Deep Learning Systems¶
Easy-to-use automatic differentiation tools (Keras, Tensorflow, PyTorch) have been the biggest cause of rapid development in DL
Why Learn?¶
Understand internals of deep learning systems helps
- build/improve deep learning systems
- so that you can use them better
- not too difficult
Frameworks¶
| Paradigm | ||
|---|---|---|
| PyTorch | Imperative | |
| chainer | Imperative | |
| Tensorflow 1.0 | Declarative | |
| Theano | Declarative | |
| Caffe | Forward & Backward layers | |
| Jax | ||
| mxnet |
| Paradigm | Advantage | Disadvantage |
|---|---|---|
| Imperative | Easy to debug Allow mixing of programming control flow and computational graph construction | Eager execution |
| Declarative | Lazy execution | Hard to debug |
| Forward & Backward layers |