Seyyed Hossein Hasanpour Coderx7

General guidelines for CPU performance on PyTorch

This file serves a BKM to get better performance on CPU for PyTorch, mostly focusing on inference or deployment. Chinese version available here.

1. Use channels last memory format

Right now, on PyTorch CPU path, you may choose to use 3 types of memory formats.

torch.contiguous_format: default memory format, also referred as NHCW.
torch.channels_last: also referred as NHWC.
torch._mkldnn: mkldnn blocked format.

How to install PyTorch v0.3.1 on RaspberryPi 3B

Prerequisites

Important I'd recommend you use at least a 16GB sd card. I tried it with an 8GB card and it baaaaarely fits. I had to uninstall a lot of packages and regularly clean up.

Go to https://www.raspberrypi.org/downloads/raspbian/ and download the Raspbian Stretch image (either one).

Use Etcher (from https://etcher.io/) to "burn" the image onto the SD card.

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.

A reason to use the integrated graphics for display is if installing the NVIDIA drivers causes the display to stop working properly. In my case, Ubuntu would get stuck in a login loop after installing the NVIDIA drivers. This happened regardless if I installed the drivers from the "Additional Drivers" tab in "System Settings" or the ppa:graphics-drivers/ppa in the command-line.

In this article, I will share some of my experience on installing NVIDIA driver and CUDA on Linux OS. Here I mainly use Ubuntu as example. Comments for CentOS/Fedora are also provided as much as I can.

	/*
	What's the fastest way to read a file?

	fast_read is a program that reads a file into memory using multiple parallel threads
	with a stochastic hill climb optimizer to find the fastest thread count and block size.

	Example run (note the effect of page cache, read is basically memcpy after that point):
	$ ./fast_read iotest_6G.dat
	Opening file iotest_6G.dat for reading
	Reading 6442450944 bytes

	import torch, torch.nn as nn, torch.nn.functional as F
	import numpy as np
	import torch.optim as optim

	# tied autoencoder using off the shelf nn modules
	class TiedAutoEncoderOffTheShelf(nn.Module):
	def __init__(self, inp, out, weight):
	super().__init__()
	self.encoder = nn.Linear(inp, out, bias=False)
	self.decoder = nn.Linear(out, inp, bias=False)

	I've tested it on Fedora 23 and Ubuntu 16.04. I'm using gcc-5.3.1, python-3.4, VS Code-1.14.0
	You can debug mixed Python/C++ in the same GUI. It also works for MPI applications. You can switch between the debuggers and corresponding call stacks.

	1. Packages needed
	1) Visual Studio Code

	2) Extensions for VS Code:
	"Python" from Don Jayamanne (I'm using 0.6.7)
	This allows VS Code act as the front end to debug python.
	This gives VS Code ability to attach to a python script that uses module "ptvsd".

	using System;

	namespace Algorithms
	{
	public static class LevenshteinDistance
	{
	/// <summary>
	/// Calculate the difference between 2 strings using the Levenshtein distance algorithm
	/// </summary>
	/// <param name="source1">First string</param>