MovingSumWithSigmoid

Description

Performs MovingSum on the input alpha, and Sigmoid computation on the input energy. The product of the two results is the output.

Input

Three inputs:

alpha: tensor of type float16 or float, with shape [bst], 1D.

energy: tensor of type float16 or float, with shape [bst], 1D.

offset: tensor of type int32, with shape [2 * batch_size], 1D. The value consists of the beam_num ([b_1,b_2, ...,b_batch_size]) and frame_num ([f_1,f_2, ...,f_batch_size]) arrays. b_1 x f_1 + b_2 x f_2 +... (constraint met) + b_batch_size x f_batch_size = bst.

Output

One output:

y: tensor of type float16 or float, with shape [bs, bt], 2D. bs = sum (beam_num), and bt = sum (frame_num).

Attribute

ksize: (mandatory) int.

Constraints

This operator is customized for specific user scenarios and is not recommended in irrelevant scenarios.

This operator supports only the Atlas inference products .

ONNX Opset Support

Opset v8/v9/v10/v11/v12/v13/v14/v15/v16