burkliu

fail to install kaldifeat in conda

25

pip install kaldifeat Looking in indexes: http://mirrors.cloud.tencent.com/pypi/simple Collecting kaldifeat Using cached http://mirrors.cloud.tencent.com/pypi/packages/2d/7b/0a5c9254de6a62cdca7bea829bd8d5d646e18508d34f9214f048ee003fed/kaldifeat-1.17.tar.gz (477 kB) Preparing metadata (setup.py) ... done Building wheels for collected packages: kaldifeat Building wheel for kaldifeat (setup.py)...

low GPU utilization rate

2

In locally sync mode, the GPU utilization is unstable (sometime 100%, sometime 0%), and the average GPU utilization is about 60%. I think this question may caused by IO when...

Attemp to add schedule sampling in ComputePredictionFunctional failed

11

https://github.com/tensorflow/lingvo/blob/2d05484a7d5d73db23f8a4b47d6d729b5e01fa6a/lingvo/tasks/asr/decoder.py#L1059 i found that ComputePredictionDynamic is too slow for schedule sampling, so i tried to add schedule sampling in the cell function RnnStep but failed following is my code ```...

about the bug when decoding streaming ASR

5

`2020-08-17T03:05:45.568884032Z File "neural_sp/neural_sp/models/seq2seq/decoders/transformer.py", line 968, in beam_search 2020-08-17T03:05:45.568887299Z rightmost_frame = max(0, aws_last_success[0, :, 0].nonzero()[:, -1].max().item()) + 1 2020-08-17T03:05:45.568890475Z RuntimeError: invalid argument 1: tensor must have one dimension at /pytorch/aten/src/TH/generic/THTensorEvenMoreMath.cpp:590` I...

How to print attention summary in tensorboard?

2

hello! I opened the event output using tensorboard, but I couldn't find the attention summary of decoder, how should I do?

关于多种原矿提炼计算器功能升级提案

17

作者你好，我是eve手游玩家，程序员，目前基于python开发了一个机器学习算法，可以自动计算优化挖取目标成品材料所需要的原矿种类以及数量，有意请与我联系，QQ616310753  

a streaming pipeline

6

Hello guys, I wrote a streaming inference pipeline in Python for this project, including torch jit script, int8 dynamic quantization, and streaming interface for the audio encoder and decoder (style...

operator optimization

3

https://github.com/ubisoft/ubisoft-laforge-ZeroEGGS/blob/4992d9ab5e28bc4bb13219a598906562f5786448/ZEGGS/modules.py#L461C5-L481 consider relace with ``` def forward(self, x): """ Forward function of Positional Encoding: x = (B, N) -- Long or Int tensor """ # initialize tensor nb_frames_max = torch.max(torch.cumsum(x,...

Failed to replication the experiment on lingvo

1

I tried to replicate your experiment on lingvo, and I tried to keep everything you mentioned same. However, I can only get minimum WER of 25% upon all my experiment....

[debug] support flow cache, for sharper tts_mel output (handle prompt bug)

1

#379 问题2的解决方案 flowmatching中的z和mu，跨chunk时对于每个index不是定值，是导致衔接处频谱模糊的因素之一（本质是flow的attention context问题，无解。需要重训模型+context cache，非常复杂） 有益效果： 1、可以改善衔接处频谱模糊、发音不清晰问题 2、可以改善衔接处断音问题 3、可以改善音量突变或音量异常问题 图中是flow的tts_mel输出，用于对比上下文及频谱模糊的问题 大图1列：不带cache；2列：带cache 小图左：前chunk最后34；中：（前+后）/2；右：后chunk开头34 可以发现带cache的，tts_mel频谱更清晰 需要注意的是，对于zeroshot，由于带了prompt输入，所以这部分的cache也需要额外考虑 备注： 1、由于后续的mel fade、hifigan cache、speech fade的挽救，该项虽然更本质，但最终听感提升概率较小 2、对于默认配置，断音出现周期为2s，若此时刻恰好没有发音，就难以发现区别，因此不保证每条case必然出现至少一处改善 3、flow matching 的输入Z和MU使用cache结果是有意义的。首先可以减少随机性、提升结果稳定性；其次未来做causal同样需要使用此cache，是可以复用的 已测有效的用例： SFT `cosyvoice = CosyVoice('pretrained_models/CosyVoice-300M-SFT') cosyvoice.inference_sft('收到好友从远方寄来的生日礼物，那份意外的惊喜与深深的祝福让我心中充满了甜蜜的快乐，笑容如花儿般绽放。', '中文女',...