diff options
author | nicholas-leonard <nick@nikopia.org> | 2014-09-03 22:49:56 +0400 |
---|---|---|
committer | nicholas-leonard <nick@nikopia.org> | 2014-09-03 22:49:56 +0400 |
commit | 5f2c6f4b906427ac4db3aaa53ac0a145bcad2d59 (patch) | |
tree | 9c94b446fac67d6510b7136a6540bfdfb4035ba9 /README.md | |
parent | a94f630694ddb5e19fec2331e8cb2ad66eab04bd (diff) |
library documentation++
Diffstat (limited to 'README.md')
-rw-r--r-- | README.md | 13 |
1 files changed, 10 insertions, 3 deletions
@@ -53,15 +53,22 @@ This section includes documentation for the following objects: <a name='nnx.SoftMaxTree'/> ### SoftMaxTree ### +A hierarchy of parameterized log-softmaxes. Used for computing the likelihood of a leaf class. +This Module should be used with the [TreeNLLCriterion](#nnx.TreeNLLCriterion). +Requires a Tensor mapping one `parent_id` to many `child_id`. +Greatly accelerates learning and testing for language models with large vocabularies. +A vocabulary hierarchy is provided via the [dp](https://github.com/nicholas-leonard/dp/blob/master/README.md) package's +[BillionWords](https://github.com/nicholas-leonard/dp/blob/master/doc/data.md#dp.BillionWords) +[DataSource](https://github.com/nicholas-leonard/dp/blob/master/doc/data.md#dp.DataSource). + Computes the log of a product of softmaxes in a path. Returns an output tensor of size 1D. -Only works with a tree (one parent per child). <a name='nnx.TreeNLLCriterion''/> ### TreeNLLCriterion ### -Measures the Negative Log Likelihood (NLL) for SoftMaxTrees. +Measures the Negative Log Likelihood (NLL) for [SoftMaxTrees](#nnx.SoftMaxTree). Used for maximizing the likelihood of SoftMaxTree outputs. -SoftMaxTree outputs a column tensor representing the log likelihood +The SoftMaxTree Module outputs a column Tensor representing the log likelihood of each target in the batch. Thus SoftMaxTree requires the targets. So this Criterion only computes the negative of those outputs, as well as its corresponding gradients. |