`flatten`¶

Collapse every non-batch axis into one feature axis.

Category: shape · Identity: flatten@1

Shape¶

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

Relation: F == prod(x[1:])

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

Arguments¶

This operator takes no scalar arguments.

Description¶

Equivalent to x.reshape(batch, -1). The input rank can be inferred backward only through the element count when the other axes are known.

Examples¶

Example 1¶

flatten()
linear()

Input ['B', 3, 8, 8] → output ['B', 2].

Network: [B, 3, 8, 8] -> [B, 2]  dtype=float32
index  name  operation  input shapes    output shapes
0      n0    flatten    x=[B, 3, 8, 8]  out=[B, 192]
1      n1    linear     x=[B, 192]      out=[B, 2]

Parameters: 386

flatten¶