search

huawei-noah / VanillaNet

MIT License
813 stars 57 forks source link

Performance on mobile/edge devices #9

Closed Monoclinic closed 1 year ago

Monoclinic commented 1 year ago

Hello, thanks for your great work! I notice that VanillaNet has more channels (512-1024-2048-4096) than traditional deep networks. As we all know, the bandwidth and computing power of cellphones and some edge devices (for example, Jetson series) are not so sufficient. However, intuitively a wider network will spend more time on memory access, and the FLOPs of VanillaNet have no advantage on deep networks. I wonder if you have tried VanillaNet under such conditions and is there any result?

HantingChen commented 1 year ago

Hello, thanks for your question and the recognition of our work! VanillaNet is mainly designed for GPUs with high parallelism, which is why it has a high number of channels, but also a high speed. On edge devices, where the level of parallelism is not as high, VanillaNet might not have an advantage. For these resource-limited devices, we would recommend GhostNet instead. You can find the GhostNet implementation here: https://github.com/huawei-noah/Efficient-AI-Backbones/tree/master/ghostnet_pytorch. Feel free to ask if you have any other questions!

Monoclinic commented 1 year ago

Thanks for your kind reply and advice.