JierunChen

Hi @1072010918 , the pre-trained weights can be loaded by specifying the weight file in the [config](https://github.com/JierunChen/FasterNet/blob/master/detection/configs/fasternet/mask_rcnn_fasternet_l_fpn_1x_coco.py) files.

@juan19941228 Yes, equation (2) only compares the memory access for spatial feature extraction and does not account for the memory access by the pointwise convolution (PWConv).

@abcsimple Hi, the comparison of equations (2) and (3) is based on a reasonable assumption that the width or the number of channels when using DWConv is generally higher than...

@wsy-yjys Hi, the slicing mode can be slower because of the feature map clone (see the code ```x = x.clone() ```). Such clone is necessary to avoid modifying the input,...

@wsy-yjys Hi, ```shortcut = x``` is a shallow copy, which can be much faster than the deep copy version of ```x = x.clone()```. Therefore, the implementation as you suggested would:...

Hi, we haven't done it yet. You are welcome to monitor any follow-up work or try it by yourself. Note that you may put some add-ons, e.g., the squeeze and...

@LKAMING97 Hi, T-shaped Conv requires non-trial implementation and thus we adopt out-of-the-box PConv and PWConv. The combination of PConv and PWConv also has lower FLOPs and fewer parameters compared to...

@LKAMING97 Hi, the PConv can be changed into 1D, whose effectiveness depends on the input redundancy of your task.

@LKAMING97 Hi, the "evaluation command" you mentioned refers to the evaluation of performance, e.g., accuracy, regardless of the latency. Therefore, "fuse_conv_bn" is not compulsory. You may also turn it on...

@1920230345 Hi, we did not conduct an ablation study on this. We suggest empirical experiments for different FasterNet variants, as further incorporating activation functions may increase or decrease the model...