`split`¶

Cut one axis into a first section and the remainder.

Category: shape · Identity: split@1

Shape¶

x -> first, rest

Relation: first[dim] == size; rest[dim] == x[dim] - size; all other axes equal

Port	Direction	Pattern	dtype
`x`	input	`x`	compute
`first`	output	`first`	compute
`rest`	output	`rest`	compute

Arguments¶

Name	Type	Default	Constraints	Description
`size` (positional)	int	inferred	>= 1; <= 2147483647	Extent of the first section. Inferred when the consumers determine it.
`dim`	int	`1`	>= 1	Axis to split; the batch axis 0 cannot be split.

Description¶

Returns (x[:size], x[size:]) along dim as views sharing autograd with the input. Exactly two non-empty sections are produced; this is not a repeated chunking. After split there is no single current tensor, so the next operation must name its input.

Examples¶

Example 1¶

Both sections feed explicit branches; split clears the current tensor.

z1, z2 = split(64)
linear(z1, 32)
features = relu()
style = linear(z2, 32)
add(features, style)

Input ['B', 128] → output ['B', 32].

Network: [B, 128] -> [B, 32]  dtype=float32
index  name  operation  input shapes          output shapes
0      n0    split      x=[B, 128]            first=[B, 64], rest=[B, 64]
1      n1    linear     x=[B, 64]             out=[B, 32]
2      n2    relu       x=[B, 32]             out=[B, 32]
3      n3    linear     x=[B, 64]             out=[B, 32]
4      n4    add        a=[B, 32], b=[B, 32]  out=[B, 32]

Parameters: 4,160

Example 2¶

The first size 96 is inferred from the remainder.

a, b = split()
out = b