Huahuan Zheng comments

Results 35 comments of


                                            Huahuan Zheng

openwrt下kcptun客户端连接服务器问题

> 可能copy paste时候不小心漏了一个0... ~~其实你只是多打了一个回车~~,像上面说的可以先试试能不能用你的quiet设置为false按理说有连接的话能看到类似stream open/close之类的信息

Count of Conformer parameters mismatch with that in the paper

This is kind of weird. I test several open-source Conformer implementation (I also implement it myself), but none of them can strictly match the reported number of parameters. Do you...

warning that forward/backward mismatch

If this is error related to the input data, it should repeat every epoch in training, but at the beginning, no warning is thrown. And as you can see, all...

warp-rnnt with compact memory layout implementation

benchmark results: (tested on RTX 3090) * forward only: https://github.com/maxwellzh/warp-rnnt/blob/master/pytorch_binding/BenchMarkForwardResults.txt * forward+backward: https://github.com/maxwellzh/warp-rnnt/blob/master/pytorch_binding/BenchMarkResults.txt

warp-rnnt with compact memory layout implementation

Maybe this one? https://arxiv.org/abs/1909.12415

warp-rnnt with compact memory layout implementation

Yes, that's it. We can add it in the README later.

warp-rnnt with compact memory layout implementation

The function `compactTensor` is just for testing. If we convert the tensor by that function every time invoking `rnnt_loss()`, the overall performance might be poor. Normally, in Joint Network, we...

warp-rnnt with compact memory layout implementation

It's something outside the rnnt-loss, so I'd like to create a new repo including these functions implementation. It will take some time for me to prepare the codes. I'll let...

warp-rnnt with compact memory layout implementation

I have released the implementation of these functions. https://github.com/maxwellzh/torch-gather