eazygrad.logsumexp¶

eazygrad.logsumexp(input: _Tensor, dim: int, keepdims: bool = False) → _Tensor[source]¶

Compute log(sum(exp(input))) in a numerically stable way.

Parameters:

input (_Tensor) – Input tensor.
dim (int) – Axis along which the reduction is performed.
keepdims (bool, default=False) – Whether to keep the reduced dimension in the output.

Returns:

_Tensor – Reduced tensor after applying the log-sum-exp operation.