Original text version of tutorial you can visit here.
The purpose of this tutorial is to learn how to install and prepare TensorFlow framework to train your own convolutional neural network object detection classifier for multiple objects, starting from scratch. At the end of this tutorial, you will have basics and a program that can identify and draw boxes around specific objects in computer screen.
There are several good tutorials available for how to use TensorFlow’s Object Detection API to train a classifier for a single or several objects, but they are not that detailed how we want they to be. More over Object Detection API seems to have been developed on a Linux-based OS. To set up TensorFlow to train and use a model on Windows, there are several workarounds that need to be used in place of commands that would work fine on Linux. Also, this tutorial provides instructions for training a classifier that can detect multiple objects, not just one.
This tutorial is written and tested for Windows 10 operating system, it should also work for Windows 7 and 8, but haven’t tested that, so I am not sure. The general procedure can also be used for Linux operating systems, but file paths and package installation commands should be changed accordingly.
TensorFlow-GPU allows your PC to use the video card to provide extra processing power while training, so it will be used for this tutorial. In my experience, using TensorFlow-GPU instead of regular TensorFlow reduces training time by a factor of about 8 (depends on used CPU and trained model).
Regular CPU version TensorFlow can also be used for this tutorial, but it will take longer and real time models may work slower. If you use regular TensorFlow, you do not need to install CUDA and cuDNN in installation step. I used newest TensorFlow-GPU v1.11 while creating this tutorial, but it also should work for future versions of TensorFlow, but I am not guaranteed.
Vision of this tutorial: to create TensorFlow object detection model, that could detect CS:GO players. In this tutorial we’ll be more focused on detecting terrorists and counter terrorists bodies and heads, then automatically aim to them and shoot. At first, it seems like an easy task, but going through this tutorial series you will see that there is more problems than it seems.
Main computer specs for this tutorial:
- OS – Windows 10 Pro
- GPU - GTX1080TI
- CPU – i5-6400
- RAM – 16GB
Hardware requirements:
- NVIDIA® GPU card with CUDA® Compute Capability 3.5 or higher
- Strong enough computer (high end CPU and at least 8GB of RAM)
Software requirements:
- 64-bit Python v3.5+ for windows.
- A supported version of Microsoft Windows.
- A supported version of Microsoft Visual Studio. Visual Studio 2017
- CUDA 9.0
- cuDNN v7.3.1 for Windows 10 and CUDA 9.0
In many places there was said that there is some problems while working on newest CUDA versions, but I took this challenge and installed CUDA v10.0 and cuDNN v7.3.1. As future versions of TensorFlow will be released, you will likely need to continue updating the CUDA and cuDNN versions to the latest supported version. If you face problems while installing CUDA, visit this documentation site. If you face problems while installing cuDNN, visit this documentation site. This tutorial is made for TensorFlow-GPU v1.11, so the “pip install tensorflow-gpu” command should automatically download and install newest 1.11 version.
If you don’t know how to install python, follow my tutorial on PyLessons page. 32-bit version doesn’t support TensorFlow so don’t even try. If you already have python installed, install required packages:
pip install pillow
pip install lxml
pip install Cython
pip install jupyter
pip install matplotlib
pip install pandas
pip install opencv-python
pip install tensorflow-gpu
Download the full TensorFlow object detection repository located at this link by clicking the “Clone or Download” button and downloading the zip file. Open the downloaded zip file and extract the “models-master” folder directly into the C:\ directory. Rename “models-master” to “TensorFlow”.
This working directory will contain the full TensorFlow object detection framework, as well as your training images, training data, trained classifier, configuration files, and everything else needed for the object detection classifier.
You can follow my tutorial with installation and repository preparation or follow original TensorFlow tutorial (Must mention it didn’t worked for me with official tutorial).
When you already have TensorFlow models on your disk, you must add object detection directories to python path (if it doesn’t work from CMD line, do it manually like I did on video tutorial):
Configure PYTHONPATH environment variable (in my case):
set PYTHONPATH=$PYTHONPATH;C:\TensorFlow \research
set PYTHONPATH=$PYTHONPATH;C:\TensorFlow \research\slim
Next, compile the Protobuf files, which are used by TensorFlow to configure model and training parameters. Unfortunately, the short protoc compilation command posted on TensorFlow’s Object Detection API installation page doesn’t work for me on Windows. So I downloaded older 3.4.0 version of protoc from here. Transfered files to same TensorFlow directory and ran the following command from the tensorflow /research/ directory:
# From tensorflow/research/
"C:/TensorFlow/bin/protoc" object_detection/protos/*.proto --python_out=.
