Deep Learning with Python

✍ Scribed by William Edison

• Publisher: William Edison
• Year: 2020
• Tongue: English
• Leaves: 148
• Category: Library

✦ Synopsis

Learn to create inventive programs on your Machine Learning&Deep Learning and Python―with no programming experience required. Discover how to configure, write Python scripts, create user-friendly GUIs.Projects include a object detection by find object with camera, tracking motion. Hands-on Deep Learning&Machine Learning and Object, face/yes detection, recognize handwriting, pedstrian detecting.

• Computer vision and application fields
• Setup Python, OpenCV, Numpy Packages for Visual Studio IDE on windows
• Machine Learning and Object Detection with 1k-Means, 2k-NN, SVM and HOG, Cascade Classifier
• Make Classifier for Object Detection Casscade Classify Tool, Collecting Data, Training Object to Classify, Labeling, Training LBG, Detection Object
• Train Object Detection Classifier Using TensorFlow, Installing Anaconda, CUDA, and cuDNN, Setting Anaconda Virtual Environment, Gathering and labeling pictures, Generating training data, Creating a label map and configuring training, Training, Exporting the inference graph, Testing object detection classifier

✦ Table of Contents

1.Computer Vision
1.1 Computer Vision
1.2 Application Fields
2.Introduction and Setup
2.1Setup Visual Studio IDE On Windows
2.2Install Python Packages
2.3Install OpenCV Packages
2.4Install NumPy Packages
3.Machine Learning and Object Detection
3.1k-Means
3.2k-NN
3.3SVM and HOG
3.4Cascade Classifier
4.Make Classify for Object Detection
4.1Install Casscade Classify Tool
4.2Collecting Data
4.3Training Object to Classify
4.4Labeling
4.5Training LBG
4.6Detection Object
5.Train Object Detection Classifier Using TensorFlow
5.1
5.2
5.3
5.4
5.5
5.6
5.7
5.8
Field of Computer Vision  
Image Processing
Process of getting better quality video using comp
Print new video after receiving video
Image Enhancement
Image Restoration
Image Compositing
Image Segmentation
Computer Vision(CV)
Get meaningful info. from video
Print data from received video
Object Detection
Object Recognition
Object Tracking
Computer Graphics(CG)
Generating video not exiting
Print new data from received data
Image Processing Case
Improve Image
Image Processing Case
Improve Image
Restore Image
Image Processing Case
Compose Images
Divide Image
Computer Vision Case
Detect Objects
Recognize Objects
Computer Vision Case
Trace Objects
Fields
Mobile
Camera filter, image messenger
Video, Broadcasting
Warping, Morphing, Face Swapping,
Chroma key, Caption
Medical
Ultra Sound, X-Ray, CT, MRI Analysis Diagnose
Sports
Posture analysis, Video Reading
Automobile
Around View, Autonomous Cars
Security, Safety, Traffics
Intruder alarm, trace suspect, protect, figure out
Industry
Process inspection, defect inspection, sense mall-
Robotics
Recognize objects and faces
Mobile
Snap Chat
Mobile
Snow Camera
Movie, Film
Warping
Morphing
Movie, Film
Face Swapping
Chroma key
Movie, Film
Chroma key
Subtitles
Medical diagnosis
Sports
Car
Around View
Autonomous Driving Cars
Security, Traffic
Security, Traffic
Industry
Process Management
Industry
Quality, Defect Inspection
Robotics
 Install Visual Studio 2019 IDE

Visual Studio is an integrated development environ
Language Support
Language : Python, C++/C#...
