TensorFlow

TensorFlow

Reference

• https://developers.googleblog.com/2017/02/announcing-tensorflow-10.html
• Blog
• https://github.com/tensorflow/tensorflow/releases/tag/v1.0.0
• Source Code
• https://www.tensorflow.org/
• Official Site
• https://hunkim.github.io/ml/
• 기본적인 머신러닝과 딥러닝 강의
• Exercises
• https://penglover.github.io/categories/ML/TensorFlow
• 간단한 한글 설명(MNIST).
• http://pythonkim.tistory.com
• 간단한 예제들.
• http://pythonkim.tistory.com/38: tensorboard 설명
• http://pythonkim.tistory.com/56: ConvNet 설명
• https://github.com/aymericdamien
• 좋은 예제 및 참고 코드들.
• https://goo.gl/ka4Pn6
• Mnist 관련 수행 코드, 우분투에서 tensorflow/GPU 설정.
• http://goodtogreate.tistory.com/entry/CNN-for-CIFAR10
• Tutorial
• CS 20SI: Tensorflow for Deep Learning Research
• Stanford 강의. 
• 이외에도, CS229(ML), CS231n(NN) 강의 추천

Versions

Tensorflow 1.3

• As of Aug. 2017, Tensorflow 1.3 was released.

• [CPU] $ pip3 install --upgrade tensorflow
• [GPU] $ pip3 install --upgrade tensorflow-gpu

Tensorflow 1.0

• As of Feb. 15, 2017, Tensorflow 1.0 was announced.
• Python API stability
• Speedup on multiple GPUs
• tf.Keras
• Installation improvements
• [Experimental] XLA (Accelerated Linear Algebra) is a domain-specific compiler for linear algebra that optimizes TensorFlow computations. 
• XLA currently supports JIT compilation on x86-64 and NVIDIA GPUs; and AOT compilation for x86-64 and ARM.

TFlearn guide

• https://chatbotslife.com/deep-learning-in-7-lines-of-code-7879a8ef8cfb?imm_mid=0f1262

TensorFlow Installation (GPU)

• 설치 환경(2017.05.30)
• Ubuntu 16.04; python3.5.2
• 공식 홈페이지: https://www.tensorflow.org/install/install_linux
• 현재 지원버전: CUDA 8.0, cuDNN 5.1
• 선택한 설치 방식: native pip
• 1) CUDA 라이브러리 설치(v8.0)
• https://developer.nvidia.com/cuda-downloads
• linux -> x86_64 -> Ubuntu -> 16.04 -> deb(network)
• 파일 다운로드 후 관련 명령어 수행
• 2) cuDNN 다운로드(v.5.1)
• https://developer.nvidia.com/cudnn
• Nvidia 개발자 계정 필요
• 다음 명령 실행
• $ tar xvzf cudnn-8.0-linux-x64-v5.1-ga.tgz
• $ sudo cp -P cuda/include/cudnn.h /usr/local/cuda/include
• $ sudo cp -P cuda/lib64/libcudnn* /usr/local/cuda/lib64
• $ sudo chmod a+r /usr/local/cuda/include/cudnn.h /usr/local/cuda/lib64/libcudnn*
• 3) 다음 명령 실행
• $ sudo apt-get install libcupti-dev
• 4) pip3 설치
• $ sudo apt-get install python3-pip python3-dev
• $pip3 -V
• 5) tensorflow-gpu 설치
• $ pip3 install tensorflow-gpu # Python 3.n; GPU support
• $ python3 -c 'import tensorflow as tf; print(tf.__version__)'
• 기타
• 실행시 SSE, AVX 등 warning 발생하는 것은 신경쓰지 않아도 됨. TF 소스 자체에서 build하면 해결된다고 하나, 굳이 놔두어도 상관없을 듯.
• 상기 설치 후 system restart 한 번 해주면, graphic driver 잘 잡힐 듯.
• IDE로는 py-charm 등 이용.
• Performance Guide
• NCHW image data format
• Winograd conv algorithm 설정(3x3 행렬연산을 위한 tensor 성능 개선)
• $ export TF_ENABLE_WINOGRAD_NONFUSED=1
• # Couldn't open CUDA library libcupti.so.8.0. LD_LIBRARY_PATH
• $ export LD_LIBRARY_PATH=/usr/local/cuda/extras/CUPTI/lib64:$LD_LIBRARY_PATH

Library

• tf.Transform
• ML 전처리를 위한 TensorFlow 라이브러리 (Beam Preprocessing).
• Users define a pipeline by composing modular Python functions, which tf.Transform then executes with Apache Beam, a framework for large-scale, efficient, distributed data processing. --> 그렇다면, 런타임(Serving)에 전처리를 위해서는 Apache Beam을 별도로 설치해야 하나? 
• tf.Transform allows users to compute summary statistics for their datasets.
• Edward
• 확률 모델링, 추론을 위한 라이브러리.