TensorFlow - Car-Occupancy Detector/Classifier from Scratch
THIS is now Version 2 of this Solution:
Smart City How might we create smart solutions to cope with the increasing individual car traffic in our cities? https://github.com/iCounterBOX/Themes-from-INTEREST/wiki/Smart-City---How-might-we-create-smart-solutions-to-cope-with-the-increasing-individual-car-traffic-in-our-cities%3F
Many thanks to this guy ( edje ) / https://www.youtube.com/watch?v=Rgpfk6eYxJA and https://github.com/EdjeElectronics/TensorFlow-Object-Detection-API-Tutorial-Train-Multiple-Objects-Windows-10
Mainly i followed this Youtube video...but here and there i struggled with different Version conflicts on TF, CUDA, cuDNN, Path-conflicts ..etc. I am quite new on this Stuf - SO this "Experiance-BLOG" is also a view out from a Newbee perspective.
Main Goal:
Main target is still to detect "HOW many People are sitting in a Car?". Via WebCam i will get alerted if ONE or TWO Persons are detected.
Step by Step:
My Setup: PC: ACER NITRO 5 / Win 10 / 32 GB RAM / GeForce GTX 1660 Ti
I made a fresh Installation of Anaconda3 & CuDnn & CUDA / Gforce
Dependency-List: https://www.tensorflow.org/install/source_windows
I decided to setup.. tensorflow_gpu-1.12.0 + Python 3.6.5 + cuDNN 7 + CUDA9
Anaconda3 (https://www.anaconda.com/)
I use Anaconda and Spyder and the Anaconda-Prompt! My first try to setup ALL within Anaconda-GUI ( TF, Cuda,..) did not work – the different dependent Versions made conflicts.
At the end I setup “SOME” in anaconda and some within the environment-Prompt – figuring out the proper Versions was the biggest pain..
IN Anaconda-GUI i create the environment
tfgpu112
This env was created with Phython 3.6 ( choosen from DropDown )
CUDA-Toolkit
For some Tests i need VS2017 - this was installed on my PC before!
Mainly i followed Marc Jay https://www.youtube.com/watch?v=Ebo8BklTtmc).
..here are some remarks i noticed during this installation:
I take the Gforce-Experience
to setup and Install the Latest drivers (https://www.nvidia.com/content/DriverDownload-March2009/confirmation.php?url=/Windows/436.48/436.48-notebook-win10-64bit-international-whql.exe&lang=us&type=geforcem)
edje ( https://www.youtube.com/watch?v=Rgpfk6eYxJA ) is running and placing all the code for the detection under and into the folder structure from the Tensor-Flow-Installation!
I decided to store those Code into a different folder! This lets Tesorflow as it is on HardDrive. The experimental code is located in my project-Folder..e.g.:
Installations made in Anaconda-GUI ( environment / tfgpu112 ):
Opencv
matplotlib
pillow
numpy
Installations made in Anaconda-Prompt ( environment / tfgpu112 ):
(
Check python --version..if wrong then install this one
conda install python==3.6.5 ( überprüfen wichtig!)
)
conda install tensorflow-gpu==1.12.0
conda install spyder
conda install pytorch torchvision -c pytorchFrom this guy ( thx 2 H. Singh / https://towardsdatascience.com/tensorflow-gpu-installation-made-easy-use-conda-instead-of-pip-52e5249374bc) we get the hint to use as much as possible from the anaconda-gui..The idea is that A-GUI is checking the heavy Version-Dependencies for us..In some case a nice thing and i use as much as possible. BUT in some case we need specific versions.In that case i had good success with those CUDA setup in A-Prompt.
also in his blog - a nice quick test if Tensorflow is basically working together with my GPU:
import tensorflow as tf
sess = tf.Session(config=tf.ConfigProto(log_device_placement=True))My biggest pain and reason why this training-Pipeline did not work was based on non-compatible Versions of the installed apps we need here! This way i wrote this little Version-Checker:
import platform
import numpy as np
import tensorflow as tf
import cv2 # opencv
import torch
import sys
sys.path.append("C:\\appl\\TensorFlow\\models\\research\\")
sys.path.append("C:\\appl\\TensorFlow\\models\\research\\object_detection\\utils")
print("V E R S I O - I N F O:")
print("OpenCV Version: {}".format(cv2.__version__))
print("Python Version: " + platform.python_version())
print("Numpy Version: " + np.__version__)
print("TensorFlow Version: " + tf.__version__)
'''
OK ABER der unten ist besser!!
#https://discuss.pytorch.org/t/pytorch-and-cuda-9-1/13126/9
print("The CUDA-VERSION: " + torch.version.cuda)
if tf.test.gpu_device_name():
print("Default GPU Device: {}".format(tf.test.gpu_device_name()))
else:
print("Please install GPU version of TF")
print("\nVerify GPU & Tensorflow: https://www.codingforentrepreneurs.com/blog/install-tensorflow-gpu-windows-cuda-cudnn \n")
from tensorflow.python.client import device_lib
print(device_lib.list_local_devices())
'''
print("\nVerify GPU & cuDNN: https://gist.github.com/wassname/34626c2d31e28ffc864fc4f3027c4489 ")
import datetime
print(datetime.datetime.utcnow().strftime('%Y%m%d_%H-%M-%S'))
print(torch._C._cudnn_version(), 'cudnn')
print(torch._C._cuda_getDriverVersion(), 'cuda driver')
print(torch._C._cuda_getCompiledVersion(), 'cuda compiled version')
#print(torch._C._nccl_version(), 'nccl')
for i in range(torch.cuda.device_count()):
print('device %s:'%i, torch.cuda.get_device_properties(i))After following the videos, blogs and hints THIS now the configuration which finally let the training-Session RUN ( from the Version-Checker above ):
Python 3.6.5 |Anaconda, Inc.| (default, Mar 29 2018, 13:32:41) [MSC v.1900 64 bit (AMD64)]
Type "copyright", "credits" or "license" for more information.
