Variable Selection Network block — layer

It receives four-dimensional vector as an input in the case of dynamic data (batch_size, timesteps, n_features, feature_dim)

Usage

layer_vsn(
  object,
  hidden_units,
  state_size,
  dropout_rate = NULL,
  use_context = FALSE,
  return_weights = FALSE,
  ...
)

Arguments

state_size: Dimensionality of the feature space, common across the model. The name comes from the original paper where they also refer to as $$d_model$$
return_weights: Return weights of the selection.

Value

A tensor of shapes:

dynamic data - (batch_size, timesteps, state_size)
static data - (batch_size, state_size)

Examples


# =========================================================================
#               THREE-DIMENSIONAL INPUT (STATIC FEATURES)
# =========================================================================

# input: (batch_size, n_features, state_size)

inp <- layer_input(c(10, 5))
out <- layer_vsn(hidden_units = 10, state_size = 5)(inp)
dim(out)
#> [1] NA  5

# =========================================================================
#               FOUR-DIMENSIONAL INPUT (DYNAMIC FEATURES)
# =========================================================================

# input: (batch_size, timesteps, n_features, state_size)

inp <- layer_input(c(28, 10, 5))
out <- layer_vsn(hidden_units = 10, state_size = 5)(inp)
dim(out)
#> [1] NA 28  5