Skip to main content

ClikaRT::ops::cross_entropy

function

cross_entropy()​

Tensor cross_entropy(
    Tensor input,
    Tensor target,
    OptionalTensor weight = {},
    std::optional<int64_t> ignore_index = std::nullopt,
    Reduction reduction = Reduction::Mean
)

Cross-entropy loss over class LOGITS; the class dim is LAST.

Equivalent to log_softmax over the class axis followed by negative-log-likelihood:

ℓ=−log⁡softmax(x)y\ell = -\log\mathrm{softmax}(x)_{y}

Parameters

  • input: logits [.., C] (channels-last class axis).
  • target: class indices [..] (integer dtype).
  • weight: optional per-class weight [C].
  • ignore_index: optional target value whose positions contribute no loss; absent = no masking.
  • reduction: None / Mean / Sum; default Mean.

Returns: the loss (0-D under Mean/Sum; [..] under None).

Throws

  • ClikaRT::Error: (INVALID_ARGUMENT) when target's shape is not input's minus the class axis, or a target index is out of range.
auto loss = ClikaRT::ops::cross_entropy(logits, labels);   // 0-D mean

Declared in ClikaRT/compute/ops.h, line 2940