This document outlines the steps to train a custom object detection model using TensorFlow and Google Cloud Platform. It includes setting up prerequisites like TensorFlow and Google Cloud SDK, creating training datasets with images and annotations, converting data to TFRecord format, training a model using transfer learning on Cloud ML Engine, exporting the trained model, and testing the model.

1. Training
Custom
Object
Detection
Model in GCP
Cloud ML
Engine

3. Prerequisite
• Install Git
• Install Tensorflow
• Requires Python 3.4, 3.5, or 3.6
• pip install tensorflow # CPU Version
• https://www.tensorflow.org/install
• Download Tensorflow models git
• git clone https://github.com/tensorflow/models
• Some other python libraries
• Pillow, lxml, contextlib2 (use pip install)
• Install Google cloud SDK
• https://cloud.google.com/sdk/docs/quickstart-windows

4. Setup Google Cloud Platform
• Create Google Cloud Platform (GCP) Account
• Create a project
• Enable Billing
• On Command prompt (Terminal)
• gcloud init
• gcloud config auth list
• gcloud config list
• gcloud config set project

5. Protobuf
• The TensorFlow Object Detection API uses .proto files which need to be
compiled into .py files. Google provides a program called Protobuf that
will batch compile these.
• cd models/research
• For windows
• Download version 3.4.0 for Windows from
https://github.com/protocolbuffers/protobuf/releases/tag/v3.4.0
• “E:Object_Detectiondemoprotoc-3.4.0-win32binprotoc.exe"
object_detectionprotos*.proto --python_out=.
• For Linux
• sudo apt-get install protobuf-compiler
• protoc object_detection/protos/*.proto --python_out=.
• You will see multiple python files now in “object_detectionprotos” folder

6. Environment Setup
• Linux
• export PYTHONPATH=$PYTHONPATH:`pwd`:`pwd`/slim
• Windows
• Add to Environment Variable
• modelsresearch
• modelsresearchslim
• Test Installation
• python object_detection/builders/model_builder_test.py
• If no errors installation is perfect

7. Create Dataset
• Create two folders:
• images
• Image Name Format: categoryName_imageID.jpg (only one "_")
• annotations
• trainval.txt
• categoryName_imageID category_ID (80% images for train/val)
• test.txt
• categoryName_imageID category_ID (20% images for test)
• list.txt
• categoryName_imageID category_ID (all images)
• xmls
• categoryName_imageID.xml
• Label_map.pbtxt
• Note:
• Make sure no image is corrupt, bounding boxes are well defined.
• Bounding box co-ordinates do not fall outside of image height and width.
• Image Height and width are correct.
• If one image has many object of same kind, only one copy of image is needed with respective object category name and all bounding box locations in
xml.
• If image has many object of different category, multiple copies of same image will be stored with one object category in image name in one copy

8. Samples
images
annotations
xmls
Sample xml
Label_map.pbtxt

9. Create TFRecord
• The Tensorflow Object Detection API uses the TFRecord format for
training and validation datasets
• python object_detection/dataset_tools/create_pet_tf_record.py
--label_map_path=E:Object_Detectiondemolabel_map.pbtxt
--data_dir=E:Object_Detectiondemo
--output_dir=E:Object_Detectiondemo
• You will see two new files at output_dir location
• pet_train.record
• pet_val.record

10. Move data to GCS (Google Cloud Storage)
• Create a new bucket
• gsutil mb gs://train_coke
• Set bucket name for future use
• set YOUR_GCS_BUCKET=gs://train_coke
• Move file to new GCS Bucket
• gsutil cp pet_train.record %YOUR_GCS_BUCKET%/data/pet_train.record
• gsutil cp pet_val.record %YOUR_GCS_BUCKET%/data/pet_val.record
• gsutil cp label_map.pbtxt %YOUR_GCS_BUCKET%/data/pet_label_map.pbtxt

11. Transfer Learning
• Upload trained model to GCS
• wget
https://storage.googleapis.com/download.tensorflow.org/models/object_detection/fast
er_rcnn_resnet101_coco_11_06_2017.tar.gz
• tar -xvf faster_rcnn_resnet101_coco_11_06_2017.tar.gz
• gsutil cp faster_rcnn_resnet101_coco_11_06_2017/model.ckpt.*
%YOUR_GCS_BUCKET%/data/
• Edit Config File
• sed -i "s|PATH_TO_BE_CONFIGURED|"${YOUR_GCS_BUCKET}" data|g"
object_detection/samples/configs/faster_rcnn_resnet101_pets.config #Linux
• gsutil cp object_detection/samples/configs/faster_rcnn_resnet101_pets.config
%YOUR_GCS_BUCKET%/data/faster_rcnn_resnet101_pets.config
• Tensorflow detection model zoo
https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/det
ection_model_zoo.md

12. Train Model
• Package the TensorFlow Object Detection API and TF Slim
• python setup.py sdist
• cd slim
• python setup.py sdist
• Enable Cloud Machine Learning Engine API
• bash object_detection/dataset_tools/create_pycocotools_package.sh /tmp/pycocotools (#Linux)
• gcloud ml-engine jobs submit training object_detection_coke
--runtime-version 1.9
--job-dir=%YOUR_GCS_BUCKET%/model_dir
--packages dist/object_detection-0.1.tar.gz,slim/dist/slim-0.1.tar.gz,pycocotools/pycocotools-2.0.tar.gz
--module-name object_detection.model_main
--region us-east1
--config object_detection/samples/cloud/cloud.yml
--
--model_dir=%YOUR_GCS_BUCKET%/model_dir
--pipeline_config_path=%YOUR_GCS_BUCKET%/data/faster_rcnn_resnet101_pets.config
• Get Logs
• gcloud ml-engine jobs stream-logs object_detection_pets

13. Tensorboard
• pip install tensorboard
• From cloudshell
• tensorboard --logdir=${YOUR_GCS_BUCKET}/model_dir --port=5000

14. Export Trained Model
• Get model checkpoint number
• Set CHECKPOINT_NUMBER=2975
• set YOUR_LOCAL_CHK_DIR=exported_model
• Copy model checkpoint files to local from GCS
• gsutil cp %YOUR_GCS_BUCKET%/model_dir/model.ckpt-
%CHECKPOINT_NUMBER%.* % YOUR_LOCAL_CHK_DIR%
• Export Inference Graph
• python object_detection/export_inference_graph.py
--input_type image_tensor
--pipeline_config_path
object_detection/samples/configs/faster_rcnn_resnet101_pets.config
--trained_checkpoint_prefix %YOUR_LOCAL_CHK_DIR%/model.ckpt-
%CHECKPOINT_NUMBER%
--output_directory exported_graphs

15. Test Model

16. References
• https://cloud.google.com/blog/products/gcp/training-an-object-
detector-using-cloud-machine-learning-engine
• https://towardsdatascience.com/google-object-detection-api-to-
detect-brand-logos-fd9e113725d8
• https://towardsdatascience.com/google-object-detection-api-to-
detect-brand-logos-part-2-384055e4e109
• Git Repo
• https://github.com/sahuvaibhav/Object_Detection