TensorFlow Object Detection API

Training Custom Object Detector Step by Step

🌱 Introduction

  • ✨ Tensorflow object detection API is a powerful tool that allows us to create custom object detectors depending on pre-trained, fine tuned models even if we don't have strong AI background or strong TensorFlow knowledge.
  • 💁‍♀️ Building models depending on pre-trained models saves us a lot of time and labor since we are using models that maybe trained for weeks using very strong machines, this principle is called Transfer Learning.
  • 🗃️ As a data set I will show you how to use OpenImages data set and converting its data to TensorFlow-friendly format.
  • 🎀 You can find this article on Medium too.

🚩 Development Pipeline

🤕 While you are applying the instructions if you get errors you can check out 🐞 Common Issues section at the end of the article

👩‍💻 Environment Preparation

🔸 Environment Info

💻 Platform
🏷️ Version
Python version
TensorFlow version

🥦 Conda env Setting

🔮 Create new env

  • 🥦 Install Anaconda​
  • 💻 Open cmd and run:
# conda create -n <ENV_NAME> python=<REQUIRED_VERSION>
conda create -n tf1 python=3.7

▶️ Activate the new env

# conda activate <ENV_NAME>
conda activate tf1

🔽 Install Packages

💥 GPU vs CPU Computing

Brain of computer
Brawn of computer
Very few complex cores
hundreds of simpler cores with parallel architecture
single-thread performance optimization
thousands of concurrent hardware threads
Can do a bit of everything, but not great at much
Good for math heavy processes

🚀 Installing TensorFlow

conda install tensorflow-gpu=1.15
conda install tensorflow=1.15

📦 Installing other packages

conda install pillow Cython lxml jupyter matplotlib
conda install -c anaconda protobuf

🤖 Downloading models repository

🤸‍♀️ Cloning from GitHub

  • A repository that contains required utils for training and evaluation process
  • Open CMD and run in E disk and run:
# note that every time you open CMD you have
# to activate your env again by running:
# under E:\>
conda activate tf1
git clone
cd models/research
🧐 I assume that you are running your commands under E disk,

🔃 Compiling Protobufs

💻 Windows
🐧 Linux
# under (tf1) E:\models\research>
for /f %i in ('dir /b object_detection\protos\*.proto') do protoc object_detection\protos\%i --python_out=.
# under /models/research
$ protoc object_detection/protos/*.proto --python_out=.

📦 Compiling Packages

# under (tf1) E:\models\research>
python build
python install

🚩 Setting Python Path Temporarily

💻 Windows
🐧 Linux
# under (tf1) E:\models\research> or anywhere 😅
set PYTHONPATH=E:\models\research;E:\models\research\slim
# under /models/research
$ export PYTHONPATH=`pwd`:`pwd`/slim
👮‍♀️ Every time you open CMD you have to set PYTHONPATH again

👩‍🔬 Installation Test

🧐 Check out that every thing is done

💻 Command

# under (tf1) E:\models\research>
python object_detection/builders/

🎉 Expected Output

Ran 17 tests in 0.833s
OK (skipped=1)

🖼️ Image Acquiring

👮‍♀️ Directory Structure

  • 🏗️ I suppose that you created a structure like:
|___ models
|___ demo
|___ annotations
|___ eval
|___ images
|___ inference
|___ OIDv4_ToolKit
|___ OpenImagesTool
|___ pre_trainded_model
|___ scripts
|___ training
📂 Folder
📃 Description
🤖 models
the repo here​
📄 annotations
will contain generated .csv and .record files
👮‍♀️ eval
will contain results of evaluation
🖼️ images
will contain image data set
▶️ inference
will contain exported models after training
🔽 OIDv4_ToolKit
the repo here (OpenImages Downloader)
👩‍🔧 OpenImagesTool
the repo here (OpenImages Organizer)
will contain files of TensorFlow model that we will retrain
👩‍💻 scripts
will contain scripts that we will use for pre-processing and training processes
🚴‍♀️ training
will contain generated check points during training

🚀 OpenImages Dataset

  • 🕵️‍♀️ You can get images in various methods
  • 👩‍🏫 I will show process of organizing OpenImages data set
  • 🗃️ OpenImages is a huge data set contains annotated images of 600 objects
  • 🔍 You can explore images by categories from here