Finally, run the following command from the C:\ TensorFlow\research directory:
python setup.py install
You can test that you have correctly installed the Tensorflow Object Detection API by running the following command:
# From tensorflow/research/object_detection
python builders/model_builder_test.py
If everything was fine you can test it out and verify your installation is working by launching the object_detection_tutorial.ipynb script with Jupyter. From the object_detection directory, issue this command:
# From tensorflow/research/object_detection
jupyter notebook
In your browser click on „object_detection_tutorial.ipynb“. Then navigate to „Cell“ in navigation bar and click on „Run All“. If everything is fine in short time you should see these nice photos:
Original text version of tutorial you can visit here.
In this part we are going to merge jupyter API code from 1-st tutorial with code from 2-nd tutorial where we tested 3 different ways of grabbing screen.
To begin, we're going to modify the notebook first by converting it to a .py file. If you want to keep it in a notebook, that's fine too. To convert, you can go to file > download as > python file. Once that's done, we're going to comment out the lines we don't need.
Once you have your converted object detection file, go to your TensorFlow installation folder: research\object_detection\data and grab mscoco_label_map.pbtxt file, place it to you working directory.
Next you should download pretrained model from here, I am using faster_rcnn_inception_v2_coco, so I recommend you to use the same, at least at the beginning. Take frozen_inference_graph.pb file and transfer it to your local working repository.
So we begin by importing time, CV2, MSS libraries. If you don't have them, install before moving forward.
I personally imported line to disable CUDA devices, because I wanted to run this example on CPU, because running tensorflow-gpu takes more time to startup in backend.
os.environ['CUDA_VISIBLE_DEVICES'] = '-1'
Then we don't need to import tarfile and import zipfile, because we are not working with these files, so we comment them out now.
Going further we comment #from matplotlib import pyplot as plt and #from PIL import Image lines, because we are doing things our way.
Next I am importing few lines from my second tutorial for grabing screen and measuring FPS:
# title of our window
title = "FPS benchmark"
# set start time to current time
start_time = time.time()
# displays the frame rate every 2 second
display_time = 2
# Set primarry FPS to 0
fps = 0
# Load mss library as sct
sct = mss.mss()
# Set monitor size to capture to MSS
monitor = {"top": 40, "left": 0, "width": 800, "height": 640}
Because we are not using notebook anymore we are not using and these lines:
#sys.path.append("..")
#if StrictVersion(tf.__version__) < StrictVersion('1.9.0'):
# raise ImportError('Please upgrade your TensorFlow installation to v1.9.* or later!')
We are not using matplotlib to display image, so we are commenting line used for that:
#get_ipython().run_line_magic('matplotlib', 'inline')
There are two lines of import before going to an actual code:
from utils import label_map_util
from utils import visualization_utils as vis_util
But if you will try to use them like this, you will get an error, so add object_detection. before utils, just like this:
from object_detection.utils import label_map_util
from object_detection.utils import visualization_utils as vis_util
Next is links to paths, if you would like to have everything in same folder, just like in my tutorial, comment all these lines:
# What model to download.
MODEL_NAME = 'ssd_mobilenet_v1_coco_2017_11_17'
MODEL_FILE = MODEL_NAME + '.tar.gz'
DOWNLOAD_BASE = 'http://download.tensorflow.org/models/object_detection/'
# Path to frozen detection graph. This is the actual model that is used for the object detection.
PATH_TO_FROZEN_GRAPH = MODEL_NAME + '/frozen_inference_graph.pb'
