Skip to contents

This is an implementation of grouped-query attention introduced by Ainslie et al., 2023. Here num_key_value_heads denotes number of groups, setting num_key_value_heads to 1 is equivalent to multi-query attention, and when num_key_value_heads is equal to num_query_heads it is equivalent to multi-head attention.

This layer first projects query, key, and value tensors. Then, key and value are repeated to match the number of heads of query.

Then, the query is scaled and dot-producted with key tensors. These are softmaxed to obtain attention probabilities. The value tensors are then interpolated by these probabilities and concatenated back to a single tensor.

Usage

layer_group_query_attention(
  object,
  head_dim,
  num_query_heads,
  num_key_value_heads,
  dropout = 0,
  use_bias = TRUE,
  kernel_initializer = "glorot_uniform",
  bias_initializer = "zeros",
  kernel_regularizer = NULL,
  bias_regularizer = NULL,
  activity_regularizer = NULL,
  kernel_constraint = NULL,
  bias_constraint = NULL,
  ...
)

Arguments

object

Object to compose the layer with. A tensor, array, or sequential model.

head_dim

Size of each attention head.

num_query_heads

Number of query attention heads.

num_key_value_heads

Number of key and value attention heads.

dropout

Dropout probability.

use_bias

Boolean, whether the dense layers use bias vectors/matrices.

kernel_initializer

Initializer for dense layer kernels.

bias_initializer

Initializer for dense layer biases.

kernel_regularizer

Regularizer for dense layer kernels.

bias_regularizer

Regularizer for dense layer biases.

activity_regularizer

Regularizer for dense layer activity.

kernel_constraint

Constraint for dense layer kernels.

bias_constraint

Constraint for dense layer kernels.

...

For forward/backward compatability.

Value

attention_output: Result of the computation, of shape (batch_dim, target_seq_len, feature_dim), where target_seq_len

is for target sequence length and feature_dim is the query input last dim. attention_scores: (Optional) attention coefficients of shape (batch_dim, num_query_heads, target_seq_len, source_seq_len).

Call Arguments

  • query: Query tensor of shape (batch_dim, target_seq_len, feature_dim), where batch_dim is batch size, target_seq_len is the length of target sequence, and feature_dim is dimension of feature.

  • value: Value tensor of shape (batch_dim, source_seq_len, feature_dim), where batch_dim is batch size, source_seq_len is the length of source sequence, and feature_dim is dimension of feature.

  • key: Optional key tensor of shape (batch_dim, source_seq_len, feature_dim). If not given, will use value for both key and value, which is most common case.

  • attention_mask: A boolean mask of shape (batch_dim, target_seq_len, source_seq_len), that prevents attention to certain positions. The boolean mask specifies which query elements can attend to which key elements, where 1 indicates attention and 0 indicates no attention. Broadcasting can happen for the missing batch dimensions and the head dimension.

  • return_attention_scores: A boolean to indicate whether the output should be (attention_output, attention_scores) if TRUE, or attention_output if FALSE. Defaults to FALSE.

  • training: Python boolean indicating whether the layer should behave in training mode (adding dropout) or in inference mode (no dropout). Will go with either using the training mode of the parent layer/model or FALSE (inference) if there is no parent layer.

  • use_causal_mask: A boolean to indicate whether to apply a causal mask to prevent tokens from attending to future tokens (e.g., used in a decoder Transformer).

See also

Other attention layers:
layer_additive_attention()
layer_attention()
layer_multi_head_attention()

Other layers:
Layer()
layer_activation()
layer_activation_elu()
layer_activation_leaky_relu()
layer_activation_parametric_relu()
layer_activation_relu()
layer_activation_softmax()
layer_activity_regularization()
layer_add()
layer_additive_attention()
layer_alpha_dropout()
layer_attention()
layer_average()
layer_average_pooling_1d()
layer_average_pooling_2d()
layer_average_pooling_3d()
layer_batch_normalization()
layer_bidirectional()
layer_category_encoding()
layer_center_crop()
layer_concatenate()
layer_conv_1d()
layer_conv_1d_transpose()
layer_conv_2d()
layer_conv_2d_transpose()
layer_conv_3d()
layer_conv_3d_transpose()
layer_conv_lstm_1d()
layer_conv_lstm_2d()
layer_conv_lstm_3d()
layer_cropping_1d()
layer_cropping_2d()
layer_cropping_3d()
layer_dense()
layer_depthwise_conv_1d()
layer_depthwise_conv_2d()
layer_discretization()
layer_dot()
layer_dropout()
layer_einsum_dense()
layer_embedding()
layer_feature_space()
layer_flatten()
layer_gaussian_dropout()
layer_gaussian_noise()
layer_global_average_pooling_1d()
layer_global_average_pooling_2d()
layer_global_average_pooling_3d()
layer_global_max_pooling_1d()
layer_global_max_pooling_2d()
layer_global_max_pooling_3d()
layer_group_normalization()
layer_gru()
layer_hashed_crossing()
layer_hashing()
layer_identity()
layer_integer_lookup()
layer_lambda()
layer_layer_normalization()
layer_lstm()
layer_masking()
layer_max_pooling_1d()
layer_max_pooling_2d()
layer_max_pooling_3d()
layer_maximum()
layer_minimum()
layer_multi_head_attention()
layer_multiply()
layer_normalization()
layer_permute()
layer_random_brightness()
layer_random_contrast()
layer_random_crop()
layer_random_flip()
layer_random_rotation()
layer_random_translation()
layer_random_zoom()
layer_repeat_vector()
layer_rescaling()
layer_reshape()
layer_resizing()
layer_rnn()
layer_separable_conv_1d()
layer_separable_conv_2d()
layer_simple_rnn()
layer_spatial_dropout_1d()
layer_spatial_dropout_2d()
layer_spatial_dropout_3d()
layer_spectral_normalization()
layer_string_lookup()
layer_subtract()
layer_text_vectorization()
layer_tfsm()
layer_time_distributed()
layer_torch_module_wrapper()
layer_unit_normalization()
layer_upsampling_1d()
layer_upsampling_2d()
layer_upsampling_3d()
layer_zero_padding_1d()
layer_zero_padding_2d()
layer_zero_padding_3d()
rnn_cell_gru()
rnn_cell_lstm()
rnn_cell_simple()
rnn_cells_stack()