🎨 Downloading By Category

​OIDv4_Toolkit is a tool that we can use to download OpenImages dataset by category and by set (test, train, validation)
💻 To clone and build the project, open CMD and run:
# under (tf1) E:\demo>
git clone
cd OIDv4_ToolKit
# under (tf1) E:\demo\OIDv4_ToolKit>
pip install -r requirements.txt
⏬ To start downloading by category:
# python downloader --classes <OBJECT_LIST> --type_csv <TYPE>
# TYPE: all | test | train | validation
# under (tf1) E:\demo\OIDv4_ToolKit>
python downloader --classes Apple Orange --type_csv validation
👮‍♀️ If object name consists of 2 parts then write it with '_', e.g. Bell_pepper

🤹‍♀️ Image Organization

🔮 OpenImagesTool

  • 👩‍💻 OpenImagesTool is a tool to convert OpenImages images and annotations to TensorFlow-friendly structure.
  • 🙄 OpenImages provides annotations ad .txt files in a format like:<OBJECT_NAME> <XMIN> <YMIN> <XMAX> <YMAX> which is not compatible with TensorFlow that requires VOC annotation format
  • 💫 To do that synchronization we can do the following
💻 To clone and build the project, open CMD and run:
# under (tf1) E:\demo>
git clone
cd OpenImagesTool/src

💻 Applying Organizing

🚀 Now, we will convert images and annotations that we have downloaded and save them to images folder
# under (tf1) E:\demo\OpenImagesTool\src>
# python -i <INPUT_PATH> -o <OUTPUT_PATH>
python -i E:\pre_trainded_model\OIDv4_ToolKit\OID\Dataset -o E:\pre_trainded_model\images
👩‍🔬 OpenImagesTool adds validation images to training set by default, if you wand to disable this behavior you can add -v flag to the command.

🏷️ Creating Label Map

  • ⛓️ label_map.pbtxt is a file that maps object names to corresponded IDs
  • ➕ Create label_map.pbtxtfile under annotations folder and open it in a text editor
  • 🖊️ Write your objects names and IDs in the following format
item {
id: 1
name: 'Hamster'
item {
id: 2
name: 'Apple'
👮‍♀️ id:0 is reserved for background, so don' t use it
🐞 Related error: ValueError: Label map id 0 is reserved for the background label

🏭 Generating CSV Files

  • 🔄 Now we have to convert .xml files to csv file
  • 🔻 Download the script script and save it under scripts folder
  • 💻 Open CMD and run:

👩‍🔬 Generating train csv file

# under (tf1) E:\demo\scripts>
python -i E:\demo\images\train -o E:\demo\annotations\train_labels.csv

👩‍🔬 Generating test csv file

# under (tf1) E:\demo\scripts>
python -i E:\demo\images\test -o E:\demo\annotations\test_labels.csv

👩‍🏭 Generating TF Records

  • 🙇‍♀️ Now, we will generate tfrecords that will be used in training precess
  • 🔻 Download script and save it under scripts folder

👩‍🔬 Generating train tfrecord

# under (tf1) E:\demo\scripts>
# python --label_map=<PATH_TO_LABEL_MAP>
# --csv_input=<PATH_TO_CSV_FILE> --img_path=<PATH_TO_IMAGE_FOLDER>
# --output_path=<PATH_TO_OUTPUT_FILE>
python --label_map=E:/demo/annotations/label_map.pbtxt --csv_input=E:\demo\annotations\train_labels.csv --img_path=E:\demo\images\train --output_path=E:\demo\annotations\train.record

👩‍🔬 Generating test tfrecord

# under (tf1) E:\demo\scripts>
python --label_map=E:/demo/annotations/label_map.pbtxt --csv_input=E:\demo\annotations\test_labels.csv --img_path=E:\demo\images\test --output_path=E:\demo\annotations\test.record

🤖 Model Selecting

  • 🎉 TensorFLow Object Detection Zoo provides a lot of pre-trained models
  • 🕵️‍♀️ Models differentiate in terms of accuracy and speed, you can select the suitable model due to your priorities
  • 💾 Select a model, extract it and save it under pre_trained_model folder
  • 👀 Check out my notes here to get insight about differences between popular models

