zhanwenchen

Updated 4/11/2018

Here's my experience of installing the NVIDIA CUDA kit 9.0 on a fresh install of Ubuntu Desktop 16.04.4 LTS.

Table of Contents

• Install NVIDIA Graphics Driver via apt-get
• Install CUDA
• Install cuDNN

Table of contents generated with markdown-toc

zhanwenchen

'use strict';
//////////////////////////////////
// How to use?
// 1. Create `sequelize-schema-file-generator.js` in your app root
// 2. Make sure you've ran the `sequelize init` before (It should create `config`,`seeders`,`migrations` folders).
// 3. Update `DATABASE_DSN` below to match your connection string (works with any database adapter that Sequelize supports)
// 4. Run it with `node sequelize-schema-file-generator.js`
// 5. Review the generated migrations inside of the `migrations` folder.
//////////////////////////////////

zhanwenchen

import os
import tensorflow as tf

trained_checkpoint_prefix = 'checkpoints/dev'
export_dir = os.path.join('models', '0') # IMPORTANT: each model folder must be named '0', '1', ... Otherwise it will fail!

loaded_graph = tf.Graph()
with tf.Session(graph=loaded_graph) as sess:
    # Restore from checkpoint
    loader = tf.train.import_meta_graph(trained_checkpoint_prefix + '.meta')

zhanwenchen

"""
Adapted from https://github.com/Vetal1977/tf_serving_example/blob/master/svnh_semi_supervised_client.py
"""

# -*- coding: utf-8 -*-
import time

from argparse import ArgumentParser
import numpy as np

zhanwenchen

# -*- coding: utf-8 -*-
import os
import sys
import logging
from flask import Flask, request, jsonify
from flask_cors import CORS, cross_origin


from translate_client import Server

zhanwenchen

#One workaround is to create clone environment, and then remove original one:
#(remember about deactivating current environment with deactivate on Windows and source deactivate on macOS/Linux)

conda create --name new_name --clone old_name --offline #use --offline flag to disable the redownload of all your packages
conda remove --name old_name --all # or its alias: `conda env remove --name old_name`

#There are several drawbacks of this method:
# time consumed on copying environment's files,
# temporary double disk usage.

zhanwenchen

# https://github.com/HDFGroup/hdf5/blob/hdf5-1_13_1/release_docs/INSTALL_parallel
# https://docs.olcf.ornl.gov/software/python/parallel_h5py.html
# https://www.pism.io/docs/installation/parallel-io-libraries.html
# using ~/local/build/hdf5 as the build directory.
# Install HDF5 1.13.1 with parallel I/O in ~/local/hdf5,

version=1.13.1
prefix=$HOME/local/hdf5
build_dir=~/local/build/hdf5
hdf5_site=https://support.hdfgroup.org/ftp/HDF5/releases/hdf5-1.13

zhanwenchen

"""
Creates an HDF5 file with a single dataset of shape (channels, n),
filled with random numbers.

Writing to the different channels (rows) is parallelized using MPI.

Usage:
  mpirun -np 8 python demo.py

Small shell script to run timings with different numbers of MPI processes:

zhanwenchen

I wrote these instructions as part of "installing PyTorch with CUDA 12.1.1".

Anyway, if you still need to compile from source… here's how:

Build Magma from source

This is a dependency of PyTorch, which is sensitive to CUDA version.

Clone Magma:

zhanwenchen

git clone https://git.videolan.org/git/ffmpeg/nv-codec-headers.git
cd nv-codec-headers
vi Makefile   # change the first line to PREFIX = ${CONDA_PREFIX}
make install
cd ..

git clone https://git.ffmpeg.org/ffmpeg.git
cd ffmpeg
git checkout n4.2.2
conda install nasm