Support Platform : Windows
Developer: Microsoft
 Pre Required Visual Studio 2019
 Setup Visual Studio 2019 On Windows
Down Load Visual Studio Community 2019
Install the Visual Studio installer
Setup Visual Studio 2019 On Windows
Choose workloads and components
 select the Python development workload and choose
Choose individual components (Optional)
Install language packs (Optional)
Select the installation location (Optional)
Greate A New Python Project
Open Visual Studio and create a new project
Open Visual Studio and create a new project
Select the Python Application and create
Visual Studio show the project structure in the so
(1) The name of the project
(2) The solution is container for one or more rel
(3) The source files .py
(4) The Python Environments Python interpreters,
Setup Opencv-contrib-python
Pre-built OpenCV packages for Python
Select the packages for your enviroment
Select the packages for your enviroment
Run pip install opencv-contrib-python
Opencv-contrib-python Installation - PyPi
Check out Installed Version
import cv2
cv2.version
Test Code Show Picture
Installing Numpy
Pre-built Numpy packages for Python
Run pip install numpy
Numpy
http://www.nu

Fundamental package for scientific computing with
Strong N-th array object
Precise broadcasting function
C/C++, Fortran Combined tool
Useful linear algebra, Fourier transform, random n
OpenCV-Python ver 2.3+
Previously it used self internal data structure
Now it uses NumPy as image data structure
Change to cv2 package
import cv2
import numpy as np
NumPy and Image
ndarray
ndim: n dimension (axes)
shape: Size of each dimension, tuple
size : # of all item, Multiple of shape
dtype : data type of item
itemsize : Bite size of each item
data : Buffer having item, Not use in real
Numpy
Producing
Produce as value
 array()
Generate as initial value
empty(), zeros(), ones(), full()
Generate as exiting array  
empty_like(), zeros_like(), ones_like(), full_like
Generate as consecutive value
arange()
Generate as random number
random.rand(), random.randn()
Generate as value
numpy.array(list [, dtype]) : Generate NumPy array
list: Python list objects with value to be used in
dtype: Data type, if omitted, it is auto decided b
int8, int16, int32, int64 : Integer with sign
uint8, uint16, uint32, uint64 : Integer without si
float16, float32, float64, float128 : Real-number
complex64, complex128, complex256 : Complex number
bool : Boolian
Generate as value
Numpy
Generate as size and initial value
np.empty ( tuple [, dtype])
All factors not initialized, tuple size
np.empty( (2,3))
np.empty( (2,3), dtype=np.int16)
np.zeros( tuple [, dtype])
All factors are 0, tuple size
np.zeros( (3,4))
np.zeros( (2,3,4) , dtype=np.int16)
np.ones ( tuple [, dtype])
All factors are 1, tuple size
np.ones( (3,4) )
np.ones( (2,3,4) , dtype=np.int16)
np.full(shape, fill_value [, dtype])
Generate array initialized with fill_value
Generate as size
Numpy
Generate as exiting array
np.empty_like(array [, dtype] )
Generate non-initialized dtype and same shape arra
zeros_like(array [, dtype])
 Generate 0(zero) initialized dtype and same shape
ones_like(array [, dtype])
Generate 1 initialized dtype and same shape array
ones_like(array, fill_value [, dtype])
Generate fill_value initialized dtype and same sha
Generate as exiting array
Generate as Sequence and Random Number
numpy.arange([start=0, ] stop [, step=1, dtype=flo
start : start value
stop : stop value, Number included in the scope ar
step : Increasing value
dtype : data type
numpy.random.rand([d0 [, d1 [..., dn]]]) :
d0, d1..dn : shape, if omitted, return a random nu
Random number b/w 0 and 1
numpy.random.randn([d0 [, d1 [..., dn]]]) :
Mean=0, Distribution=1, Random number following st
Numpy
Generate as Sequence and Random Number
Change Data Type
a.astype(‘typename’) , a.astype(np.dtype)
np.uintXX(), np.intXX(), np.floatXX(), np.complexX
Numpy
Change Dimension
np.reshape(array, newshape [, order])
arr.reshape(newshape [, order])
array : Array the objects to rearrange
newshape :  Size of new array, tuple
-1 : The column size is not decided
order : {‘C’, ‘F’, ‘A’}  C:C-like, F:Fotran-like,
np.ravel(array)
Change to 1 dimension
arr.T : Transposed array
