Anton Bakhtin

I've finally added information about finally of different models. Check out https://github.com/yandex/faster-rnnlm#experiments

The model is downloaded from benchmark's repo. I used quite straightforward interpolation, i.e. mean per-word probability, as it's describe in the benchmarks' paper.

Hi! 1. -direct 1000 requires 1000 \* 10000000 \* 4 (float) ~= 4GB memory on your GPU. As a result maxent weights fail to be copied to your GPU. A...

That's weird. Does the rnnlm actually learn anything in less then a second? Does valid entropy decrease?

Yeap, that's normal. Validation for Hierarchical Softmax is a few orders faster than for HS. The problem with NCE is that nobody guarantees that predicted probabilites would be stochastic, i.e....

NCE validation uses only simple operations (like matrix multiplication) and could be efficiently implemented on GPU. As for training, some operations works faster on GPU (matrix multiplication) and some operations...

Hi! There is not flag to use text format. All the information about the network is stored in NNet class. Once it is constructed, you can dump it in a...

In kaldi faster-rnnln is used for rescoring only. For instance, checkout librispeech recipe: https://github.com/kaldi-asr/kaldi/blob/master/egs/librispeech/s5/run.sh#L333 On Fri, Dec 25, 2015 at 4:10 AM, xiaofeixia1234 [email protected] wrote: > Merry Christmas! > I...

Which compile version to you use? Can you try g++-4.8?

Is this fixed?