Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/10362
This diff implements a manual export from PyText's CRF module to the caffe2 CRF layer.
Note that most of the changes in caffe2/python/crf.py are just formatting changes, the only relevant change is the new class CRFUtils.
Reviewed By: hikushalhere
Differential Revision: D9234126
fbshipit-source-id: 1a67d709034660e8b3d5ac840560b56de63e3f69
Summary:
This is a prototype for joint intents + slots modeling workflow, it has the following:
1- New data readers and data processors to process joint labels in parallel
2 - New JointNN model
3- New Fblearner workflow (jointnn) for joint modeling experimentations
This is still work in progress, sending the diff to start the discussion about the interface and what we need to support in our joint modeling efforts.
P.S. The number of lines in this diff is multiplied by 3 since caffe2 is mirrored in both fbandroid and fbobjc. I will highlight the most important parts so that people are not confused.
Differential Revision: D5725243
fbshipit-source-id: ecc5322f937ad0fddaf200a9e090b3573a69f994
Summary: Inegration of the CRF Layer in DeepText wordmodels + Implementing the viterbi decode operator in C++ instead of python so that the CRF models can be deployed in production.
Differential Revision: D4912196
fbshipit-source-id: 64f499a1bd47e811e7a96dde839904dcd05cacb3
Summary:
Uses the cudnnTransformTensor function. It works by shuffling the strides according to the transpose axis. Significant speedup over current GPU version .
+ moves the transpose test under utility_ops, because hypothesis_test is too big
Reviewed By: jamesr66a
Differential Revision: D4810993
fbshipit-source-id: 82577c4ced1389e70bd5992820ae4d8297a3817f
