INSTALL.md

Install Guide

Pre Requiresite

• A Windows 10 Pro PC
• Preferably a dedicated NVIDIA GPU
• Administrator access to download and install

0. Software Requirement

1. Cmder (Download the Full Version)
2. Visual Studio Code
3. Visual Studio 2015/2017 Visual C++ Build Tool (Follow the install process, The total install is about 1.5gigs)

1. Python

1. Download Anaconda Python 3.6 and run installation with default settings. Note the installation path which will be used in steps 2 and 3
2. Add the installation directory path to your Path (either user or system variables) in the Environment Variable settings in System Properties. For us it was C:\Users\{yourUserName}\AppData\Local\Continuum\anaconda3
3. Also add the .\anaconda3\Scripts to your Path (either user or system variables) in the Environment Variable settings in System Properties. For us it was C:\Users\{yourUserName}\AppData\Local\Continuum\anaconda3\Scripts
4. To test python, run python in your cmder, and you should see Python 3.6.x |Anaconda, Inc. ... and a REPL prompt.
5. Type exit() to exit the REPL.
6. To test pip, run pip -V and you should see pip 10.0.x ...
7. If you have pip 9.0.x run python -m pip install -U pip to update to the latest version.

2. CUDA (if you have a compatible NVIDIA GPU)

0. What GPU do I have?
1. Go to CUDA 9.0 Download - Follow the link to download CUDA 9.0, the latest version of CUDA (9.2) is not supported for Tensorflow yet. See GH Issue
2. Pick the right configuration for your machine
3. Download the Base Installer (Express option is ok)
4. Download Patch 2 (Released Mar 5, 2018) (Express option is ok)
5. Download cuDNN for CUDA 9.0 Note: You will need to create a membership account for download access
6. Follow the instruction here Note: Complete 4.2 to 4.3, skip 4.3.5 (adding to VS Project)

3. TensorFlow

1. With Python, and CUDA/cuDNN installed, run pip install --upgrade tensorflow-gpu to install the GPU Version of Tensorflow.
2. Check that your machine has successfully installed tensorflow. In cmder run python to open the python repl.
3. Inside the python repl run import tensorflow (you may see a float error, this is expected)
4. Run print(tensorflow.__version__) and you should see 1.8.0
5. Run exit() to exit the python repl

tensorflow-gpu is now downloaded and installed successfully. Note: The following output during installation is okay

notebook 5.4.0 requires ipykernel, which is not installed.
jupyter 1.0.0 requires ipykernel, which is not installed.
jupyter-console 5.2.0 requires ipykernel, which is not installed.
ipywidgets 7.1.1 requires ipykernel>=4.5.1, which is not installed.

4. Install OpenCV for Python

1. Run pip install --upgrade opencv-contrib-python
2. Test the installation by running the python REPL (python >>> import cv2 >>> print(cv2.__version__)). You should see 3.4.1

5. Install Object Detection API for Tensorflow

The Object Detection API provides established models optimized for image/video analysis. Developers can modify and train these models to fit their specific needs.

1. Anaconda should come with these packages but verify they are installed. To do this run:

   pip install Cython
   pip install pillow
   pip install lxml
   pip install jupyter
   pip install matplotlib
   

2. Create a project folder (i.e C:\Users\{YourUserName}\projects\obj-det-proj)
3. Go to https://github.com/tensorflow/models and clone this repository (~500mb) into your project folder (i.e C:\Users\{YourUserName}\projects\obj-det-proj\models)

ProtoBuf

1. Download the protobuf win-zip 3.4.0 file here and extract it in a protoc folder in your project folder (i.e C:\Users\{YourUserName}\projects\obj-det-proj\protoc)
2. Go back to {project}/models/research, and run "C:\Users\{YourUserName}\projects\obj-det-proj\protoc\bin\protoc" object_detection/protos/*.proto --python_out=.

COCO API Install

1. Go back to your project folder (ie C:\Users\{YourUserName}\projects\obj-det-proj), we'll denote this path as {project} in this doc going forward.
2. Clone the Coco API repo by running git clone https://github.com/cocodataset/cocoapi.git
3. Go into {project}/cocoapi/PythonAPI, and open up setup.py in VS Code.
4. Remove the arguments in the extra_compile_args array (line 12), so the line should look like extra_compile_args=[], Save, and exit.
5. In the {project}/cocoapi/PythonAPI directory, run python setup.py build_ext --inplace
6. You should see Finished generating code /n copying build\lib.win-amd64-3.6\pycocotools\_mask.cp36-win_amd64.pyd -> pycocotools
7. In the {project}/cocoapi/PythonAPI directory, run rm -rf build to remove the build directory
8. In the {project}/cocoapi/PythonAPI directory, Run cp -r pycocotools ../../models/research/.

Note: folder cocoapi and folder models should both be in the same directory

Add PYTHONPATH environement variable

1. Go to environement variable settings and create a new variable named PYTHONPATH with three values:

• the path of your python.exe
• the path of your {project}/models/research
• path of {project}/models/research/slim.

Concate the three paths with semi-colons without spaces in between.

For example ours looks like:

C:\Users\{YourUserName}\AppData\Local\Continuum\anaconda3;
C:\Users\{YourUserName}\projects\obj-det-proj\models\research;
C:\Users\{YourUserName}\projects\obj-det-proj\models\research\slim

Restart your cmder (exit and open cmder back up), for the environment variable to take effect.

running echo %PYTHONPATH% should print something similar: C:\Users\{YourUserName}\AppData\Local\Continuum\anaconda3;C:\Users\{YourUserName}\projects\obj-det-proj\models\research;C:\Users\{YourUserName}\projects\obj-det-proj\models\research\slim

Test Object Detection API

1. Go to {project}/models/research
2. Run python object_detection/builders/model_builder_test.py
3. You should see

...............
----------------------------------------------------------------------
Ran 15 tests in 0.184s

OK

Note: If you see some error with FutureWarning and float types, it is ok.