Comparison Between Deep Learning Packages Yi Wu 9/22/2018
Outlines Background Caffe Theano TensorFlow Work Preliminary Results 9/22/2018
Background 9/22/2018
Background 9/22/2018
Background 9/22/2018
Caffe From UC Berkeley Written in C++ Python bindings 9/22/2018
Caffe Network structures are stored in .proto files 9/22/2018 http://caffe.berkeleyvision.org/tutorial/net_layer_blob.html
Caffe Model Zoo https://github.com/BVLC/caffe/wiki/Model-Zoo 9/22/2018
Caffe (+) Train models without writing any code (+) Lots of pre-trained models, good for testing and fine-tuning (+) Python interface (-) Cumbersome for big networks (-) No good for recurrent networks 9/22/2018
Theano From Yoshua Bengio’s group at University of Montreal symbolic computation compatible with Python 9/22/2018
Theano gradient on anything http://www.marekrei.com/blog/theano-tutorial/ 9/22/2018
Theano black box magic http://www.marekrei.com/blog/theano-tutorial/ 9/22/2018
Theano (+) Python + numpy (+) computation graph (+) RNN fits nicely in computation graph (-) Hard to manually optimize 9/22/2018
TensorFlow From Google Computation graphs Scalable 9/22/2018
TensorFlow X h h2 y w_h2 w_h w_0 https://github.com/nlintz/TensorFlow-Tutorials/blob/master/04_modern_net.py 9/22/2018
TensorFlow X h h2 y w_h2 w_h w_0 define graph nodes https://github.com/nlintz/TensorFlow-Tutorials/blob/master/04_modern_net.py 9/22/2018
TensorFlow X h h2 y w_h2 w_h w_0 define graph edges https://github.com/nlintz/TensorFlow-Tutorials/blob/master/04_modern_net.py 9/22/2018
TensorFlow Tensorboard http://thenewstack.io/look-inside-tensorflow-googles-open-source-deep-learning-framework/ 9/22/2018
Work Image classification (CNN) data: PASCAL 2012 model: VGG16 21 classes 11540 images (random 1000) model: VGG16 9/22/2018
Work Hardware data: 1000 images test for speed CPU: I7 4GHz GPU: GTX 1080 RAM: 32G data: 1000 images test for speed CPU: Caffe vs Theano vs TensorFlow GPU: Theano vs TensorFlow test for easiness to use installation code test & debug visualization 9/22/2018
Preliminary Result 9/22/2018
Preliminary Result training with GPU predicted class takes 2 sec 9/22/2018