JierunChen

Hi, could you be more specific on your question? PConv should work with multiple GPUs. And there have been some works successfully reproducing our results, e.g., the [repo](https://github.com/liaomingg/FasterNet).

> "RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cuda:1 and cuda:0! (when checking argument for argument weight in method wrapper__cudnn_convolution)" this...

Hi, you could try enabling the clip_grad with proper value (e.g., 1) in the [cfg](https://github.com/JierunChen/FasterNet/blob/master/cfg/fasternet_t0.yaml) file.

@WYHZQ Hi, table 6 specifies the training and evaluation settings only for ImageNet-1k classification but not for detection. And it's not unusual that the optimal value of drop_path_rate varies for...

Hi, we used the widely used version of ILSVRC2012. We downloaded it from the [webpage](https://image-net.org/challenges/LSVRC/2012/index.php) after registration. And yes, feel free to apply FasterNet on MOT dataset as long as...

@870231652 Hi, you may consider the following: - The ReLu function should be removed. - ```.eval()``` mode should be enabled for model inference. - Benchmark mode can be enabled for...

@870231652 Hi, the PConv's statistics in the paper are measured by setting the partial ratio r=1/4 where in your case r=1/8. When using the 'slicing' mode, the line ```x =...

> hello,I did the same experiment. I converted it to onnx format and tested it on cpu. The speed was not as fast as that of DW conv Hi, the...

@870231652 Hi, for the PConv itself, $c_{in}=c_{out}=c_p$.

@orange-36 Hi, thanks for your correction. We'll update the paper accordingly.