In this article

Padding

|
  • Share via
  •  Edit this article

Padding

[class] Padding

The Func Padding pads the input_tensor with the specified padding. The argument is given in the following format. 

(output_tensor, input_tensor, *index_components, *padding_starts, *padding_ends)
where each (length index_components), (length padding_starts), and (length padding_ends) == ndim(input_tensor) == ndim(output_tensor).

With JIT=1, everything is expected to be fused in a single EXPR like: 

out = where(start_0 < index_components[0] < end_0 and start_1 < index_components[1] < end_1 and ..., input_tensor[i=i'-start_0, j=j'_start_1, ...], value)

[function] !padding

(!padding x padding &key (value 0.0))
Pads the tensor x with the specified padding and value.

Padding is specified as: ((pad_0_1 pad_0_2) (pad_1_0 pad_1_0) ...). If T is provided instead of (pad_0_0 pad_0_1), that means no padding is applied to the axis. (that is, T = (0 0))

Each start_0 and pad_0 is expected as a positive integer (caten will not check this!), or a scalar int32/int64 tensor.

[function] !padding2d

(!padding2d x padding &key (value 0.0))
Pads the last two dimension of the tensor x with the specified padding and value..

padding is specified as: (padding_left padding_right padding_top padding_bottom).