Dahan Gong
Dahan Gong
Yes it works. Thanks for your work. Though I think it's very natural and easy for `F5` to keep an old list - it just needs to: 1) copy the...
Could you do a simple test to find what Vimium got? * open https://gdh1995.cn/vimium-c/keyboard-test.html * click Vimium's icon and save options to make Vimium disabled on the page * reload...
Then this is an issue of Firefox itself, or does another extension hook the keydown events?
You may use ``, and it works like `` at most places (when no `map someCommand`). And `KeyboardUtils.isEscape` supports `mapkey ` in key mappings.
What will the test page log for `Ctrl+[` on your Firefox when Vimium is disabled?
During my tests on Firefox 122, Win11, `keydown#key=Escape` will be blocked by `permissions.default.shortcuts=2`, while `Ctrl+[` and `Ctrl+]` works as intended.
TensorRT 基本上是个黑盒子；Nvidia 有自己的 profiler 工具来衡量各层速度，不过吞吐量我不确定要怎么算。 python: https://docs.nvidia.com/deeplearning/tensorrt/api/python_api/infer/Core/Profiler.html C++: https://docs.nvidia.com/deeplearning/tensorrt/api/c_api/classnvinfer1_1_1_i_profiler.html#details
@kpzhang93 谢谢您的工作！我有两个关于pad的问题想请教您： * 请问pad操作后的tmpim变量可以初始化为127.5吗？因为我看后边会把它缩放到以[-1, 1]，而现在的zeros相当于直接填充了黑色，不知道灰色会不会好些呢？ * first stage得到的total_boxes后，现在是先做了一次类似bbreg的操作才传给pad，而second到third的时候就直接调用bbreg了，我能找到的区别只有“前者的`bbw`、`bbh` 比 `bbreg` 里的w和h 要小1”，请问这个区别是有益的吗？是不是first stage模型训练的时候就是这么训练的，导致不适合再加1了呢？
原理和torch的qat是一致的，都是额外加onnx node来存储量化信息，所以理论上肯定能在 trt 跑，只是仓库里目前的代码不一定跑得通，可能需要你看着改改
啊我也不咋了解torch的qat，只是按官网例子跑过几天，有两年没接触了没法解答