👩‍🔧 Model Configuration

⏬ Downloading config File

  • 😎 We have downloaded the models (pre-trained weights) but now we have to download configuration file that contains training parameters and settings
  • 👮‍♀️ Every model in TensorFlow Object Detection Zoo has a configuration file presented here​
  • 💾 Download the config file that corresponds to the models you have selected and save it under training folder

👩‍🔬 Updating config File

You have to update the following lines:
🙄 Take a look at Loss exploding issue​
// number of classes
num_classes: 1 // set it to total number of classes you have
// path of pre-trained checkpoint
fine_tune_checkpoint: "E:/demo/pre_trained_model/ssd_mobilenet_v1_quantized_300x300_coco14_sync_2018_07_18/model.ckpt"
// path to train tfrecord
tf_record_input_reader {
input_path: "E:/demo/annotations/train.record"
// number of images that will be used in evaluation process
eval_config: {
metrics_set: "coco_detection_metrics"
use_moving_averages: false
// I suggest setting it to total number of testing set to get accurate results
num_examples: 11193
eval_input_reader: {
tf_record_input_reader {
// path to test tfrecord
input_path: "E:/demo/annotations/test.record"
// path to label map
label_map_path: "E:/demo/annotations/label_map.pbtxt"
// set it to true if you want to shuffle test set at each evaluation
shuffle: false
num_readers: 1
🤹‍♀️ If you give the whole test set to evaluation process then shuffle functionality won't affect the results, it will only give you different examples on TensorBoard

👶 Training

  • 🎉 Now we have done all preparations
  • 🚀 Let the computer start learning
  • 💻 Open CMD and run:
# under (tf1) E:\models\research\object_detection\legacy>
# python --train_dir=<DIRECTORY_TO_SAVE_CHECKPOINTS>
# --pipeline_config_path=<PATH_TO_CONFIG_FILE>
python --train_dir=E:/demo/training --pipeline_config_path=E:/demo/training/ssd_mobilenet_v1_quantized_300x300_coco14_sync.config
  • 🕐 This process will take long (You can take a nap 🤭, but a long nap 🙄)
  • 🕵️‍♀️ While model is being trained you will see loss values on CMD
  • ✋ You can stop the process when the loss value achieves a good value (under 1)

👮‍♀️ Evaluation

🎳 Evaluating Script

  • 🤭 After training process is done, let's do an exam to know how good (or bad 🙄) is our model doing
  • 🎩 The following command will use the model on whole test set and after that print the results, so that we can do error analysis.
  • 💻 So that, open CMD and run:
# under (tf1) E:\models\research\object_detection\legacy>
# python --logtostderr --pipeline_config_path=<PATH_TO_CONFIG_FILE>
python --pipeline_config_path=E:/demo/training/ssd_mobilenet_v1_quantized_300x300_coco14_sync.config --checkpoint_dir=E:/demo/training --eval_dir=E:/demo/eval

👀 Visualizing Results

  • ✨ To see results on charts and images we can use TensorBoard for better analyzing
  • 💻 Open CMD and run:

👩‍🏫 Training Values Visualization

  • 🧐 Here you can see graphs of loss, learning rate and other values
  • 🤓 And much more (You can investigate tabs at the top)
  • 😋 It is feasable to use it while training (and exciting 🤩)
# under (tf1) E:\>
tensorboard --logdir=E:/demo/tarining

👮‍♀️ Evaluation Values Visualization

  • 👀 Here you can see images from your test set with corresponded predictions
  • 🤓 And much more (You can inspect tabs at the top)
  • ❗ You must use this after running evaluation script
# under (tf1) E:\>
tensorboard --logdir=E:/demo/eval
  • 🔍 See the visualized results on localhost:6006 and
  • 🧐 You can inspect numerical values from report on terminal, result example:
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.708
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.984
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.868
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.289
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.623
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.767
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.779
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.781
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.781
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.300
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.703
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.824
  • 🎨 If you want to get metric report for each class you have to change evaluating protocol to pascal metrics by configuring metrics_set in .config file:
eval_config: {
metrics_set: "weighted_pascal_voc_detection_metrics"

