Fangjun Kuang

加载 PNNX 导出的模型时出现 layer aten::exp not exists or registered

5

## error log | 日志或报错信息 | ログ ``` layer aten::exp not exists or registered layer pnnx.Expression not exists or registered ``` ## model | 模型 | モデル 1. original model...

Decoding speed

14

I am using a pretrained model from https://huggingface.co/speechbrain/asr-transformer-transformerlm-librispeech with `EncoderDecoderASR` to decode the `test-clean` and `test-other` datasets from LibriSpeech on a v100 NVIDIA GPU with a batch size of 2....

Error in cub::DeviceHistogram::HistogramEven

3

The following code fails when invoking `cub::DeviceHistogram::HistogramEven`. **NOTE**: It fails **ONLY** for some values of `n` and `dim` in the code below. @danpovey ```bash (py38) fangjun:~/open-source/k2/build_debug$ ./bin/cu_cub_test Invoking DeviceHistogramInitKernel() Invoking...

Don't install torchaudio automatically if torch is installed

8

The following code https://github.com/lhotse-speech/lhotse/blob/08a613a06d258ab72679f0d2e952ec668afbe185/setup.py#L140-L156 installs `torchaudio` automatically for the user, which is.not.correct, I believe. The reasons are: (1) The following exception message is never seen by the user since the...

Check the number of expected utterances in data preparation

1

I suggest adding some checks like below to dataset preparation ```python3 assert len(manifests[part]['recordings']) == expected_number_of_recordings # we can also check expected durations ``` With such checks, we can avoid issues...

Pipe the output of ffmpeg to cat before reading it.

4

See https://github.com/lhotse-speech/lhotse/issues/666#issuecomment-1100574606

Got warnings when loading sampler's state_dict

11

I get the following warning while trying to use https://github.com/k2-fsa/icefall/pull/259 to restore the state dict of a sampler from a checkpoint. ``` lhotse/dataset/sampling/simple.py:144: UserWarning: SimpleCutSampler.load_state_dict(): Inconsistent time_constraint: expected TimeConstraint(max_duration=10, max_samples=None,...

Misleading feature type

3

In https://github.com/lhotse-speech/lhotse/blob/d9c4141319adb39f64684c762aa541467d25f7fc/lhotse/kaldi.py#L144-L145 It uses `kaldiio` as the feature type. However, https://github.com/lhotse-speech/lhotse/blob/d9c4141319adb39f64684c762aa541467d25f7fc/lhotse/features/base.py#L407-L408 says possible types are `fbank`, `mfcc`, etc.

See https://github.com/csukuangfj/kaldi_native_io `kaldi_native_io` is implemented in C++ and wrapped to Python. ## Example usage ### Read/write `kaldi::Matrix` ```python import numpy as np import kaldi_native_io base = "float_matrix" wspecifier = f"ark,scp,t:{base}.ark,{base}.scp"...

Book not completed

2

I really appreciate your work. Is there any plan to continue this book? For example, the following pages are still missing: - http://www.bravegnu.org/gnu-eprog/c-library.html - http://www.bravegnu.org/gnu-eprog/inline-assembly.html