log_softmax_op

→TensorImplPtr

int log_softmax_op(const int & a, int axis)

source edit

Python wrappers

Numerically stable log-softmax $y = x - \mathrm{logsumexp}(x, \text{axis})$ .

See Also

softmax_op : Probability-domain variant. LogSoftmaxBackward : Autograd node implementing the gradient rule.

Parameters

aTensorImplPtr

Input logits.

axisint

Axis along which to normalise. Negative values count from the end.

Returns

TensorImplPtr

Log-softmax output with the same shape and dtype as a.

Notes

Prefer log_softmax_op followed by nll_loss over log(softmax_op(...)) followed by nll_loss — the former is numerically stable for very confident logits, the latter is not.