Change Dimension
Numpy
Broadcasting Calculation
Scaler and Calculation
Array and calculation, two arrays should have same
Matrix
Matrix Addition : +
Matrix Multiply: np.dot(a, b), a.dot(b)
Transposed Matrix: a.T , a.transpose()
Numpy
Indexing/Slicing
1 Dimension
a[0] : Certain Parameter
a[1 : 5] : 1~4 Parameter
a[5: ] : 5 to end
a[:5] : Beginning to 4
a[ : ] : All parameter
2 Dimension
a[2, 3 ] : 2 column 3 row
a[2:4,  3] : 2~3 column 3 row
a[ 2, 1:4] : 2 column 1~3 row
a[2, :] :2 column and all row
a[:, 2] : All column and 2 row
a[1:3, : ] : 1~2 column and all row
Indexing/Slicing
Numpy
Fancy Indexing
Generate data fit to condition or initialize
Condition calculation:  > , >=, <, <=, ==, !=
Bool type (True/False)
a[a > 4]
a[a > 4] = 0
Fency Indexing
Merge
numpy.hstack(arrays) : Merge arrays horizontally  
numpy.vstack(arrays) : Merge arrays vertically  
numpy.concatenate(arrays, axis=0) : Merge arrays a
numpy.stack( arrays, axis=0) : Merge arrays as new
arrays : Merge objects, tuple
Numpy
Merge
Numpy
splitting
np.hsplit(array, indice) = np.split(axis=1)
np.vsplit(array, indice) = np.split(axis=0)
np.split(array, indice, axis=0)
Numpy Search
numpy.where(condition [, t, f]) : Finding items fi
return : Tuple, array filled with index or changed
condition : Condition that will be used in search
t, f : Value or array set according to condition,
t : Value or array fit to condition
f : Value or array unfit to condition
numpy.nonzero(array) : Turnaround index that is no
numpy.all(array [, axis]) : Search true for all ar
array : Array of search subject
axis : Axis to search, if omitted, search all axis
numpy.any(array [, axis]) : Search for true in any
Numpy
Search
Numpy
Search
Numpy
Basic Statistics Function
numpy.sum(array [, axis]) : add array
numpy.mean(array [, axis]) : Average of array
numpy.amin(array [, axis]) : Min. value of array
numpy.min(array [, axis]) : Same as numpy.amin()
numpy.amax(array [, axis]) : Max value of array
numpy.max(array [, axis]) : Same as numpy.amax()
array : Arry of calculation objects
axis : Base axis to calculate, if it is not define
Numpy
Creating Gray Image
n x m
n : height, rows
m : width, columns
value : 0~255 (dtype=np.uint8)
Creating Color Image
n x m x 3
n : height, rows
m : width, columns
3 : color channels, Default = [Blue, Green, Red]
value : 0~255 (dtype=np.uint8)
Numpy Image
AI
In 1956 Dartmouth University in USA, Prof. John Mc
General AI: AI thinking like human
Narrow AI: AI only performing certain function i.e
Machine Learning
Analyze data with algorithms, learn the analyzed d
Various Algorithm
Deep Learning
One algorithm of machine learning
Artificial neural network: it is a network inspire
In 2012, Prof. Andrew NG of Google and Stanford re
Machine Learning
Through preconditioning process of data, transfer
Work analyzing characteristics of various algorith
Of the data set, 90% is for training set and 10% f
Supervised Learning
Learning with data with label (answer)
classification: name to label
Regression: numerical label
Linear regression: Score to study time (0~100)
H(x) = Wx + b, H: hypothesis, W: weight, b:bias
Binary classification: Pass/Fail to study time
Multi-labeled classification: GPA to study time (A
Unsupervised Learning
Learning without label
Clustering: Concern only in which data go to which
Realizing OpenCV
Legacy Routine: Realize even before the standardiz
core Module
k-Means, Mahalanobis
Standardized Interface
ml Module, Run StatModel Abstract Class Inheritanc
Normal Bayes(Normal/Pure size Bayes divider)
Decision trees
Boosting
Random tree
Expected Value - maximize
k-Nearest Neighbor
Mumltilayer Perceptron (MLP, Multilayer perceptron
Suport Vector Machine(SVM)
Objective Detector
Upper level class using learning method
CascadeClassifier, Latent SVM(DPMDetector), Bag of
K-Means Clustering
Clustered algorithm, non-guide learning
Cluster mixed data as wanted
Algorithm
Step1.
Set two center points(C1, C2) randomly.
Step2.
Calculate distance at the two center points.
If test data is close to C1, 0, close to C2, 1.
If center point is C3, C4, mark as  2,3.
Step3
Picture mark 0 data close to C1 as red, and 1 data
K-Means Clustering
Algorithm
Step4.
