Setup Tensorflow Objection Detection API

Prerequisite: Ensure protocol buffer is installed.Instructions for installing on Centos7/RHEL7/RHEL8 is given here

#test if protoc is installed.
▶ protoc --version
libprotoc 3.11.1

If you are using TensorFlow with GPU then CUDA10.0 must be installed and the following envs should be set.

▶ export PATH=/usr/local/cuda-10.0/bin${PATH:+:${PATH}}
▶ export LD_LIBRARY_PATH=/usr/local/cuda-10.0/lib64${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}

Clone tensorflow/models repo

▶ mkdir obj_detection
▶ cd obj_detection
▶ git clone https://github.com/tensorflow/models.git
▶ mkdir workspace
▶ ll
total 0
drwxrwxr-x 7 suku suku 249 Jan  6 21:06 models
drwxrwxr-x 7 suku suku  97 Jan  6 22:18 workspace

Create virtualenv

▶ virtualenv env
▶ source env/bin/activate
▶ ll
total 0
drwxrwxr-x 7 suku suku  80 Jan  6 22:16 env
drwxrwxr-x 7 suku suku 249 Jan  6 21:06 models
drwxrwxr-x 7 suku suku  97 Jan  6 22:18 workspace
▶ python -V
Python 3.6.8

Build Object Detection Python module & compile the protobufs

▶ cd models/research
▶ protoc object_detection/protos/*.proto --python_out=.
▶ export PYTHONPATH=$(pwd):$(pwd)/slim/:$PYTHONPATH
▶ python3 setup.py install
...
Using /obj_detection/env/lib/python3.6/site-packages
Finished processing dependencies for object-detection==0.1
▶ python3 setup.py build

Install dependencies

▶ cd ../../
▶ ls
drwxrwxr-x 7 suku suku  80 Jan  6 23:01 env
drwxrwxr-x 7 suku suku 249 Jan  6 22:59 models
drwxrwxr-x 7 suku suku  97 Jan  6 23:07 workspace
▶ pip3 install absl-py tensorflow-gpu==1.15.0 matplotlib image Cython jupyterlab pandas
▶ pip3 install pycocotools

Test if object_detection api is working.

You should get an OK at the end.

▶ python3 -c "from pycocotools import mask as mask ;print(mask.__author__)"
tsungyi
▶ python3 -c "import tensorflow as tf; print(tf.__version__); print(tf.test.is_gpu_available())"
1.15.0
True
▶ python3 models/research/object_detection/builders/model_builder_test.py
...
[       OK ] ModelBuilderTest.test_unknown_faster_rcnn_feature_extractor
[ RUN      ] ModelBuilderTest.test_unknown_meta_architecture
[       OK ] ModelBuilderTest.test_unknown_meta_architecture
[ RUN      ] ModelBuilderTest.test_unknown_ssd_feature_extractor
[       OK ] ModelBuilderTest.test_unknown_ssd_feature_extractor
----------------------------------------------------------------------
Ran 17 tests in 0.126s

OK (skipped=1)

Training Custom Object Detector

Create workspace folders

▶ cd workspace
▶ mkdir -p images annotations pre-trained-model scripts my_project
▶ ll
total 0
drwxrwxr-x 2 suku suku 115 Jan  6 23:05 annotations
drwxrwxr-x 4 suku suku  31 Jan  6 23:04 images
drwxrwxr-x 2 suku suku   6 Jan  6 23:02 my_project
drwxrwxr-x 3 suku suku  46 Jan  6 23:03 pre-trained-model
drwxrwxr-x 2 suku suku  55 Jan  6 23:04 scripts

Download SSD Model

▶ cd pre-trained-model
▶ wget http://download.tensorflow.org/models/object_detection/ssd_inception_v2_coco_2017_11_17.tar.gz
▶ tar -xvf ssd_inception_v2_coco_2017_11_17.tar.gz
▶ ll ./
total 266M
drwxr-x--- 3 suku suku  161 Nov 17  2017 ssd_inception_v2_coco_2017_11_17
-rw-rw-r-- 1 suku suku 266M Nov 17  2017 ssd_inception_v2_coco_2017_11_17.tar.gz
▶ rm -fr ssd_inception_v2_coco_2017_11_17.tar.gz
▶ ll ./ssd_inception_v2_coco_2017_11_17
total 197M
-rw-r----- 1 suku suku   77 Nov 17  2017 checkpoint
-rw-r----- 1 suku suku  98M Nov 17  2017 frozen_inference_graph.pb
-rw-r----- 1 suku suku  96M Nov 17  2017 model.ckpt.data-00000-of-00001
-rw-r----- 1 suku suku  18K Nov 17  2017 model.ckpt.index
-rw-r----- 1 suku suku 3.6M Nov 17  2017 model.ckpt.meta
drwxr-x--- 3 suku suku   45 Nov 17  2017 saved_model

Prepare the Datasets using labelImg

Separate the data into 2 sets train and test.