IPython 7.8.0 -- An enhanced Interactive Python.
runfile('D:/ALL_DEVEL_VIP/ALL_python/edje_objDetClassifier/getSomeVersionInfos_V1.py', wdir='D:/ALL_DEVEL_VIP/ALL_python/edje_objDetClassifier')
V E R S I O N - I N F O:
OpenCV Version: 3.3.1
Python Version: 3.6.5
Numpy Version: 1.16.5
TensorFlow Version: 1.12.0
Verify GPU & cuDNN: https://gist.github.com/wassname/34626c2d31e28ffc864fc4f3027c4489
20191021_14-15-11
7005 cudnn
10010 cuda driver
9000 cuda compiled version
device 0: _CudaDeviceProperties(name='GeForce GTX 1660 Ti', major=7, minor=5, total_memory=6144MB, multi_processor_count=24)After all this setup and specific adaptions i described ( and still following the steps from mentioned Video/Edje ) the training is running:
python train.py --logtostderr --train_dir=training/ --pipeline_config_path=training/faster_rcnn_inception_v2_pets.configwe are now able to train our Model to detect how many Persons are sitting in a car ( e.g.).
ONE:
TWO-Person:
Preselecting CAR´s from a Video-File to make this "labelImg" a bit easier
Preparing and labeling the Training-Images is hard work. But as CarOccupency needs CARS we can use pretrained Models (TF) to select those cars from a Video and store the single CAR-Pictures in a folder. The Separation in the two other Folders ( ONE and TWO ) i made manually after running this script: a_carOccupancy_PreSelector_Cars_V2.py - it produces single Car-Pictures and store them in the folder ./detectedImages/
IMPORTANT: this script is using visualization_utils_ck4CarsSinglePic.py from the TF Utils folder. Its a modified copy of visualization_utils.py!!
(..you need to copy this then from the utils-folder here into the utils folder from your TF-installation!!)
the parts where i made adaptions is labeled with "ckoss"...
if WriteSinglePic2File:
cv2.imshow('Write PIC 2 File', cv2.resize(crop_img, (960, 720))) # ckoss
global _cCnt #https://stackoverflow.com/questions/46018872/accessing-variables-defined-in-enclosing-scope
_cCnt += 1
newImgPath = './detectedImages/car' + str(_cCnt) + '.jpg'
#print('new filename: ' + newImgPath)
cv2.imwrite(newImgPath,cv2.resize(crop_img, (960, 720))) # <<<<---- WRITE THE SINGLE PIC TO folderAfter this step i used Labelimg (https://github.com/tzutalin/labelImg)..
the HotKey´s i like:
Create a rect box ( w ) / Next Image ( d ) / prev Image ( a )
I first wanted to know if this training and detection is basically working - And still have to experiment here with the resultion and stuff of the extracted Pictures out from the CAR-Video ( Any Hint highly welcome !!) - So i currently ONLY have 30 Pics!!
Nevertheless - Training was running fine and this frozen inference graph was build ( video from edje !!)
python export_inference_graph.py --input_type image_tensor --pipeline_config_path training/faster_rcnn_inception_v2_pets.config --trained_checkpoint_prefix training/model.ckpt-XXXX --output_directory inference_graphSome attention on the folders we need for deployment
Sometimes the deployment may break or is running on err during training etc etc..some of the folders need to get cleaned or sub folder deleted..
for new deployment delete subfolder saved_model
Training folder only need labelmap.pbtxt and faster_rcnn_inception_v2_pets.config for a fresh training... a training may fail if there are checkpoints stored from an other train-run. In some case we need old checkpoints if we wand to continue a training. Folder after a successful training:
Current issue:
the current version is only working with edje_objDetClassifierCarOccu\Object_detection_webcam.py - So i run the Car-Video on my PC and via webcam (a cam with wire so i can view my PC-Screen!! ) we get the detected ONE/TWO - frames - FINE! Basically working but we will need hundreds of those windshield-pics ( i have 30 ! ).
RESUME
Its working - Training from ZERO, Predictions are possible...For me as newbie on MachineLearning a MileStone :-) But shure THIS is now just the beginning..Lots of the Libs, Classes etc are still BlackBoxes for me..and i am still on "hallo World" level with this.
LINK´s ( CREDITS / So Many thanks to the makers who shared their know how with us ):
GURU / edje: https://www.youtube.com/watch?v=Rgpfk6eYxJA
https://www.youtube.com/watch?v=bYqvx_DM45U
https://www.youtube.com/watch?v=nZUxoHPFf4w
https://www.tensorflow.org/install/source_windows
Marc Jay
https://www.youtube.com/watch?v=Ebo8BklTtmc).
This .md is written with Markdown-Monster https://weblog.west-wind.com/posts/2016/Nov/04/Introducing-Markdown-Monster-a-new-Markdown-Editor