A (No-Cross) Compile of PyTorch and OpenCV using AWS A1 instances

a1_gist.md

This gist is a collection of bash scripts and a Jupyter noteookbooks to support the blog entitled:

"A (No-Cross) Compile of PyTorch and OpenCV using AWS A1 instances"

opencv_install.sh

#!/bin/bash

# References
# https://docs.opencv.org/master/d2/de6/tutorial_py_setup_in_ubuntu.html

set -x
set -e

if [ "$1" == "a1" ]; then
    threads=16
elif [ "$1" == "u96" ]; then
    threads=4
elif [ "$1" == "install" ]; then
	make install
	ldconfig
else
    echo '$0 <a1|u96|install>'
    exit 0
fi 


# if needed
# sudo apt-get update
# sudo apt-get -y install build-essential cmake zip python3-opencv python3-dev python3-numpy

opencv_version=3.4.3
wget -O opencv.zip https://github.com/opencv/opencv/archive/${opencv_version}.zip  
wget -O opencv_contrib.zip https://github.com/opencv/opencv_contrib/archive/${opencv_version}.zip

unzip opencv.zip 
unzip opencv_contrib.zip

mkdir opencv-${opencv_version}/build
cd opencv-${opencv_version}/build

cmake -D CMAKE_BUILD_TYPE=RELEASE \
	-D CMAKE_INSTALL_PREFIX=/usr/local \
	-D BUILD_WITH_DEBUG_INFO=OFF \
	-D BUILD_DOCS=OFF \
	-D BUILD_EXAMPLES=OFF \
	-D BUILD_TESTS=ON \
	-D BUILD_opencv_ts=OFF \
	-D BUILD_PERF_TESTS=OFF \
	-D INSTALL_C_EXAMPLES=OFF \
	-D INSTALL_PYTHON_EXAMPLES=OFF \
	-D ENABLE_NEON=ON \
	-D WITH_LIBV4L=ON \
	-D WITH_GSTREAMER=ON \
	-D BUILD_opencv_dnn=OFF \
	-D OPENCV_EXTRA_MODULES_PATH=../../opencv_contrib-${opencv_version}/modules \
        ../

make -j$threads
cd ../..

pytorch_build.sh

#!/bin/bash

# Following a nice blog here:
# https://medium.com/hardware-interfacing/how-to-install-pytorch-v4-0-on-raspberry-pi-3b-odroids-and-other-arm-based-devices-91d62f2933c7

set -x
set -e

if [ "$1" == "a1" ]; then
    threads=16
	sudo apt-get update
    sudo apt-get -y install build-essential libopenblas-dev libblas-dev m4 cmake cython python3-dev python3-yaml python3-setuptools python3-pip 
    pip3 install wheel

elif [ "$1" == "u96" ]; then
    threads=4
	sudo apt-get update
	sudo apt-get install libopenblas-dev
else
	echo "$0 <a1|u96>"
	exit 0
fi 
mkdir pytorch_install && cd pytorch_install
git clone --recursive https://github.com/pytorch/pytorch
cd pytorch

export NO_CUDA=1
export NO_DISTRIBUTED=1
export NO_MKLDNN=1 
export NO_NNPACK=1
export NO_QNNPACK=1
export MAX_JOBS=$threads

