global_avg_pool

Average each channel over height and width, producing one value per channel.

Category: spatial · Identity: global_avg_pool@1

Shape

x[B, C, H, W] -> out[B, C]

Port	Direction	Pattern	dtype
x	input	x[B, C, H, W]	compute
out	output	out[B, C]	compute

Arguments

This operator takes no scalar arguments.

Description

Computes out[b, c] = mean(x[b, c, :, :]) on a [B, C, H, W] input, returning [B, C]: the spatial axes are reduced away rather than kept as size-1 axes, so the result feeds linear directly.

This is adaptive_avg_pool(1) followed by flatten, written as one operation because the pattern is the standard head of a convolutional classifier. There are no parameters and train and eval behave identically.

Only rank-4 inputs are accepted, which keeps the axis convention unambiguous; reduce a [B, C, L] sequence with a reshape to [B, C, L, 1] first, or use adaptive_avg_pool.

Shape inference: C flows in both directions, while H and W are not inferable backward — every spatial extent produces the same output — so they must come from the network input or the preceding operation. The mean is accumulated in the input dtype.

Examples

Example 1

A convolutional trunk followed by a global-pooled classifier head; the head's in_features resolves to the 16 channels.

conv(16, kernel_size=3, padding=1)
relu()
global_avg_pool()
linear()

Input ['B', 3, 16, 16] → output ['B', 10].

Network: [B, 3, 16, 16] -> [B, 10]  dtype=float32
index  name  operation        input shapes       output shapes
0      n0    conv             x=[B, 3, 16, 16]   out=[B, 16, 16, 16]
1      n1    relu             x=[B, 16, 16, 16]  out=[B, 16, 16, 16]
2      n2    global_avg_pool  x=[B, 16, 16, 16]  out=[B, 16]
3      n3    linear           x=[B, 16]          out=[B, 10]

Parameters: 618

Example 2

Any spatial extent collapses to one value per channel.

global_avg_pool()

Input ['B', 32, 7, 7] → output ['B', 32].

Network: [B, 32, 7, 7] -> [B, 32]  dtype=float32
index  name  operation        input shapes     output shapes
0      n0    global_avg_pool  x=[B, 32, 7, 7]  out=[B, 32]

Parameters: 0

Example 3

Spatial size no longer constrains the head, so the trunk can downsample freely.

conv(8, kernel_size=4, stride=2, padding=1)
group_norm(4)
relu()
global_avg_pool()
linear()

Input ['B', 3, 32, 32] → output ['B', 4].

Network: [B, 3, 32, 32] -> [B, 4]  dtype=float32
index  name  operation        input shapes      output shapes
0      n0    conv             x=[B, 3, 32, 32]  out=[B, 8, 16, 16]
1      n1    group_norm       x=[B, 8, 16, 16]  out=[B, 8, 16, 16]
2      n2    relu             x=[B, 8, 16, 16]  out=[B, 8, 16, 16]
3      n3    global_avg_pool  x=[B, 8, 16, 16]  out=[B, 8]
4      n4    linear           x=[B, 8]          out=[B, 4]

Parameters: 444