one_hot_op

→TensorImplPtr

int one_hot_op(const int & input, int num_classes, Dtype out_dtype)

source edit

Python wrappers

pylucid.nn.functional.one_hotfunction

Encode integer indices into one-hot float vectors.

Returns a tensor of shape (*input.shape(), num_classes) where the last axis is the one-hot indicator

y_{\ldots, c} = \mathbf{1}\bigl[c = \text{input}_{\ldots}\bigr], \quad c \in \{0, 1, \ldots, \text{num\_classes} - 1\}.

The output dtype is out_dtype (typically a floating-point type so the result can feed directly into a loss function). Indices outside the $[0, \text{num\_classes})$ range produce all-zero rows. No autograd node is attached — this op is inference-only.

Parameters

inputTensorImplPtr

Integer tensor of arbitrary shape.

num_classesint

Length of the one-hot dimension.

out_dtypeDtype

Element type of the output (e.g. Dtype::Float32).

Returns

TensorImplPtr

One-hot tensor of shape (*input.shape(), num_classes) and dtype out_dtype.