`pad`¶

Enlarge trailing axes with constant, reflected, or replicated borders.

Category: shape · Identity: pad@1

Shape¶

x -> out

Relation: out[axis] == x[axis] + left + right for each padded axis; all other axes equal

Port	Direction	Pattern	dtype
`x`	input	`x`	compute
`out`	output	`out`	compute

Arguments¶

Name	Type	Default	Constraints	Description
`padding` (positional)	ints	`()`	>= 0; <= 2147483647	(left, right) pairs starting at the last axis, as torch.nn.functional.pad orders them. Non-negative; two values per padded axis.
`value`	float	`0.0`	—	Fill value for mode='constant'.
`mode`	str	`"constant"`	one of `constant`, `reflect`, `replicate`	Border rule: 'constant' fills with value, 'reflect' mirrors without repeating the edge, 'replicate' repeats the edge.

Positional values fill padding.

Description¶

torch.nn.functional.pad(x, padding, mode=mode, value=value).

padding lists (left, right) pairs starting from the last axis, exactly as F.pad orders them:

rank 2 [B, F]: pad(left, right) widens the feature axis F.
rank 3 [B, C, L] (equivalently [B, T, D]): pad(left, right) widens the last axis L.
rank 4 [B, C, H, W]: pad(left, right, top, bottom) widens W first and then H.

The batch axis is never padded, so padding holds at most rank - 1 pairs; anything longer, or an odd number of values, is E_ARGUMENT. Each padded axis grows by left + right and every other axis is unchanged, and the relation runs in both directions: a known input extent fixes the output, and a known output extent fixes the input.

mode='constant' fills the new positions with value and works for every supported rank. mode='reflect' and mode='replicate' follow the PyTorch restriction to spatial padding: a rank-3 tensor padded on its last axis, or a rank-4 tensor padded on H and W; other rank/padding combinations are rejected with E_ARGUMENT. reflect additionally requires each pad width to be smaller than the extent it mirrors (E_CONSTRAINT), because the border itself is not repeated. value must stay at its default for the non-constant modes, which PyTorch does not accept.

There are no parameters, behavior is identical in train and eval mode, the output keeps the input dtype, and the gradient of the padded positions is discarded (constant) or accumulated back onto the mirrored or repeated source positions (reflect, replicate).

Examples¶

Example 1¶

One column on each side of W and one row on each side of H, so the 3x3 convolution preserves the image size.

pad(1, 1, 1, 1)
conv(4, kernel_size=3)

Input ['B', 3, 8, 8] → output ['B', 4, 8, 8].

Network: [B, 3, 8, 8] -> [B, 4, 8, 8]  dtype=float32
index  name  operation  input shapes      output shapes
0      n0    pad        x=[B, 3, 8, 8]    out=[B, 3, 10, 10]
1      n1    conv       x=[B, 3, 10, 10]  out=[B, 4, 8, 8]

Parameters: 112

Example 2¶

Mirrors two positions onto each end of L in a [B, C, L] signal.

pad(2, 2, mode="reflect")

Input ['B', 3, 8] → output ['B', 3, 12].

Network: [B, 3, 8] -> [B, 3, 12]  dtype=float32
index  name  operation  input shapes  output shapes
0      n0    pad        x=[B, 3, 8]   out=[B, 3, 12]

Parameters: 0

Example 3¶

Backward inference: the projection width 8 follows from the padded output.

linear()
pad(1, 1)

Input ['B', 16] → output ['B', 10].

Network: [B, 16] -> [B, 10]  dtype=float32
index  name  operation  input shapes  output shapes
0      n0    linear     x=[B, 16]     out=[B, 8]
1      n1    pad        x=[B, 8]      out=[B, 10]

Parameters: 136

Example 4¶

A one-position border of ones on the right and bottom edges.

pad(0, 1, 0, 1, value=1.0, mode="constant")

Input ['B', 2, 4, 4] → output ['B', 2, 5, 5].

Network: [B, 2, 4, 4] -> [B, 2, 5, 5]  dtype=float32
index  name  operation  input shapes    output shapes
0      n0    pad        x=[B, 2, 4, 4]  out=[B, 2, 5, 5]