# List of the strings that is used to add correct label for each box.
PATH_TO_LABELS = os.path.join('data', 'mscoco_label_map.pbtxt')
And replace them with my used path lines (don't forget to add NUM_CLASSES = 99 line)
MODEL_NAME = 'inference_graph'
PATH_TO_FROZEN_GRAPH = 'frozen_inference_graph.pb'
PATH_TO_LABELS = 'mscoco_label_map.pbtxt'
NUM_CLASSES = 99
Next you can comment all [6] part, because we won't use it:
#opener = urllib.request.URLopener()
#opener.retrieve(DOWNLOAD_BASE + MODEL_FILE, MODEL_FILE)
#tar_file = tarfile.open(MODEL_FILE)
#for file in tar_file.getmembers():
# file_name = os.path.basename(file.name)
# if 'frozen_inference_graph.pb' in file_name:
# tar_file.extract(file, os.getcwd())
Next, add 3: label_map, categories and category_index lines before detection_graph code:
label_map = label_map_util.load_labelmap(PATH_TO_LABELS)
categories = label_map_util.convert_label_map_to_categories(label_map, max_num_classes=NUM_CLASSES, use_display_name=True)
category_index = label_map_util.create_category_index(categories)
and in part [8] comment or delete category_index line:
category_index = label_map_util.create_category_index_from_labelmap(PATH_TO_LABELS, use_display_name=True)
Comment all lines in part [7] where images were loaded to numpy array:
#def load_image_into_numpy_array(image):
# (im_width, im_height) = image.size
# return np.array(image.getdata()).reshape(
# (im_height, im_width, 3)).astype(np.uint8)
Next you can delete PATH_TO_TEST_IMAGES_DIR, TEST_IMAGE_PATHS and IMAGE_SIZE lines, but if you will leave them it wound effect our code:
PATH_TO_TEST_IMAGES_DIR = 'test_images'
TEST_IMAGE_PATHS = [ os.path.join(PATH_TO_TEST_IMAGES_DIR, 'image{}.jpg'.format(i)) for i in range(1, 3) ]
IMAGE_SIZE = (12, 8)
At the end of code, not to make any mistakes you can replace all [12] block code with this code:
with detection_graph.as_default():
with tf.Session(graph=detection_graph) as sess:
while True:
# Get raw pixels from the screen, save it to a Numpy array
image_np = np.array(sct.grab(monitor))
# to ger real color we do this:
image_np = cv2.cvtColor(image_np, cv2.COLOR_BGR2RGB)
#image = Image.open(image_path)
# the array based representation of the image will be used later in order to prepare the
# result image with boxes and labels on it.
#image_np = load_image_into_numpy_array(image)
# Expand dimensions since the model expects images to have shape: [1, None, None, 3]
image_np_expanded = np.expand_dims(image_np, axis=0)
# Actual detection.
output_dict = run_inference_for_single_image(image_np, detection_graph)
# Visualization of the results of a detection.
vis_util.visualize_boxes_and_labels_on_image_array(
image_np,
output_dict['detection_boxes'],
output_dict['detection_classes'],
output_dict['detection_scores'],
category_index,
instance_masks=output_dict.get('detection_masks'),
use_normalized_coordinates=True,
line_thickness=8)
#plt.figure(figsize=IMAGE_SIZE)
#plt.imshow(image_np)
cv2.imshow(title, cv2.cvtColor(image_np, cv2.COLOR_BGR2RGB))
fps+=1
TIME = time.time() - start_time
if (TIME) >= display_time :
print("FPS: ", fps / (TIME))
fps = 0
start_time = time.time()
# Press "q" to quit
if cv2.waitKey(25) & 0xFF == ord("q"):
cv2.destroyAllWindows()
break
So I tried to use this slow object detection method on image where you can see crowd of people walking across the street:
And here is the results of frames per second working with TensorFlow CPU version. In average, it is about 7 seconds to receive one frame per second. So if we would like to use it for real time purpose, this would be impossible to do something useful with it. So we need to make it work much faster, we are doing so in second part below.
In previous part we ran actual pretrained object detection, but our code is messy and detection was working really slow. In this part we are cleaning the messy code and making some code modifications that our object detection would work in much faster way.
At first I went through all code and deleted all unecassary code, so instead of using object_detection_tutorial_grabscreen.py, better take object_detection_tutorial_grabscreen_pretty.py it will be much easier to understand how it works. All code is in 3-4 part folder.
After cleaning the code, I started to make some changes to it. Mainly what I done is that I deleted def run_inference_for_single_image(image, graph): function and added needed lines to main while loop, and this changed object detection speed. Not taking into details I will upload code part I changed:
# # Detection
with detection_graph.as_default():
with tf.Session(graph=detection_graph) as sess:
while True:
# Get raw pixels from the screen, save it to a Numpy array
image_np = np.array(sct.grab(monitor))
# To get real color we do this:
image_np = cv2.cvtColor(image_np, cv2.COLOR_BGR2RGB)
# Expand dimensions since the model expects images to have shape: [1, None, None, 3]
image_np_expanded = np.expand_dims(image_np, axis=0)
# Actual detection.
image_tensor = detection_graph.get_tensor_by_name('image_tensor:0')
boxes = detection_graph.get_tensor_by_name('detection_boxes:0')
scores = detection_graph.get_tensor_by_name('detection_scores:0')
classes = detection_graph.get_tensor_by_name('detection_classes:0')
num_detections = detection_graph.get_tensor_by_name('num_detections:0')