👒 Model Exporting

  • 🔧 After training and evaluation processes are done, we have to make the model in such a format that we can use
  • 🦺 For now, we have only checkpoints, so that we have to export .pb file
  • 💻 So, open CMD and run:
# under (tf1) E:\models\research\object_detection>
# python --input_type image_tensor
# --pipeline_config_path <PATH_TO_CONFIG_FILE>
# --trained_checkpoint_prefix <PATH_TO_LAST_CHECKPOINT>
# --output_directory <PATH_TO_SAVE_EXPORTED_MODEL>
python --input_type image_tensor --pipeline_config_path=E:/demo/training/ssd_mobilenet_v1_quantized_300x300_coco14_sync.config --trained_checkpoint_prefix E:/demo/training/model.ckpt-16438 --output_directory E:/demo/inference/ssd_v1_quant
  • If you are using SSD and planning to convert it to tflite later you have to run
# under (tf1) E:\models\research\object_detection>
# python --input_type image_tensor
# --pipeline_config_path <PATH_TO_CONFIG_FILE>
# --trained_checkpoint_prefix <PATH_TO_LAST_CHECKPOINT>
# --output_directory <PATH_TO_SAVE_EXPORTED_MODEL>
python --input_type image_tensor --pipeline_config_path=E:/demo/training/ssd_mobilenet_v1_quantized_300x300_coco14_sync.config --trained_checkpoint_prefix E:/demo/training/model.ckpt-16438 --output_directory E:/demo/inference/ssd_v1_quant

📱 Converting to tflite

  • 💁‍♀️ If you want to use the model in mobile apps or tflite supported embedded devices you have to convert .pb file to .tflite file

📙 About TFLite

  • 📱 TensorFlow Lite is TensorFlow’s lightweight solution for mobile and embedded devices.
  • 🧐 It enables on-device machine learning inference with low latency and a small binary size.
  • 😎 TensorFlow Lite uses many techniques for this such as quantized kernels that allow smaller and faster (fixed-point math) models.
  • 📍 Official site​
🍫 Converting Command
  • 💻 To apply converting open CMD and run:
# under (tf1) E:\>
# toco --graph_def_file=<PATH_TO_PB_FILE>
# --output_file=<PATH_TO_SAVE> --input_shapes=<INPUT_SHAPES>
# --input_arrays=<INPUT_ARRAYS> --output_arrays=<OUTPUT_ARRAYS>
# --inference_type=<QUANTIZED_UINT8|FLOAT> --change_concat_input_ranges=<true|false>
# --alow_custom_ops
# args for QUANTIZED_UINT8 inference
# --mean_values=<MEAN_VALUES> std_dev_values=<STD_DEV_VALUES>
toco --graph_def_file=E:\demo\inference\ssd_v1_quant\tflite_graph.pb --output_file=E:\demo\tflite\ssd_mobilenet.tflite --input_shapes=1,300,300,3 --input_arrays=normalized_input_image_tensor --output_arrays=TFLite_Detection_PostProcess,TFLite_Detection_PostProcess:1,TFLite_Detection_PostProcess:2,TFLite_Detection_PostProcess:3 --inference_type=QUANTIZED_UINT8 --mean_values=128 --std_dev_values=128 --change_concat_input_ranges=false --allow_custom_ops

🐞 Common Issues

🥅 nets module issue

ModuleNotFoundError: No module named 'nets'
This means that there is a problem in setting PYTHONPATH, try to run:
(tf1) E:\models\research>set PYTHONPATH=E:\models\research;E:\models\research\slim

🗃️ tf_slim module issue

ModuleNotFoundError: No module named 'tf_slim'
This means that tf_slim module is not installed, try to run:
(tf1) E:\models\research>pip install tf_slim

🗃️ Allocation error

2020-08-11 17:44:00.357710: I tensorflow/core/common_runtime/] Stats:
Limit: 10661327
InUse: 10656704
MaxInUse: 10657688
NumAllocs: 2959
MaxAllocSize: 3045064
For me it is fixed by minimizing batch_size in .config file, it is related to your computations resources
train_config: {
batch_size: 128

❗ no such file or directory error tensorflow.python.framework.errors_impl.notfounderror no such file or directory