python3 setup.py bdist_wheel

pytorch_test_on_u96.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## PyTorch on PYNQ v2.5 and the Ultra96v1\n",
    "\n",
    "This notebook is a simple smoke test for a PyTorch installation running on the PYNQ v2.5 image.  PyTorch cannot build on the off-the-shelf PYNQ image due to the small core/memory count on the Ultra96.  Therefore I used a AWS A1 instance for building the pytorch bdist_wheel.  \n",
    "\n",
    "This notebook is part of a PYNQ blog describing how to use the A1's to help build big-softare for embedded platforms like the Ultra96.  I am only running two simple examples below to show the installation working on the board.  \n",
    "\n",
    "- example #1 from [here](https://medium.com/hardware-interfacing/how-to-install-pytorch-v4-0-on-raspberry-pi-3b-odroids-and-other-arm-based-devices-91d62f2933c7)\n",
    "- examples #2 from [here](https://dev.to/nestedsoftware/pytorch-hello-world-37mo)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "torch version: 1.4.0a0+b2b1601\n",
      "\n"
     ]
    }
   ],
   "source": [
    "# sanity check on import from the A1 built, Ultra96 installed torch package\n",
    "# I built from the HEAD of the pytorch repo, hence the git hash.\n",
    "import torch\n",
    "print(f'torch version: {torch.__version__}\\n')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "tensor([[0.1611, 0.4185, 0.2366],\n",
      "        [0.0886, 0.0956, 0.1774],\n",
      "        [0.7228, 0.3696, 0.9871],\n",
      "        [0.9236, 0.4115, 0.4540],\n",
      "        [0.7930, 0.3345, 0.5961]])\n"
     ]
    }
   ],
   "source": [
    "# example 1 - simple tensor build \n",
    "a = torch.rand(5,3)\n",
    "print (a)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "network topology: Net(\n",
      "  (hidden_layer): Linear(in_features=1, out_features=1, bias=True)\n",
      "  (output_layer): Linear(in_features=1, out_features=1, bias=True)\n",
      ")\n",
      "w_l1 = 1.58\n",
      "b_l1 = -0.14\n",
      "w_l2 = 2.45\n",
      "b_l2 = -0.11\n",
      "a_l2 = 0.8506\n",
      "updated_w_l1 = 1.5814\n",
      "updated_b_l1 = -0.1383\n",
      "updated_w_l2 = 2.4529\n",
      "updated_b_l2 = -0.1062\n",
      "updated_a_l2 = 0.8515\n"
     ]
    }
   ],
   "source": [
    "# example 2 - taking an optimization step\n",
    "import torch.nn as nn\n",
    "import torch.optim as optim\n",
    "\n",
    "\n",
    "class Net(nn.Module):\n",
    "    def __init__(self):\n",
    "        super(Net, self).__init__()\n",
    "        self.hidden_layer = nn.Linear(1, 1)\n",
    "        self.hidden_layer.weight = torch.nn.Parameter(torch.tensor([[1.58]]))\n",
    "        self.hidden_layer.bias = torch.nn.Parameter(torch.tensor([-0.14]))\n",
    "\n",
    "        self.output_layer = nn.Linear(1, 1)\n",
    "        self.output_layer.weight = torch.nn.Parameter(torch.tensor([[2.45]]))\n",
    "        self.output_layer.bias = torch.nn.Parameter(torch.tensor([-0.11]))\n",
    "\n",
    "    def forward(self, x):\n",
    "        x = torch.sigmoid(self.hidden_layer(x))\n",
    "        x = torch.sigmoid(self.output_layer(x))\n",
    "        return x\n",
    "\n",
    "\n",
    "net = Net()\n",
    "print(f\"network topology: {net}\")\n",
    "\n",
    "print(f\"w_l1 = {round(net.hidden_layer.weight.item(), 4)}\")\n",
    "print(f\"b_l1 = {round(net.hidden_layer.bias.item(), 4)}\")\n",
    "print(f\"w_l2 = {round(net.output_layer.weight.item(), 4)}\")\n",
    "print(f\"b_l2 = {round(net.output_layer.bias.item(), 4)}\")\n",
    "\n",
    "# run input data forward through network\n",
    "input_data = torch.tensor([0.8])\n",
    "output = net(input_data)\n",
    "print(f\"a_l2 = {round(output.item(), 4)}\")\n",
    "\n",
    "# backpropagate gradient\n",
    "target = torch.tensor([1.])\n",
    "criterion = nn.MSELoss()\n",
    "loss = criterion(output, target)\n",
    "net.zero_grad()\n",
    "loss.backward()\n",
    "\n",
    "# update weights and biases\n",
    "optimizer = optim.SGD(net.parameters(), lr=0.1)\n",
    "optimizer.step()\n",
    "\n",
    "print(f\"updated_w_l1 = {round(net.hidden_layer.weight.item(), 4)}\")\n",
    "print(f\"updated_b_l1 = {round(net.hidden_layer.bias.item(), 4)}\")\n",
    "print(f\"updated_w_l2 = {round(net.output_layer.weight.item(), 4)}\")\n",
    "print(f\"updated_b_l2 = {round(net.output_layer.bias.item(), 4)}\")\n",
    "\n",
    "output = net(input_data)\n",
    "print(f\"updated_a_l2 = {round(output.item(), 4)}\")"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.6.5"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}