def prune_fc_layer_with_craig(layer: nn.Linear,
prune_percent_per_layer: float,
similarity_metric: Union[Text, Dict] = "",
prune_type: Text = "craig",
**kwargs) -> Tuple[List[int], List[float]]:
# Get CRAIG subset.
subset_nodes: List
subset_weights: List
subset_nodes, subset_weights = get_layer_craig_subset(
layer=layer,
original_num_nodes=layer.out_features,
prune_percent_per_layer=prune_percent_per_layer,
similarity_metric=similarity_metric,
prune_type=prune_type,
**kwargs)
# Remove nodes+weights+biases, and adjust weights.
num_nodes: int = len(subset_nodes)
# Prune current layer.
# Multiply weights (and biases?) by subset_weights.
subset_weights_tensor = torch.tensor(subset_weights)
layer.weight = nn.Parameter(layer.weight[subset_nodes] *
subset_weights_tensor.reshape((num_nodes, 1)))
if layer.bias is not None:
layer.bias = nn.Parameter(layer.bias[subset_nodes] *
subset_weights_tensor)
layer.out_features = num_nodes
return subset_nodes, subset_weights
def _create_mean_predictor(self):
"""Creates a new predictor using the mean parameters across the predictor heads."""
weights = []
biases = []
# Collect weights/biases from each predictor head and create tensors
for i in range(self.n_participants):
weights.append(self.predictor_heads[i].weight)
biases.append(self.predictor_heads[i].bias)
weights = torch.stack(weights)
biases = torch.stack(biases)
# Create new linear predictor and set weights/biases to means
predictor_heads_mean = Linear(self.hidden_dim, self.output_dim)
predictor_heads_mean.weight = Parameter(weights.mean(0))
predictor_heads_mean.bias = Parameter(biases.mean(0))
return predictor_heads_mean
def _create_sampled_predictor(self):
"""Creates a new predictor using parameters sampled from the prior for random effects."""
# Sample parameters and extract weight and bias parameters from flattened list
sampled_params = MultivariateNormal(self.mean, self.cov).sample()
flattened_mlp_params = sampled_params[
: ((self.hidden_dim + 1) * self.output_dim)
]
mlp_params = flattened_mlp_params.reshape(
(self.output_dim, self.hidden_dim + 1)
)
weight, bias = mlp_params[:, :-1], mlp_params[:, -1]
# Create new linear predictor and set weights/biases to sampled values
predictor_head_sampled = Linear(self.hidden_dim, self.output_dim)
predictor_head_sampled.weight = Parameter(weight)
predictor_head_sampled.bias = Parameter(bias)
return predictor_head_sampled
