Lecture 8: Deep Learning Software

CS231n

Lecture 8: Deep Learning Software

CPU < GPU

CPU: Fewer cores, but each core is much faster and much more capable; great at sequential tasks
GPU: More cores, but each core is much slower and “dumber”; great for parallel tasks

Deep Learning Frameworks

computational graph的概念适合深度学习

TensorFlow

静态图
源于Theano
- 定义tf.placeholder() tf.get_variable()
- 可以手动计算、更新参数，但是最好使用提供的接口
- 使用tf.Session()
- 使用封装好的层tf.layers

keras是tensorflow的高层封装
tensorboard

PyTorch

动态图
源于Torch
类似tensorboard的Visdom

Caffe

比较

Compare	TensorFlow	PyTorch	Caffe
Pros	safe, wide use	best	production
Cons	not perfect	new,can be rough