search

1ytic / warp-rnnt

CUDA-Warp RNN-Transducer
MIT License
211 stars 41 forks source link

question about the gather arguments #25

Closed wl-junlin closed 2 years ago

wl-junlin commented 2 years ago

core_gather.cu is a memory-efficient version that expects log_probs with the shape (N, T, U, 2) only for blank and labels values. It shows excellent performance with a large vocabulary.

Hello, what dose the gather mean? Or when should I set gather to True? Does the log_probs shape (N, T, U, 2) means only have two classes, blank and whatever labels?

1ytic commented 2 years ago

This flag decreases the memory allocation for the array of gradients. You can look at the implementation. For example, if you have a large output vocabulary V, I highly recommend using this flag.