# Visualization of the results of a detection.
(boxes, scores, classes, num_detections) = sess.run(
[boxes, scores, classes, num_detections],
feed_dict={image_tensor: image_np_expanded})
vis_util.visualize_boxes_and_labels_on_image_array(
image_np,
np.squeeze(boxes),
np.squeeze(classes).astype(np.int32),
np.squeeze(scores),
category_index,
use_normalized_coordinates=True,
line_thickness=3)
# Show image with detection
cv2.imshow(title, cv2.cvtColor(image_np, cv2.COLOR_BGR2RGB))
# Bellow we calculate our FPS
fps+=1
TIME = time.time() - start_time
if (TIME) >= display_time :
print("FPS: ", fps / (TIME))
fps = 0
start_time = time.time()
# Press "q" to quit
if cv2.waitKey(25) & 0xFF == ord("q"):
cv2.destroyAllWindows()
break
Same as in 3-rd part, we are testing how fast it is working. To compare results we got in 3-rd tutorial part we are taking the same picture, with the same object count in it. In bellow image, you can see significant difference comparing what we had before, it is in average 1 FPS. If you will run it on GPU you will get from 5 to 10 times boost.
Welcome to part 5 of the TensorFlow Object Detection API tutorial series. In this part and few in future, we're going to cover how we can track and detect our own custom objects with this API.
I am doing this by using the pre-built model to add custom detection objects to it. That’s a decent jump from my findings, and it’s quite hard to locate any full step-by-step tutorials, so hopefully I can help you with that. Once you finish this tutorial you will have the ability to train for any custom object you can think of (and create data for) - that’s an awesome skill to have in my opinion.
Alright, so this is my overview of the steps needed to do in this tutorial:
- Collect at least 500 images that contain your object - The bare minimum would be about 100, ideally more like 1000 or more, but, the more images you have, the more tedious step 2 will be.
- Annotate/label the images. I am personally using LabelImg. This process is basically drawing boxes around your object(s) in an image. The label program automatically will create an XML file that describes the object(s) in the pictures.
- Split this data into train/test samples. Training data should be around 80% and testing around 20%.
- Generate TF Records from these splits.
- Setup a .config file for the model of choice (you could train your own from scratch, but we'll be using transfer learning).
- Train our model.
- Export inference graph from new trained model.
- Detect custom objects in real time.
TensorFlow needs hundreds of images of an object to train a good detection classifier, best would be at least 1000 pictures for one object. To train a robust classifier, the training images should have random objects in the image along with the desired objects, and should have a variety of backgrounds and lighting conditions. There should be some images where the desired object is partially obscured, overlapped with something else, or only halfway in the picture.
In my classifier, I will use four different objects I want to detect (terrorist, terrorist head, counter-terrorist and counter-terrorist head). I used to play CSGO with harmless bots to gather hundreds of these images playing on few different game maps, also I was grabbing some pictures from google, to make them different from my game.
Here comes the best part. With all the pictures gathered, it’s time to label the desired objects in every picture. LabelImg is a great tool for labeling images, and its GitHub page has very clear instructions on how to install and use it.
Download and install LabelImg, when running this, you should get a GUI window. From here, choose to open dir and pick the directory that you saved all of your images to. Now, you can begin to annotate images with the create rectbox button. Draw your box, add the name in, and hit ok. Save, hit next image, and repeat! You can press the w key to draw the box and do ctrl+s to save faster. For me it took in average 1 hour for 100 images, this depends on object quantity you have in image. Keep in mind, this will take a while!
LabelImg saves a .xml file containing the label data for each image. These .xml files will be used to generate TFRecords, which are one of the inputs to the TensorFlow trainer. Once you have labeled and saved each image, there will be one .xml file for each image in the \test and \train directories.
Once you have labeled your images, we're going to separate them into training and testing groups. To do this, just copy about 20% of your images and their annotation XML files to a new dir called test and then copy the remaining ones to a new dir called train.
With the images labeled, it’s time to generate the TFRecords that serve as input data to the TensorFlow training model. This tutorial uses the xml_to_csv.py and generate_tfrecord.py scripts from EdjeElectronics repository, with some slight modifications to work with our directory structure.
First, the image .xml data will be used to create .csv files containing all the data for the train and test images. From the main folder, if you are using the same file structure issue the following command in command prompt: python xml_to_csv.py.
This creates a train_labels.csv and test_labels.csv file in the CSGO_images folder. If you are using different files structure, please change xml_to_csv.py accordingly. To avoid of using cmd I created short .bat script called xml_to_csv.bat.
Next, open the generate_tfrecord.py file in a text editor, this file I also took from EdjeElectronics repository. Replace the label map with your own label map, where each object is assigned with an ID number. This same number assignment will be used when configuring the labelmap.pbtxt file.
