JierunChen

Hi, thanks for your interest in the work. If you are referring 128, 256, 512, and 1048 as the network widths of four stages in a single network, it has...

@xyh001007 Hi, PWConv refers to the pointwise convolution or the conv 1x1.

@jackiezhang11 Hi, both 416 and 448 can be the input size. The detection performance would degrade without pre-training. And yes, FasterNet should be applicable as a backbone for YOLOV3.

Yes, please refer to the relevant sections in the paper and [code](https://github.com/JierunChen/FasterNet/tree/master/detection).

Hi, thanks for your interest in this work. The "slicing" mode w/o ```x = x.clone()``` in its implementation can be faster than the "split_cat" mode, particularly for GPUs and large...

Hi, we did not apply BN after PConv in FasterNet, but after the Conv $1\times1$. In your case, you may apply the BN on only the channels processed and exclude...

Hi, if you are replacing the regular Conv in ResNet with PConv, don't forget to remove the activation layer as it helps the ResNet but hurts the expressiveness of the...

> > Hi, if you are replacing the regular Conv in ResNet with PConv, don't forget to remove the activation layer as it helps the ResNet but hurts the expressiveness...

Hi, as we said in the ablation study of the paper, for the activation function, we empirically found that GELU fits FasterNet-T0/T1 models more efficiently than ReLU. It, however, becomes...

Hi, this work shares the same motivation as GhostNet to leverage the similarity of feature maps. However, there are some key differences in terms of the design: 1. A Ghost...