Average red and blue, set is as new center point a
Repeat step 2 to 4, until each center point is fix
Or, repeat it until it satisfies certain criteria;
Center point is dots with smallest value of distan
K-Means Clustering
cv2.kmeans(data, K, bestLabels, criteria, attempts
data : np.float32, should be saved at 1column
k : # of cluster wanted
bestLabels : Result data, None
criteria : termination criteria, tuple(type, max_i
attempts : # of run to be done through different i
flags : Selecting initial center points, 
cv2.KMEASNS_PP_CENTERS
cv2.KMEASNS_RANDOM_CENTERS
cv2.KMEANS_USE_INITIAL_LABELS 
Turnaround:
retval : Square of center points and each data dis
bestLabels : Labels, 0, 1, ...
centers : Array of cluster’s center point
Cluster of random number example
K-Means Clustering
Cluster of random number example
K-Means Clustering
사진속 영상 색상 군집화 Example
16 color clustering
k-NN(k-Nearest Neighbour)
If new point is entered into exiting points that i
Categorizing to the closest member
K is setting the scope of closest member
You can give different weight to close member and
knn = cv2.ml.KNearest_create()
knn.train(trainData, cv2.ml.ROW_SAMPLE, responses)
ret, results, neighbours ,dist = knn.findNearest(n
k-NN(k-Nearest Neighbour)
Simple Example
Simple Example
k-NN(k-Nearest Neighbour)
Move Categorizing Example
k-NN(k-Nearest Neighbour)
Move Categorizing Example
Recognize Handwriting
5000 letters of handwritten images
500 letters per number (0 ~ 9)
20 x 20=400 pixels by each letter
5000 X 400 Vector
4500 : train set
500 : test set
k-NN(k-Nearest Neighbour)
Recognize Handwriting Training
k-NN(k-Nearest Neighbour)
Recognize Handwriting Training
Recognize Handwriting  <1/3>
k-NN(k-Nearest Neighbour)
Recognize Handwriting  <2/3>
Recognize Handwriting  
Support Vector Machine
One of the classifications
k-NN needs to calculate all data distance, it take
Distinguish parameters, find lines and recognizing
SVM,NuSVM, LinearSVM (Specialized in linear kernel
Lines that can be categorized into two is called D
Find line first in all SVM  parameters because it
No need training data; if it is close to the count
Support Vectors: Blue circle filled and read quadr
Support Planes: Line that passes
Blue Circle: Wx + b   > 1
Red Data: Wx + b < -1
W : weight, vector decides orientation
b: bias, Decide location
Decision Boundary :
Crossing center point of two hyperplane
Wx +b = 0
SVM
Example <1/2>
SVM
Example
HOG
Histogram of Oriented Gradient Feature Descriptor
Histogram for edge slope orientation
Suggested to recognize pedestrian  
Normalized Histogram
hogDesc = cv2.HOGDescriptor(winSize,blockSize,bloc
windsize : Size of HOG extraction area
blockSize : Size of normalization area
blockStride : Size of normalization overlapping bl
cellSize : Size of HOG calculation
nbins : # of histogram level
hodDesc : HOG specialized narrator
Vector Size :
HOG SVM
MNIST Handwriting Study  <1/3>
HOG SVM
MNIST Handwriting Study <2/3>
HOG SVM
MNIST Handwriting Study
Recognize Handwriting  <1/3>
HOG SVM
Recognize Handwriting  <2/3>
HOG SVM
Recognize Handwriting  
Pedstrian Detecting
HOG training object to detect pedestrian
cv2.HOGDescriptor_getDefaultPeopleDetector()
cv2.HOGDescriptor_getDaimlerPeopleDetector()
cv2.HOGDescriptor.setSVMDetector()
cv2.HOGDescriptor.detectMultiScale
Detect Pedestrian
Pedstrian Detecting
Detect Pedestrian
Cascade
Tree base object divider
Effective in recognizing faces
Use Haar Feature while adding or subtracting certa
Offer trainer and detector
Detector: Recognize pre-trained object file
haarcascades.xml
Trainer: New image training
Cascade
classifier = cv2.CascadeClassifier('file.xml')
classifier.detectMultiScale(img, scaleFactor, minN
img : Input image
scaleFactor: Limit image enlarging size 1.3~1.5
Use for scale pyramid; if value is big, chances fo
minNeighbors: # of pixels to maintain
 If the number is big, quality gets better, but #
flags : For old API; not used now
minSize, maxSize: Ignore detection if it goes over
Cascade
Face Detection Example
Face Detection Example
Cascade
Face Detection Example
Face Detection Example
Face Recognition
Recognizing your own face
Take your own photo in multiple angles
Train with the pictures take
mode = cv2.face.LBPHFaceRecognizer_create()
model.train(train_data, label_data)
Save the training result in xml
mode.write(‘file.xml’)
Read the saves xml to read faces
mode = cv2.face.LBPHFaceRecognizer_create()
Mode.read(‘file.xml’)
result : model.predict(face)
result : (label, confidence)
label : Label of most similar photo to the registe
