`clamp`¶

Saturate every element into the closed interval [min, max].

Category: arithmetic · Identity: clamp@1

Shape¶

x[B, ...] -> out[B, ...]

Port	Direction	Pattern	dtype
`x`	input	`x[B, ...]`	compute
`out`	output	`out[B, ...]`	compute

Arguments¶

Name	Type	Default	Constraints	Description
`min` (positional)	float	required	—	Lower bound; elements below it become min.
`max` (positional)	float	required	—	Upper bound; elements above it become max. Must be >= min.

Description¶

Elementwise out = min(max(x, min), max), computed as torch.clamp(x, min=min, max=max) out of place.

The operation is shape preserving: every supported rank ([B, F], [B, T, D], [B, C, H, W]) passes through unchanged, and no axis has a special meaning. Both bounds are required floats and min <= max is checked when arguments are normalized (E_ARGUMENT otherwise).

The gradient is one strictly inside the interval and zero outside it, so a saturated element stops contributing to its input's gradient; elements exactly on a bound keep a gradient of one, matching torch.clamp. Behavior is identical in train and eval mode and there are no parameters. Computation stays in the input dtype, so float16 and bfloat16 activations are clamped without an upcast.

Examples¶

Example 1¶

A ReLU6-style activation on the hidden features.

linear(64)
clamp(0.0, 6.0)
linear()

Input ['B', 128] → output ['B', 10].

Network: [B, 128] -> [B, 10]  dtype=float32
index  name  operation  input shapes  output shapes
0      n0    linear     x=[B, 128]    out=[B, 64]
1      n1    clamp      x=[B, 64]     out=[B, 64]
2      n2    linear     x=[B, 64]     out=[B, 10]

Parameters: 8,906

Example 2¶

Saturating an image tensor into [-1, 1].

conv(8, kernel_size=3, padding=1)
clamp(-1.0, 1.0)

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

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

Parameters: 224

Example 3¶

On a [B, T, D] sequence the bounds apply to every element.

linear(8)
clamp(0.0, 1.0)
linear()

Input ['B', 4, 6] → output ['B', 4, 3].

Network: [B, 4, 6] -> [B, 4, 3]  dtype=float32
index  name  operation  input shapes  output shapes
0      n0    linear     x=[B, 4, 6]   out=[B, 4, 8]
1      n1    clamp      x=[B, 4, 8]   out=[B, 4, 8]
2      n2    linear     x=[B, 4, 8]   out=[B, 4, 3]

Parameters: 83

clamp¶