▶ cd ../
▶ ls
annotations  images  my_project  pre-trained-model  scripts
▶ cd images
▶ mkdir -p test train
▶ ll images
total 16K
drwxr-xr-x 2 suku suku 4.0K Jan  6 21:11 test
drwxr-xr-x 2 suku suku 8.0K Jan  6 21:11 train

#on mac
virtualenv -p python3 env
source env/bin/activate
labelImg

Convert the xml files to csv files

▶ ll scripts
total 12K
-rw------- 1 suku suku 4.1K Jan  6 21:20 generate_tfrecord.py
-rw------- 1 suku suku 2.4K Jan  6 21:20 xml_to_csv.py
▶ python3 scripts/xml_to_csv.py -i images/test -o annotations/test_labels.csv
Successfully converted xml to csv.

▶ python3 scripts/xml_to_csv.py -i images/train -o annotations/train_labels.csv
Successfully converted xml to csv.

Generate tf records from datasets

create a Label map file which maps indices to category names.

▶ cat annotations/label_map.pbtxt
item{
       id:1
       name:'dynatrace_logo'
      }
item{
       id:2
       name: 'dynatrace_text'
      }

▶ ll annotations
total 20K
-rw------- 1 suku suku  111 Jan  6 23:05 label_map.pbtxt
-rw-rw-r-- 1 suku suku 2.8K Jan  6 23:04 test_labels.csv
-rw-rw-r-- 1 suku suku  12K Jan  6 23:04 train_labels.csv
▶ python3 scripts/generate_tfrecord.py --label0=dynatrace_logo --label1=dynatrace_text --csv_input=annotations/test_labels.csv --img_path=images/test --output_path=annotations/test.record
Successfully created the TFRecords: /home/suku/development/tf_odapi/obj_detection/workspace/annotations/test.record

▶ python3 scripts/generate_tfrecord.py --label0=dynatrace_logo --label1=dynatrace_text --csv_input=annotations/train_labels.csv --img_path=images/train --output_path=annotations/train.record
Successfully created the TFRecords: /home/suku/development/tf_odapi/obj_detection/workspace/annotations/train.record

▶ ll annotations
total 22M
-rw------- 1 suku suku  111 Jan  6 22:18 label_map.pbtxt
-rw-rw-r-- 1 suku suku 2.8K Jan  6 22:24 test_labels.csv
-rw-rw-r-- 1 suku suku 2.8M Jan  6 22:26 test.record
-rw-rw-r-- 1 suku suku  12K Jan  6 21:10 train_labels.csv
-rw-rw-r-- 1 suku suku  20M Jan  6 22:27 train.record

Setup the training projects with relevant files

open ssd_inception_v2_coco.config and make change where "CHANGE HERE" comment is present.

▶ cp ../models/research/object_detection/legacy/train.py my_project/
▶ cp ../models/research/object_detection/export_inference_graph.py my_project/
▶ cd my_project
▶ wget https://raw.githubusercontent.com/tensorflow/models/master/research/object_detection/samples/configs/ssd_inception_v2_coco.config
▶ ll ./
total 16K
-rw------- 1 suku suku 4.5K Jan  6 21:25 ssd_inception_v2_coco.config
-rw-rw-r-- 1 suku suku 6.8K Jan  6 21:27 train.py
-rw-rw-r-- 1 suku suku 7.1K Jan  6 21:38 export_inference_graph.py

Train the model

▶ python3 train.py --pipeline_config_path=./ssd_inception_v2_coco.config --train_dir=./training  --alsologtostderr
...
I0106 22:33:28.245677 140687540409920 learning.py:507] global step 59: loss = 16.1677 (0.050 sec/step)
INFO:tensorflow:global step 60: loss = 12.5695 (0.052 sec/step)
I0106 22:33:28.298278 140687540409920 learning.py:507] global step 60: loss = 12.5695 (0.052 sec/step)
INFO:tensorflow:global step 61: loss = 11.3254 (0.056 sec/step)
I0106 22:33:28.355229 140687540409920 learning.py:507] global step 61: loss = 11.3254 (0.056 sec/step)
^CTraceback (most recent call last):
[CtrL+C] to terminate the training

