`reshape`¶

View the tensor with new non-batch dimensions, preserving the element count.

Category: shape · Identity: reshape@1

Shape¶

x -> out

Relation: prod(x[1:]) == prod(out[1:]); out[1:len(shape)+1] == shape

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

Arguments¶

Name	Type	Default	Constraints	Description
`shape` (positional)	ints	`()`	>= 1; <= 2147483647	Leading non-batch dimensions. Remaining dimensions are inferred.

Positional values fill shape.

Description¶

Returns x.reshape(batch, *shape). The batch axis is never reshaped. Give the leading dimensions positionally, reshape(512, 4, 4), or as shape=(512, 4, 4); exactly one omitted factor can be solved from the element count. Three leading dimensions fix an image [B, C, H, W]; shorter prefixes take their rank from the neighbouring operations. The plan records the full resolved shape.

Examples¶

Example 1¶

The projection width 128 follows from the reshape target.

linear()
reshape(8, 4, 4)

Input ['B', 16] → output ['B', 8, 4, 4].

Network: [B, 16] -> [B, 8, 4, 4]  dtype=float32
index  name  operation  input shapes  output shapes
0      n0    linear     x=[B, 16]     out=[B, 128]
1      n1    reshape    x=[B, 128]    out=[B, 8, 4, 4]

Parameters: 2,176

Example 2¶

Height and width are inferred from the element count and the output contract.

reshape(512)
conv(3, kernel_size=1)

Input ['B', 8192] → output ['B', 3, 4, 4].

Network: [B, 8192] -> [B, 3, 4, 4]  dtype=float32
index  name  operation  input shapes      output shapes
0      n0    reshape    x=[B, 8192]       out=[B, 512, 4, 4]
1      n1    conv       x=[B, 512, 4, 4]  out=[B, 3, 4, 4]

Parameters: 1,539

Example 3¶

A bare reshape flattens when the consumer fixes rank 2.

reshape()
linear()

Input ['B', 2, 4, 4] → output ['B', 10].

Network: [B, 2, 4, 4] -> [B, 10]  dtype=float32
index  name  operation  input shapes    output shapes
0      n0    reshape    x=[B, 2, 4, 4]  out=[B, 32]
1      n1    linear     x=[B, 32]       out=[B, 10]

Parameters: 330

Example 4¶

A one-dimension prefix with a [B, T, D] consumer yields a sequence of 16 tokens.

reshape(16)
linear(8)

Input ['B', 64] → output ['B', 16, 8].

Network: [B, 64] -> [B, 16, 8]  dtype=float32
index  name  operation  input shapes  output shapes
0      n0    reshape    x=[B, 64]     out=[B, 16, 4]
1      n1    linear     x=[B, 16, 4]  out=[B, 16, 8]

Parameters: 40

reshape¶