Age | Commit message (Collapse) | Author |
|
|
|
|
|
|
|
|
|
mostly useful to make model capable of scoring lower-order n-grams:
use dropout (p=0.95) during training, and pad the n-gram to score
with 0 embeddings (the embedding at null_index will be set to 0).
this is an alternative to obtaining the weighted average of all
input embeddings for padding, as done by (Vaswani et al. 2013)
|
|
|
|
c++11
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
no relative paths to Eigen
|
|
that here so that it only relies on the Makefile
|
|
|
|
|
|
|
|
|
|
files for training
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
hidden layer)
|
|
|
|
|
|
|
|
(set num_hidden to 0; num_output_embeddings will be size of remaining layer)
|
|
existing model.
(use cases: continue training with more epochs, or save memory by using different subsets of data for each epoch)
|
|
system for it.
This reverts commit 52861f20f427ffa403966583ab2b637fd264228c.
|
|
nplm's 3rdparty directory
|
|
osx compile
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|