confidence : distance (0 : 100%,  ∞ : 0%)
Recognize Faces
Take Photo
Input name and ID (in number)
Take 100 photos
Including front angle, take the photos in variety
Various faces
Recognize Faces
Training
Recognize Faces
Result
Face Mosaic
Give auto mosaic effects to photos with human face
Result Example:
Hannibal Mask
Place Hannibal bask from the Silence of The Lambs
Result Example
Face Distortion
Make a camera that distorts human face like Snapch
Example:
Casscade Classify Tool

Official Support
Windows 7 or Above
Created by Amin Ahmadi
Cascade Trainer GUI is a program that can be used
Setup Cascade Trainer GUI

Download Cascade Trainer GUI 3.3.1 
Next for setup - Cascade Trainer GUI
Setup Cascade Trainer GUI
Select Components and Next
 Create Name
Setup Cascade Trainer GUI
Ready to Install
Installation Completed
Collecting Data
Take Photo
Take 100 photos Including positive images and nega
Including front angle, take the photos in variety
Various pinochle deck
Collecting Data
Take Video
Take video Including positive images and negative
Including front angle, take video in variety of an
Various pinochle deck
Training Casscade Classify
Create two folders inside it
“p” folder: Contain positive images samples the im
“n” folder: Contain negative images is the images
Image classify
Open casscade trainger tool
Training Casscade Classify
Image classify
Press tab Cropper
Select input settings from image in a folder or vi
Training Casscade Classify
Image classify
Press browse input folder path contain image
Press browse out put to folder path to “p” positiv
Training Casscade Classify
Image classify
Click left mouse to chose positive image and press
Continue press next to other positive image
Training Casscade Classify
Image classify
Press browse out put to folder path to “n” negativ
Click left mouse to chose negative image and press
Training Casscade Classify
Image classify
Crop 400 positive image in “p” positive folder
Crop 1000 nagative image in “n” negative folder
Training Casscade Classify
Training classifiers
Press to training tab
Browse folder you have created for the classifier
Training Casscade Classify
Training classifiers
Setting common
Pre-calculation buffer size is the speed of train
Setting Cascade
Training Casscade Classify
Setting common
Sample with make sure not to set it to a very big
Sample height
Setting feature type to HAAR or LBP
 HAAR: Classifiers are very accurate but require a
 LBP: Classifiers less accurate but train much qui
Setting boost
Training Casscade Classify
Training classifiers
Press start to training classifiers
Classifier folder
Stage#.xml: The files are temporary files that won
Params.xml: Reminder file contains the parameters
“neg.lst”, “pos.lst” and “pos_samples.vec” are tem
“cascade.xml” is the actual cascade classifier.
Training Casscade Classify
Test your classifiers
 Press browse button
 Browse to path to cascade XML file in the Cascade
 Single Image: Image will be used as the scene in
 Images in a Folder: A folder containing many imag
 Video: A video file will be used for testing the
 Setting output
 Images Folder: Detected objects will be saved to
 Video: A video with the detected objects with a
1.Install Casscade Classify Tool
2.Collecting Data
3.Training Casscade Classify
4.Labeling
5.Training LBG
6.Detection Object
Labeling
Create camera object detection automation
Base_dir: Path creat labeling file
Cascade classifier: Path of classifiercascade.xml
Target_cnt: Number of pinochle deck
Code auto pinochle deck detected example #1
Labeling
Code auto pinochle deck detected example #2
Labeling
Run python scrip code auto pinochle deck detected
Input ID and Name of pinochle deck
Input ID: #1
Name: Jacks
Labeling
Camera open and auto detected pinochle deck with I
File Jacks_1 had been auto great inside D:/ Labeli
400 Image Jacks with variety of angles
1.Install Casscade Classify Tool
2.Collecting Data
3.Training Casscade Classify
4.Labeling
5.Training LBG
6.Detection Object
Training LBG
Source code training LBG
Training LBG
Training LBG
Starting LBG model training
Training LBG
Model trained successfully
Auto created “.xml file” content in labeling file
1.Install Casscade Classify Tool
2.Collecting Data
3.Training Casscade Classify
4.Labeling
5.Training LBG
6.Object Recognize
Object Recognize
Source code for object recognize #1
Object Recognize
Source code for object recognize #2
Object Recognize
Run API
Result object recognize by camera
Train Object Detection Classifier Using TensorFlo
1.
