wangwei created SINGA-316:
-----------------------------
Summary: Add SigmoidCrossEntropy
Key: SINGA-316
URL: https://issues.apache.org/jira/browse/SINGA-316
Project: Singa
Issue Type: New Feature
Reporter: wangwei
Assignee: wangwei
SigmoidCrossEntropy is a loss function for samples with multiple labels.
For each sample, denote the final layer's output as x (a vector of length L),
and y as the ground truth label vector (a binary vector of length L)
where L is the label set size.
SigmoidCrossEntropy computes
p_i = sigmoid(x_i)
loss_i = -y_i log p_i -(1-y_i)log (1-p_i)
A threshold t_i can be tuned to decide the label assignment, e.g. y_i = 1 if
p_i > t_i.
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
