BatchNormalization

Description

Normalizes the input tensors in batches.

Input

• X: a 4D tensor of type float16 or float, in ND format.
• scale (optional): tensor of type float, specifying the scale factor. The data format can be ND.
• B (optional): tensor of type float, specifying the offset. The data format can be ND.
• mean (optional): tensor of type float, specifying the mean value. The data format can be ND.
• var (optional): tensor of type float, specifying the variance. The data format can be ND.

Attribute

• epsilon (optional): float, added to var to avoid dividing by zero. Defaults to 0.0001.
• momentum: float, not supported currently.

Output

• Y: tensor of type float16 or float. The data format can be ND.
• running_mean (optional): tensor of type float, indicating the mean value. The data format can be ND.
• running_var (optional): tensor of type float, indicating the variance. The data format can be ND.
• saved_mean (optional): tensor of type float. The saved mean value is used to speed up gradient calculation during training (not supported in versions later than v14). The data format can be ND.
• saved_var (optional): tensor of type float. The saved variance is used to speed up gradient calculation during training (not supported in versions later than v14). The data format can be ND.

Constraints

None.

ONNX Opset Support

Opset v8/v9/v10/v11/v12/v13/v14/v15/v16/v17/v18