Fangjun Kuang
Fangjun Kuang
`ParallelRunner.NewStream()` currently always returns the default stream 0x0 as it uses the default argument `num_work_items == 0`. So in which case will we not use the default value?
The documentation at https://k2.readthedocs.io/en/latest/ has not been updated for a long time and it's quite outdated.
Enable packing from https://github.com/Tencent/ncnn/wiki/element-packing It was disabled since ncnn did not support LSTM with projection. I have updated http://github.com/csukuangfj/ncnn to use the LSTM with projection from ncnn. For a test...
Training command: ```bash export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7" ./pruned_transducer_stateless7/train.py \ --world-size 8 \ --num-epochs 20 \ --use-fp16 1 \ --max-duration 250 \ --exp-dir pruned_transducer_stateless7/exp \ --feedforward-dims "512,512,1024,1024,512" \ --master-port 12535 \ --training-subset L...
In the last few days, a few people (>3) have encountered the following error after updating their lhotse to the latest version: ``` Segmentation fault (core dump) ``` when they...
https://github.com/k2-fsa/icefall/issues/626#issuecomment-1293840023 The above error can be easily caught by CI tests. We should add tests for every released model mentioned in `RESULTS.md`. https://github.com/k2-fsa/icefall/tree/master/.github/scripts is a good start.
# WER comparison | | test-clean| test-other| comment| |---|---|---|---| |basline| 2.78 | 7.36 | --iter468000 --avg 16, greedy search| |this PR|3.36|8.20| --iter 468000 --avg 16, greedy search| It performs better...
Will post the results soon.
Combines #560 and #558