For example, if you are training your own classifier, you will replace the following code in generate_tfrecord.py:
# TO-DO replace this with label map
def class_text_to_int(row_label):
if row_label == 'c':
return 1
elif row_label == 'ch':
return 2
elif row_label == 't':
return 3
elif row_label == 'th':
return 4
else:
return None
Then, generate the TFRecord files by starting my created generate_tfrecord.bat file, which is issuing these commands from local folder:
python generate_tfrecord.py --csv_input=CSGO_images\train_labels.csv --image_dir= CSGO_images \train --output_path= CSGO_images\train.record
python generate_tfrecord.py --csv_input= CSGO_images \test_labels.csv --image_dir= CSGO_images \test --output_path= CSGO_images\test.record
These lines generates a train.record and a test.record files in training folder. These will be used to train the new object detection classifier.
The last thing to do before training is to create a label map and edit the training configuration file. The label map tells the trainer what each object is by defining a mapping of class names to class ID numbers. Use a text editor to create a new file and save it as labelmap.pbtxt in the training folder. (Make sure the file type is .pbtxt). In the text editor, copy or type the label map in the same format as below (the example below is the label map for my CSGO detector). The label map ID numbers should be the same as defined in the generate_tfrecord.py file.
item {
id: 1
name: 'c'
}
item {
id: 2
name: 'ch'
}
item {
id: 3
name: 't'
}
item {
id: 4
name: 'th'
}
Finally, the object detection training pipeline must be configured. It defines which model and what parameters will be used for training. This is the last step before running actual training.
Navigate to your TensorFlow research\object_detection\samples\configs directory and copy the faster_rcnn_inception_v2_coco.config file into the CSGO_training directory. Then, open the file with a text editor, I personally use notepad++. There are needed several changes to make to this .config file, mainly changing the number of classes, examples and adding the file paths to the training data.
By teh way, the paths must be entered with single forward slashes "/", or TensorFlow will give a file path error when trying to train the model. The paths must be in double quotation marks ( " ), not single quotation marks ( ' ).
Line 10. Change num_classes to the number of different objects you want the classifier to detect. For my CSGO object detection it would be num_classes : 4
Line 107. Change fine_tune_checkpoint to:
fine_tune_checkpoint : "faster_rcnn_inception_v2_coco_2018_01_28/model.ckpt"
Lines 122 and 124. In the train_input_reader section, change input_path and label_map_path to:
input_path: "CSGO_images/train.record"
label_map_path: "CSGO_training/labelmap.pbtxt"
Line 128. Change num_examples to the number of images you have in the CSGO_images\test directory. I have 113 images, so I change it to: "num_examples: 113"
Lines 136 and 138. In the eval_input_reader section, change input_path and label_map_path to:
input_path: "CSGO_images/test.record"
label_map_path: "CSGO_training/labelmap.pbtxt"
Save the file after the changes have been made. That’s it! The training files are prepared and configured for training. One more step left before training.
I have not been able to get model_main.py to work correctly yet, I run in to errors. However, the train.py file is still available in the /object_detection/legacy folder. Simply move train.py from /object_detection/legacy into the /object_detection folder. Move our created CSGO_images and CSGO_training folders into the /object_detection folder and then continue with following line in cmd from object_detection directory:
python train.py --logtostderr --train_dir=CSGO_training/ --pipeline_config_path= CSGO_training/faster_rcnn_inception_v2_pets.config
If everything has been set up correctly, TensorFlow will initialize the training. The initialization can take up to 30 to 60 seconds before the actual training begins. When training begins, it will look like this:
In picture above, each step of training reports the loss. It will start high and get lower and lower as training progresses. For my training it started at about 2-3 and quickly dropped below 0.5. I recommend allowing your model to train until the loss consistently drops below 0.05, which may take about 30,000 steps, or about few hours (depends on how powerful your CPU or GPU is). The loss numbers may be different while different model is used. Also it depends from the objects you want to detect.
You can and you should view the progress of the training by using TensorBoard. To do this, open a new window of CMD and change to the C:\TensorFlow\research\object_detection directory (or directory you have) and issue the following command: C:\TensorFlow\research\object_detection>tensorboard --logdir=CSGO_training
This will create a local webpage on your local machine at YourPCName:6006, which can be viewed through a web browser. The TensorBoard page provides information and graphs that show how the training is progressing. One of the most important graph is the Loss graph, which shows the overall loss of the classifier over time:
The training routine periodically saves checkpoints about every ten minutes. You can terminate the training by pressing Ctrl+C while in the command prompt window. I usually wait until checkpoint has been saved to terminate the training. Then you can terminate training and start it later, and it will restart from the last saved checkpoint.
That’s all for this tutorial, in next tutorial checkpoint at the highest number of steps will be used to generate the frozen inference graph and test it out