Joan Puigcerver

I would argue that if you want absolutely deterministic results, use `cudnn.convert(model, nn)`. NVIDIA cuDNN is not deterministic, so I think it's better not to give "false expectations" to the...

Copy & paste from the cuDNN 5.1 user guide: ``` However, bit-wise reproducibility is not guaranteed across versions, as the implementation of a given routine may change. With the current...

Hi, Thanks for your interest. If you are able to compose HCLG statically, it is going to be faster than the dynamic composition, for sure. The latgen-lazylm-faster-mapped tool was created...

Hello all, Make sure that you have replaced the epsilons in the input of your LM fst (G.fst) with a disambiguation symbol (i.e. #0), as well as to have produce...

@shakingWaves @daanzu It is true that, in general, not all WFST admit determinization. However, if you are using traditional models for G (n-grams) and L, and you have introduced the...

Fixed! Thanks for finding the bug. Let us know if you have any further issue. 2018-03-30 3:08 GMT+02:00 Dayvid Welles : > Hi! I'm trying to execute the RIMES recipe...

On a second thought, it would be even better if the user just specified the number of MB. That way the memory required would only depend on the model.

Hi @WenmuZhou, Not exactly. You need to make sure that you place the tensors in the appropriate devices. Take a look at this piece of code, were I use both...

Hi Adam, Yes. The Python interface for OpenFst has been broken at least since version 1.7.1. Last time I checked, 1.7.9 was still broken. This library uses wraps that interface,...

Hi! I can't manage to compile OpenFst 1.7.7 (nor 1.7.6) on my machine, as soon as I figure out what's going on I'll update the project. Unfortunately, I am short...