Dear all,
Last year Kyle told us about sklearn-theano which can be used to bridge between scikit-learn and Theano. Now, he's ready to tell us how we can even bridge Theano with Caffe (!) In other words, we can use GoogLeNet easily together with Theano now. Isn't it awesome? Furthermore, it's going to be the thirteenth this Friday (huh?)
The talk will be on Friday at 13.30 in the usual place AA3195. See below for the details.
- Cho
===
Speaker: Kyle Kaster
Date/Time: 13 Apr (Friday) @13.30
Place: AA3195
Title: Theano + Caffe, with a side of GoogLeNet, in sklearn-theano
Abstract:
applying pretrained neural networks to data using python, numpy and
theano. sklearn-theano strives to conform to the scikit-learn
interface, and can be flexibly added into machine learning pipelines
as an upfront transform followed by standard classifiers and
transformers from scikit-learn. Many of the models in sklearn-theano
also work as standalone classifiers and object localizers.
We have recently added support for a potentially huge number of
pretrained models by adding a bridge between Theano and Caffe. This
talk will go over some interesting notes about the architecture of
GoogLeNet itself, the key things necessary to bring *new* models into
sklearn-theano (or theano in general), and some discussion about where
we plan to go next.