---
kind: concept
slug: grouped-convolution
title: Grouped convolution
summary: >-
  Splits a convolution's channels into G groups, each filtered only on its own slice, so FLOPs and parameters fall roughly by 1/G. Groups do not exchange information by themselves. DualConv adds a full 1×1 branch over all input channels to keep that information flowing. The reported gains are author-reported and come from their own setups.
tags: [computer-vision, convolution, efficient-inference]
papers: ["2202.07481"]
related: [depthwise-separable-convolution, efficient-vision-backbone, convolutional-neural-network]
status: draft
updated_at: 2026-11-15
---

## Overview

A standard convolution lets every output filter read all input channels. A
grouped convolution splits the input and output channels into G groups, so each
filter reads only its own group's slice of the input. The cost falls to about
1/G of the standard layer [arxiv:2202.07481 §III.A] [eq. 9].

The price is that groups never see each other's channels. The usual fix is a
channel shuffle that reorders channels between layers so later groups mix
earlier ones [arxiv:2202.07481 §II.B]. This page is so far based on one card,
DualConv, which is the only card that lists this slug. Its design is one answer
to the same problem: it keeps the grouped 3×3 branch, drops the shuffle, and
adds a full 1×1 branch [arxiv:2202.07481 §III.A].

## Variants

### Plain grouped 3×3 convolution

Each group of N/G filters reads M/G of the M input channels, with 3×3 kernels.
The paper's comparison gives its FLOP ratio to a standard layer as 1/G
[arxiv:2202.07481 eq. 9]. Its drawback is the lack of information flow between
groups, which the paper names as the reason for the channel shuffle
[arxiv:2202.07481 §II.B].

### Grouped convolution with channel shuffle

A channel shuffle after the grouped layer moves channels between groups, so that
information exchanges across them [arxiv:2202.07481 §II.B]. The card records no
cost figure for the shuffle itself, and no new result for this variant in the
paper.

### DualConv: grouped 3×3 plus full 1×1

Card: [DualConv](/papers/2202.07481).

Each DualConv layer sums two branches with the standard layer's input and output
shape [arxiv:2202.07481 §III.A]:

- a 3×3 grouped branch in which the 3×3 slice moves from group to group,
  without the shifted layout of heterogeneous convolution [fig. 2];
- a 1×1 branch in which every filter reads all M input channels
  [fig. 1(e)].

Its FLOP ratio to a standard 3×3 layer is 1/G + 1/K², so 1/G + 1/9 for K = 3
[arxiv:2202.07481 eq. 5]. For large G the cost approaches the 1/9 floor, which
is the cost of a 1×1 layer. The paper gives no parameter formula.

Reported results (author-reported, see the card for setups): VGG-16 on
CIFAR-10 at G=4 reaches 94.14% against 93.95% with 114.24M FLOPs against
313.21M [arxiv:2202.07481 tab. I]. MobileNetV1 on CIFAR-100 at G=4 reaches
72.24% against 68.13% [arxiv:2202.07481 tab. IV]. YOLO-V3 on PASCAL VOC 2007 at
G=4 reaches 45.64 mAP against 41.24 with about half the FLOPs [arxiv:2202.07481
tab. VII].

## When to use which

The choice depends on the backbone, the cost target and the hardware. Rows cite
the paper's own tables.

| Situation | Use | Why |
|---|---|---|
| Cheaper drop-in for a 3×3 layer in a classic CNN, cost is the main constraint | DualConv with G chosen from the table | Cost ratio 1/G + 1/9 [arxiv:2202.07481 eq. 5]; VGG-16 at G=4 cuts FLOPs well below the baseline with higher reported accuracy [arxiv:2202.07481 tab. I] |
| Accuracy is the priority on ResNet-50 at CIFAR-10 | DualConv at G=8 | Reported 94.33% against 94.08%, at 923M FLOPs against 1.30G [arxiv:2202.07481 tab. I] |
| Starting from a MobileNet-style backbone | Check the depthwise separable baseline first | MobileNetV1 rows report more FLOPs and parameters than the baseline at G=4 [arxiv:2202.07481 tab. II] [arxiv:2202.07481 tab. IV] [arxiv:2202.07481 tab. VI] |
| Lightweight target on CIFAR-100 MobileNetV2 | G=4 with an accuracy check | FLOPs fall from 64.96M to 31.07M, but accuracy drops from 68.54% to 67.86% [arxiv:2202.07481 tab. IV] |
| ImageNet MobileNetV2 | Avoid, or verify on your task | The best DualConv row is 3.77 Top-1 points below baseline at G=2 [arxiv:2202.07481 tab. VI] |
| Latency on CPU is the constraint | Measure each G on the target | The best CPU time is at G=16 for VGG-16 and G=8 for ResNet-50, not at the most accurate setting [arxiv:2202.07481 tab. V] [arxiv:2202.07481 §IV.G] |
| A reproducible reference implementation is needed | Write your own from the description | No official code was found [arxiv:2202.07481 §III.A]; see the card |

## Implementation sketch

Language-agnostic pseudocode of the DualConv layer, from the paper's forward
description. The plain grouped convolution is the first branch alone, without
the 1×1 sum.

```pseudocode
# DualConv layer  [arxiv:2202.07481 §III.A] [fig. 1(e)]
# Input:  x (B, M, H, W); N filters split into G groups; 3x3 kernels
# Output: y (B, N, H, W)

def dual_conv(x, weights_3x3, weights_1x1, groups):
    # Grouped 3x3 branch: each group of N/G filters reads M/G channels,
    # and the channel slice moves from group to group  [§III.A]
    y_3x3 = grouped_conv(x, weights_3x3, groups)    # (B, N, H, W)
    # Full 1x1 branch: every filter reads all M input channels  [§III.A]
    y_1x1 = conv1x1(x, weights_1x1)                 # (B, N, H, W)
    # Sum the two branches; no channel shuffle follows  [§III.A]
    return y_3x3 + y_1x1
```

## Contradictions

- **VGG-16 on CIFAR-100 and the group count.** The text says accuracy improves
  as G grows to 8 [arxiv:2202.07481 §IV.C]. The table gives the best value at
  G=2 and a steady decline after it [arxiv:2202.07481 tab. III].
- **FLOP reduction and MobileNetV1.** The cost ratio in [eq. 5] compares
  against a standard convolution. The MobileNetV1 baseline replaced here is
  already depthwise separable [arxiv:2202.07481 §IV.B]. Against that baseline the
  table rows show higher FLOPs and parameters for DualConv [arxiv:2202.07481
  tab. II] [arxiv:2202.07481 tab. VI]. The paper does not reconcile the two.
- **FLOP count and measured speed.** The FLOP ratio falls as G grows, but CPU
  timings do not fall monotonically, and GPU timing did not separate the models
  [arxiv:2202.07481 tab. V] [arxiv:2202.07481 §IV.G].
- **HetConv reproduction.** The paper says its reproduced HetConv is 0.2–0.76%
  worse than the originally reported figures [arxiv:2202.07481 §IV.A]. The
  original figures are not in the paper's tables, so the gap cannot be checked
  from them.

## Papers

- [DualConv](/papers/2202.07481): adds a full 1×1 branch to a grouped 3×3 convolution and drops the channel shuffle; reports accuracy gains on several classifiers and YOLO-V3, with losses on others.
