Piotr Żelasko
Piotr Żelasko
No, it hasn’t been merged — I didn’t find any difference with this implementation in quick testing. Can you describe your environment a bit more? What’s your sampler, max_duration, num_workers,...
Absolutely! I'd be happy to accept contributions with tutorials for the topics you listed.
I agree with both of you, and I generally found that the simplest way to handle this is to resume the training with a different random seed. I recommend ditching...
Yes, but there's a caveat -- since `DynamicBucketingSampler` reads the manifests sequentially top to bottom, the setting of `shuffle_buffer_size` is going to matter, and the higher you set it, the...
> Piotr, I think I can think of a way to do this efficiently enough, even when using existing manifests. Suppose you have some cuts.jsonl.gz as your manifest for the...
Please see this PR, I added the CutSet constructor that'd work well with manifest shards: https://github.com/lhotse-speech/lhotse/pull/1085
My 2c, surely Dan, Fangjun, and others could tell you more: besides the need to re-write the kernels for metal, you'd have to somehow work around the design based on...
Looks like a bug in Lhotse, will fix. You can probably solve this by setting env var LHOTSE_DILL_ENABLED=1 or using the cuts = cuts.reverb_rir() API.
Was it faster without RIR or MUSAN? What’s the number of data loading workers and max duration?
One clean solution could be to add a method such as `.downcast(self, dtype=...)` to most modules that casts all parameters and buffers to a correct precision for a given dtype....