Piotr Żelasko
Piotr Żelasko
As lilcom is often used through Lhotse, several Lhotse users have reported that the decompression is computationally expensive; some of my colleagues performed dataloader profiling and found it takes 50%...
Issue: when running `pip install kaldifeat`, I'm getting the error that it can't find MKL libraries to link against. ``` [ 50%] Linking CXX shared library ../../lib/libkaldifeat_core.dylib ld: library not...
It's nothing super important -- but it would have been convenient to support gzip-compressed ARPA LMs. If it was in Python you could use `import gzip`, but if it's native...
### 🐛 Describe the bug Technically it's not a bug, but it was the most fitting category. I benchmarked torchaudio vs ffmpeg for reading a long OPUS file (> 1h...
When running `nara_wpe.torch_wpe.wpe_v6` with PyTorch 1.11, I'm seeing the following warning: ``` UserWarning: torch.solve is deprecated in favor of torch.linalg.solveand will be removed in a future PyTorch release. torch.linalg.solve has...
I think one of the most common issues with bucketing samplers (dynamic or static) is that they tend to sample mini-batches with long utterances that blow up CUDA memory. This...
The main idea is to automatically decrease the batch size in presence of "very long" examples. Two main questions are: - how much to decrease the batch size - how...
@danpovey it may address the issue described in #678; but I haven't tested it beyond running unit tests successfully. I added a background thread for collect_cuts_in_buckets. Threading should be sufficient,...
I'd like to compile a list of topics that would be good tutorial material, I can contribute some of those, and ideally we'd also get some community contributions. I've refreshed...