🤯 LossTensor is inf issue

LossTensor is inf or nan. : Tensor had NaN values
  • 👀 Related discussion is here, it is common that it is an annotation problem
  • 🙄 Maybe there is some bounding boxes outside the image boundaries
  • 🤯 The solution for me was minimizing batch size in .config file

🙄 Ground truth issue

The following classes have no ground truth examples
  • 👀 Related discussion is here​
  • 👩‍🔧 For me it was a misspelling issue in label_map file,
  • 🙄 Pay attention to small and capital letters

🏷️ labelmap issue

ValueError: Label map id 0 is reserved for the background label
  • 👮‍♀️ id:0 is reserved for background, We can not use it for objects
  • 🆔 start IDs from 1

🔦 No Variable to Save issue

Value Error: No Variable to Save
  • 👀 Related solution is here​
  • 👩‍🔧 Adding the following line to .config file solved the problem
train_config: {
fine_tune_checkpoint_type: "detection"

🧪 pycocotools module issue

ModuleNotFoundError: No module named 'pycocotools'
💻 Windows
🐧 Linux
  • 👀 Related discussion is here​
  • 👩‍🔧 Applying the downloading instructions provided here solved the problem for me (on Windows 10)
$ conda install -c conda-forge pycocotools

🥴 pycocotools type error issue

pycocotools typeerror: object of type cannot be safely interpreted as an integer.
  • 👩‍🔧 I solved the problem by editing the following lines in script under pycocotools package (by adding casting)
  • 👮‍♀️ Make sure that you are editting the package in you env not in other env.
self.iouThrs = np.linspace(.5, 0.95, int(np.round((0.95 - .5) / .05)) + 1, endpoint=True)
self.recThrs = np.linspace(.0, 1.00, int(np.round((1.00 - .0) / .01)) + 1, endpoint=True)

💣 Loss Exploding

INFO:tensorflow:global step 440: loss = 2106942657570782838784.0000 (0.405 sec/step)
INFO:tensorflow:global step 440: loss = 2106942657570782838784.0000 (0.405 sec/step)
INFO:tensorflow:global step 441: loss = 7774169971762292326400.0000 (0.401 sec/step)
INFO:tensorflow:global step 441: loss = 7774169971762292326400.0000 (0.401 sec/step)
INFO:tensorflow:global step 442: loss = 25262924095336287830016.0000 (0.404 sec/step)
INFO:tensorflow:global step 442: loss = 25262924095336287830016.0000 (0.404 sec/step)
🙄 For me there were 2 problems:
  • Some of annotations were wrong and overflow the image (e.g. xmax > width)
  • I could check that by inspecting .csv file
  • Example:
  • Learning rate in .config file is too big (the default value was big 🙄)
  • The following values are valid and tested on mobilenet_ssd_v1_quantized (Not very good 🙄)
learning_rate: {
cosine_decay_learning_rate {
learning_rate_base: .01
total_steps: 50000
warmup_learning_rate: 0.005
warmup_steps: 2000

🥴 Getting convolution Failure

Error : Failed to get convolution algorithm. This is probably because cuDNN failed to initialize, so try looking to see if a warning log message was printed above.
  • It may be a Cuda version incompatibility issue
  • For me it was a memory issue and I solved it by adding the following line to script
os.environ['TF_FORCE_GPU_ALLOW_GROWTH'] = 'true'

📦 Invalid box data error

raise ValueError('Invalid box data. data must be a numpy array of '
ValueError: Invalid box data. data must be a numpy array of N*[y_min, x_min, y_max, x_max]
  • 🙄 For me it was a logical error, in test_labels.csv there were some invalid values like: file123.jpg,134,63,3,0,0,-1029,-615
  • 🏷 So, it was a labeling issue, fixing these lines solved the problem

🔄 Image with id added issue

raise ValueError('Image with id {} already added.'.format(image_id))
ValueError: Image with id 123.png already added.
  • ☝ It is an issue in .config caused by giving value to num_example that is greater than total number of test image in test directory
eval_config: {
metrics_set: "coco_detection_metrics"
use_moving_averages: false
num_examples: 1265 // <--- this value was greater than total test images

🧐 References