype: string or keras.mixed_precision.DTypePolicy. The dtype to use for model computations and weights. Note that some computations, such as softmax and layer normalization, will always be done at float32 precision regardless of dtype.

Example

input_data = {
    "token_ids": np.ones(shape=(1, 12), dtype="int32"),
    "padding_mask": np.array([[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0]]),
}

# Pretrained DeBERTa encoder.
model = keras_nlp.models.DebertaV3Backbone.from_preset(
    "deberta_v3_base_en",
)
model(input_data)

# Randomly initialized DeBERTa encoder with custom config
model = keras_nlp.models.DebertaV3Backbone(
    vocabulary_size=128100,
    num_layers=12,
    num_heads=6,
    hidden_dim=384,
    intermediate_dim=1536,
    max_sequence_length=512,
    bucket_size=256,
)
# Call the model on the input data.
model(input_data)

[source]

from_preset method

DebertaV3Backbone.from_preset()

Instantiate DebertaV3Backbone model from preset architecture and weights.

Arguments

  • preset: string. Must be one of "deberta_v3_extra_small_en", "deberta_v3_small_en", "deberta_v3_base_en", "deberta_v3_large_en", "deberta_v3_base_multi".
  • load_weights: Whether to load pre-trained weights into model. Defaults to True.

Examples

# Load architecture and weights from preset
model = keras_nlp.models.DebertaV3Backbone.from_preset(
    "deberta_v3_extra_small_en"
)

# Load randomly initialized model from preset architecture
model = keras_nlp.models.DebertaV3Backbone.from_preset(
    "deberta_v3_extra_small_en",
    load_weights=False
)
Preset name Parameters Description
deberta_v3_extra_small_en 70.68M 12-layer DeBERTaV3 model where case is maintained. Trained on English Wikipedia, BookCorpus and OpenWebText.
deberta_v3_small_en 141.30M 6-layer DeBERTaV3 model where case is maintained. Trained on English Wikipedia, BookCorpus and OpenWebText.
deberta_v3_base_en 183.83M 12-layer DeBERTaV3 model where case is maintained. Trained on English Wikipedia, BookCorpus and OpenWebText.
deberta_v3_large_en 434.01M 24-layer DeBERTaV3 model where case is maintained. Trained on English Wikipedia, BookCorpus and OpenWebText.
deberta_v3_base_multi 278.22M 12-layer DeBERTaV3 model where case is maintained. Trained on the 2.5TB multilingual CC100 dataset.

token_embedding property

keras_nlp.models.DebertaV3Backbone.token_embedding

A keras.layers.Embedding instance for embedding token ids.

This layer embeds integer token ids to the hidden dim of the model.