FilesLast commitLast update
15 days ago
22 days ago
22 days ago
22 days ago
22 days ago
1 month ago
22 days ago
18 days ago
README

Relu6D

English | 简体中文

Applicable Products

Product Support
Ascend 950PR&950DT products √
Atlas A3 products √
Atlas A2 products √
Atlas 200I/500 A2 inference products √
Atlas inference products √
Atlas training products √

Function

  • Operator function: Relu6D is the ReLU6 activation function with a scaling coefficient. It elementwise truncates the input tensor from below first (negative values become 0) and then truncates it from above (limits it to 6*scale). When scale ≥ 0, the output value range is [0, 6*scale]; when scale < 0, the output always equals 6*scale. When scale = 1.0 (the default), it is equivalent to standard ReLU6. It suits activation layers of lightweight vision networks such as the MobileNet series, limiting the activation value range and improving numerical stability in low-precision quantization scenarios.
  • Formula:

Relu6D(x)=min⁡(max⁡(x, 0), 6⋅scale)Relu6D(x) = \min(\max(x,\ 0),\ 6 \cdot scale)

In the formula, xx is the input tensor, scalescale is the threshold scaling coefficient (default 1.01.0), and yy is the output tensor. When scale ≥ 0, the value range is [0, 6⋅scale][0,\ 6 \cdot scale]; when scale < 0, the output is always 6⋅scale6 \cdot scale.

Parameters

Parameter name Input/Output/Attribute Description Data type Data format
x Input The input feature tensor, corresponding to x in the formula. FLOAT16, FLOAT, INT32, BFLOAT16 ND
y Output The output tensor, corresponding to y in the formula, with the same shape and dtype as input x. FLOAT16, FLOAT, INT32, BFLOAT16 ND
scale Optional attribute The threshold scaling coefficient, corresponding to scale in the formula; the upper threshold = 6*scale. Default value: 1.0 (corresponding to standard ReLU6). FLOAT -

Constraints

  • The shapes of input x and output y must be the same.
  • Up to 8-dimensional tensors are supported.
  • Dynamic shapes and dynamic ranks are supported.
  • Empty tensors return successfully directly: tiling sets blockNum to 0 and allocates no data processing resources, and both Kernel Init and Process return early without initiating GM reads or writes.
  • Invalid range of the total element count: the total number of flattened elements of x (the shape product) overflows int64.
  • scale constraints: the upper threshold is 6*scale. When scale = 0, the upper threshold is 0 and the output is always 0. When scale < 0, the upper threshold 6*scale < 0 and the output always equals 6*scale (both are mathematically valid boundary semantics). For int32, the threshold is truncated toward zero. Invalid ranges: scale is NaN/±Inf, or for int32 6*scale exceeds [-2147483648, 2147483648).

Invocation Description

Invocation method Sample code Description
GE graph mode test_geir_relu6_d.cpp Calls the Relu6D operator by building the graph through the operator IR.

Compile and run the sample (execute in the repository root directory):

bash build.sh --run_example relu6_d graph