Keras, PyTorch

Keras

• (2017.09.03) Keras shoot-out: TensorFlow vs MXNet
• Keras를 TensorFlow 또는 MXNet 연동하고 각각의 성능을 간단하게 비교 
• API on top of either TensorFlow or Theano
• Primary maintainer is François Chollet, a Google engineer.
• Focused only on deep learning algorithms

1 2 3 4 5 6 7 8

import keras.layers as L import keras.models as M my_input = L.Input(shape=(100,)) intermediate = L.Dense(10, activation='relu')(my_input) my_output = L.Dense(1, activation='softmax')(intermediate) model = M.Model(input=my_input, output=my_output)

cs

PyTorch

• Facebook
• just-in-time graph compilation
• It doesn’t treat graphs as separate and opaque objects. Instead, you can assemble tensor computations ad hoc in very flexible ways.
• 즉, 동적인 연산 그래프(dynamic computation graphs, DCG)를 이용함.
• c.f. Theano 및 TensorFlow는 정적인 연산 그래프를 이용.
• Multi-GPU support, though Tensorflow still wins for larger distributed systems.

Lasagna

Kur

• Deep Learning API 
• supports Theano, TensorFlow, PyTorch
• supports multi-GPU