2.
3.
4.
5.
6.
7.
8.
Installing Anaconda
Setup Anaconda on windows
Python 3.7
Download 64-Bit or 32-Bit graphical installer
Run the downloaded executable (.exe)
Click next to setup
Installing Anaconda
Setup Anaconda on windows
Read the licensing terms and click “I Agree”.
Select an install for “Just Me” unless you’re inst
Installing Anaconda
Setup Anaconda on windows
Select a destination folder to install Anaconda an
Adding Anaconda to the PATH environment variable
Installing Anaconda
Setup Anaconda on windows
Installing
Finish
Installing CUDA
Install Cuda Tool Kit
Download Version 8.0
Note: Currently version 9 is not supported by Tens
Setup CUDA Package
Installing CUDA
Install Cuda Tool Kit
Agree software license agreement
Visual studio integration
Installing CUDA
Install Cuda Tool Kit
Instaling
Set environment variables
Go to Start and Search "environment variables"
click the Environment Variables button
Click on the Path system variable and select edit
add the following paths
 C:\Program Files\NVIDIA GPU Computing Toolkit\CU
 C:\Program Files\NVIDIA GPU Computing Toolkit\CU
 C:\Program Files\NVIDIA GPU Computing Toolkit\CU
Installing CUDNN 6.0
Download Cudnn 6.0 for cuda version 8.0 library
go to 
Create a user profile if needed
Select CUDNN 6.0 for CUDA tool kit 8.
Extract the zip file contents and place them somew
Add a path to the bin folder. For example, mine is
 C:\cuda\bi
Update your GPU Driver
go to 
Select your GPU version to download
Install the driver
Train Object Detection Classifier Using TensorFlo
1.
2.
3.
4.
5.
6.
7.
8.
TensorFlow
https://www.tensorflow.org
TensorFlow is machine learning library
Official Support
Native : C/C++
Bindings: Python, Java
Support Platform : Windows, Linux, Mac, iOS, Andro
TensorFlow was developed by the 
Free 
Environment
Installing TensorFlow CPU
Create a new conda virtual environment
Open a new Anaconda/Command Prompt window
Create a new virtual environment called “tensorflo
Installing TensorFlow CPU
Create a new conda virtual environment
”y” for installed package
Activate the environment
Installing TensorFlow CPU
Create a new conda virtual environment
Install TensorFlow CPU for Python
Install TensorFlow V1.5 in this environment
Installing TensorFlow CPU
Create a new conda virtual environment
Test your installation
Installing TensorFlow GPU
TensorFlow GPU is slightly more involved than that
Need of installing the relevant Graphics and CUDE
Systems prerequisites
Install CUDA Tool Kit
Download Version 8 here:
Currently version 9 is not supported by Tensorflow
Set Your Environment Variables
Go to Start and Search "environment variables"
Click the Environment Variables button
Click on the Path system variable and select edit
Add the following path
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA
Download CUDNN 6.0
Go to 
Create a user profile if needed
Select CUDNN 6.0 for CUDA tool kit 8.
