Open syoamakase opened 4 years ago
I confirmed that the performances of Tedlium2 were good in PyTorch's version 1.3.0
and 1.5.1
when I used both single gpu (batch size = 30) and multi gpus (batch size = 40).
I consider the PyTorch version is relevant to this bad results.
I highly recommend to avoid pytorch==1.4.0
!!!!
In some examples, we cannot confirm the same results of README.md. I think that
batch size
,pytorch version
, andsingle or multi gpu(s)
are relevant to the reproductivity.Tedlium 2:
batch_size = 30
,pytorch==1.4.0
, andsingle gpu
: bad results (WER is about 40%)batch_size = 40
,pytorch==1.4.0
, andmulti gpus
: extremely bad results (the same outputs)Please report if you cannot confirm the reproductivity.