Yongchuan Cui
Yongchuan Cui
fixed. change `np.dot(delta, activations[-2].transpose())` to `np.outer(delta, activations[-2].transpose())` The shape of the matrices are ` (a, )` and `(b, )`, respectively. When we use `np.dot` while the matrices are 1d both,...
> 微信群已满，请加QQ交流群912434813 QQ群搜不到？
sorry for my mistake. Indeed it's clock skew. @vanpelt @nate-wandb
> hi!, @cuiyc2000 ,How can I solve this clock skew problem, I have the same problem as you? The clock sync tool `ntpdate` works fine for me to solve this...
> 试试`GhostModule( inp, oup, kernel_size=1, ratio=2, dw_size=3, stride=1, relu=False)` > > 另外，EfficientNet里PWConv换成GhostModule，DWConv就不要再换成GhostModule了 是的，我只把PWConv替换成GhostModule。按您给的参数还是不行诶，和Conv2d差别很大
> 网络首尾层的Conv不要换，影响较大。另外，log能发来看看吗？ 首尾的1x1卷积比较耗时吧，我想用GhostModule替换掉。我想要牺牲精度提高EfficientNet的效率，如果不替换首尾，那该替换哪呢？
这是替换之前的： ``` loading annotations into memory... Done (t=0.01s) creating index... index created! loading annotations into memory... Done (t=0.00s) creating index... index created! [Warning] Ignoring Error(s) in loading state_dict for EfficientDetBackbone:...
> 替换后的backbone也得先imagenet预训练，然后再在detection上finetune吧 哦哦，好 那么这个将原来的1x1卷积替换为GhostModule应该是可行的吧
> @jinfagang Still the same error. You didn't modify the code in `detector/nms/src/nms_kernel.cu` at branch master.
> And it turns out that this is a real show-stopper for A40 and A100 cards. They require CUDA 11.1 or higher, thus PyTorch 1.11 or higher, which does not...