Extract the zip file contents and place them somew
Add a path to the bin folder. For example, mine is
C:\cuda\bin
Installing TensorFlow GPU
Update your GPU Driver
Go to 
Select your GPU version to download
Install the driver
Install Anaconda python 3.5
Go to 
Choose python 3.5 (64bit) and install
Check the "add python to your PATH" option during
Create a new Conda virtual environment
Open a new Anaconda/Command Prompt window
Type the following command:
The above will create a new virtual environment wi
Activate the newly created virtual environment by
Installing TensorFlow GPU
Install TensorFlow GPU for Python
Open a new Anaconda/Command Prompt window and acti
Once open, type the following on the command line:
Wait for the installation to finish
Installing TensorFlow GPU
Test your Installation
Open a new Anaconda/Command Prompt window and acti
Start a new Python interpreter session by running:
Once the interpreter opens up, type:
If the above code shows an error, then check to ma
Then run the following:
The ouput should be
TensorFlow Models Installation
Install Prerequisites:
Downloading the TensorFlow Models-Master
Create a new folder under a path of your choice an
To download the models you can simply download(
You should now have a single folder named models u
Downloading  Object Detection API Model tutorial's
Links download: (
TensorFlow Models Installation
Downloading Object Detection API Model tutorial's
Extract all the contents directly into the C:\tens
\object_detection folder should look like:
COCO API installation
Pre-trained classifiers with specific neural netwo
Download the Faster-RCNN-Inception-V2-COCO
Link download:
Extract the faster_rcnn_inception_v2_coco_2018_01_
Adding necessary Environment Variable
Created that points to the \models, \models\resear
Every time the "tensorflow1" virtual environment i
Run commands
Checking python path environment variable
Protobuf Installation/Compilation
The Tensorflow Object Detection API uses Protobufs
This should be done as follows:
Download the latest -win32.zip release (e.g. 
Create a folder in C:\Program Files and name it Go
Extract the contents of the downloaded -win32.zip
Add C:\Program Files\Google Protobuf\bin to your P
In a new Anaconda/Command Prompt, cd into TensorFl
Change directories to the \models\research directo
Run the compilation process:
This creates a name_pb2.py file from every name.pr
Protobuf Installation/Compilation
Install the object_detection Python package
Running the following from C:\tensorflow1\models\r
Test your Installation
Open a new Anaconda/Command Prompt window and acti
Cd into C:\tensorflow1\models\research\object_det
This should start a new jupyter notebook server on
Press button Cell and chose Run All
Test your Installation
Test TensorFlow setup to verify it works show the
If not work see the
Appendix: Common Errors
Object_detection/protos/.proto: No such file or di
Run cmd
Change your directory to the research folder you d
Downloaded:
Get the location of the protroc.exe file that you
Get protoc.exe 3.4 and run compilation
Appendix: Common Errors
Numpy Version
TensorFlow V1.5 work on numpy version V1.14.5
Update numpy version to V1.14.5
Train Object Detection Classifier Using TensorFl
1.
2.
3.
4.
5.
6.
7.
8.
Gathering and labeling pictures
Gather Pictures
Gathering and labeling pictures
Gather Pictures
LabelImg Installation
For Windows and Linux you can download the precomp
Links download:
Run labelImg.exe
LabelImg Installation
Point it to your \images\train directory, and then
LabelImg saves a .xml file containing the label da
Generate Training Data
The image .xml data will be used to create .csv fi
This creates a train_labels.csv and test_labels.cs
Label Map: generate_tfrecord.py
Label for training a classifier
Generate the TFRecord files by issuing these comma
Greate Label Map
Label Map: The label map tells the trainer what e
Configure Training
Configure training: Finally, the object detection
Navigate to C:\tensorflow1\models\research\object_
Make the following changes to the faster_rcnn_ince
Line 9. Change num_classes to the number of differ
Line 106. Change fine_tune_checkpoint to:fine_tune
Lines 123 and 125. In the train_input_reader secti
input_path : "C:/tensorflow1/models/research/objec
label_map_path: "C:/tensorflow1/models/research/ob
Line 130. Change num_examples to the number of ima
Lines 135 and 137. In the eval_input_reader sectio
input_path : "C:/tensorflow1/models/research/objec
label_map_path: "C:/tensorflow1/models/research/ob
Save the file after the changes have been made. Th
Run the Training
From the \object_detection directory
Command to begin training:
TensorFlow will initialize the training. The initi
Each step of training reports the loss. It will st
Note: The loss numbers will be different if a diff
Run the Training
You can view the progress of the training job by u
This will create a webpage on your local machine a
Eport Inference Graph
Now that training is complete, the last step is to
This creates a frozen_inference_graph.pb file in t
Use Your New Trained Object Detection Classifier
The object detection classifier is all ready to go
Before running the Python scripts, you need to mod
To test your object detector, move a picture of th
To run any of the scripts, type “idle” in the Anac
If everything is working properly, the object dete