Inspired by Universal Inverted Bottleneck (UIB) in MobileNetV4, we build our RepGhostNetV2 by utilizing RepGhost bottleneck from RepGhost. As shown in the Figure blow, RepGhostNetV2 differs to MobileNetV4 only in the shortcut connections.
The shortcut connections greatly facilitate the optimizations of CNNs, and even not affect the latency. It is worth benchmarking its key role in light-weight CNNs. And more variants may be possible.
We directly use the training settings from RepGhost repo to evaluate their performances. Refer to train.sh
for more training details.
Re-parameterization is not applied here during training.
The latency are evaluated on an iPhone12 based on ModelBench.
Model | Size | Params | FLOPs | Latency(ms) | Top-1 Acc.(%) |
---|---|---|---|---|---|
RGv2-S | 224 | 3,772,744 | 186,006,848 | 1.01 | 72.3 |
MNv4-S | 224 | 3,772,744 | 186,006,848 | 1.01 | 72.3 |
RGv2-M | 256 | 9,714,232 | 1,080,245,504 | 1.48 | 79.5 |
MNv4-M | 256 | 9,714,232 | 1,080,245,504 | 1.48 | 79.5 |
RGv2-L | 384 | 32,589,584 | 6,376,221,952 | 3.63 | 81.5 |
MNv4-L | 384 | 32,589,584 | 6,376,221,952 | 3.63 | 81.5 |
- Tune the training setting to reproduce results of MobileNetV4.
- Release checkpoints.
- Support Mobile MQA attention block.
- More variants of shortcut connections in light-weight CNNs.
If RepGhostNetV2 helps your research or work, please consider citing:
@misc{chen2024repghostnetv2,
author = {Chengpeng Chen},
title = {RepGhostNetV2: When RepGhost Meets MobileNetV4},
year = {2024},
publisher = {GitHub},
journal = {GitHub repository},
url = {https://github.com/ChengpengChen/RepGhostNetV2},
}
If you have any questions, please contact chencp@live.com.