▶ ls
training  ssd_inception_v2_coco.config  train.py export_inference_graph.py

▶ ll training
total 248M
-rw-rw-r-- 1 suku suku   81 Jan  6 21:36 checkpoint
-rw-rw-r-- 1 suku suku  19M Jan  6 21:36 events.out.tfevents.1578364596.localhost.localdomain
-rw-rw-r-- 1 suku suku  17M Jan  6 21:36 graph.pbtxt
-rw-rw-r-- 1 suku suku 204M Jan  6 21:36 model.ckpt-50.data-00000-of-00001
-rw-rw-r-- 1 suku suku  53K Jan  6 21:36 model.ckpt-50.index
-rw-rw-r-- 1 suku suku 9.7M Jan  6 21:36 model.ckpt-50.meta
-rw------- 1 suku suku 4.8K Jan  6 21:36 pipeline.config

Create saved_model for exporting the model.

▶ python3 export_inference_graph.py --input_type=image_tensor --pipeline_config_path=./ssd_inception_v2_coco.config --trained_checkpoint_prefix=./training/model.ckpt-50 --output_directory=./inference
...
INFO:tensorflow:Writing pipeline config file to ./inference/pipeline.config
I0106 22:35:18.542130 140537440007744 config_util.py:190] Writing pipeline config file to ./inference/pipeline.config


▶ ll
total 696K
-rw-rw-r-- 1 suku suku 7.1K Jan  6 23:09 export_inference_graph.py
drwxr-xr-x 3 suku suku  184 Jan  6 23:11 inference
-rw-rw-r-- 1 suku suku 672K Jan  6 23:13 inference.ipynb
-rw------- 1 suku suku 4.8K Jan  6 23:09 ssd_inception_v2_coco.config
drwxrwxr-x 2 suku suku   40 Jan  6 23:21 test_images
drwxr-xr-x 2 suku suku  217 Jan  6 23:10 training
-rw-rw-r-- 1 suku suku 6.8K Jan  6 23:09 train.py

▶ ll inference
total 105M
-rw-rw-r-- 1 suku suku   77 Jan  6 21:41 checkpoint
-rw-rw-r-- 1 suku suku  52M Jan  6 21:41 frozen_inference_graph.pb
-rw-rw-r-- 1 suku suku  51M Jan  6 21:41 model.ckpt.data-00000-of-00001
-rw-rw-r-- 1 suku suku  18K Jan  6 21:41 model.ckpt.index
-rw-rw-r-- 1 suku suku 1.7M Jan  6 21:41 model.ckpt.meta
-rw-rw-r-- 1 suku suku 3.9K Jan  6 21:41 pipeline.config
drwxr-xr-x 3 suku suku   45 Jan  6 21:41 saved_model


▶ ll inference/saved_model
total 52M
-rw-rw-r-- 1 suku suku 52M Jan  6 21:41 saved_model.pb
drwxr-xr-x 2 suku suku   6 Jan  6 21:41 variables

Test the model

To test Inference code create a new environment

▶ cd ../../../
▶ ll
total 4.0K
drwxrwxr-x 7 suku suku  80 Jan  6 23:01 env
drwxrwxr-x 7 suku suku 249 Jan  6 22:59 models
drwxrwxr-x 7 suku suku  97 Jan  6 23:07 workspace

▶ virtualenv env2
▶ source env2/bin/activate
▶ ll
total 4.0K
drwxrwxr-x 7 suku suku  80 Jan  6 23:01 env
drwxrwxr-x 5 suku suku  56 Jan  6 23:15 env2
drwxrwxr-x 7 suku suku 249 Jan  6 22:59 models
-rw-rw-r-- 1 suku suku  15 Jan  6 22:58 README.md
drwxrwxr-x 7 suku suku  97 Jan  6 23:07 workspace

▶ pip3 install absl-py tensorflow-gpu=="2.*"  matplotlib image Cython jupyterlab pandas
Ensure that PYTHONPATH has not changed.Objection detection apis will be used in inference.
▶ jupyter lab
Open inference.ipynb

Few Useful Commands:

ll *.jpg |wc -l ; ll *.png |wc -l ; ll *.xml |wc -l
ls -1 *.png | xargs -n 1 bash -c 'convert "$0" "${0%.png}.jpg"'
ls -1 *.jpg | head -n 100 | xargs -n 1 bash -c 'cp "$0" ../100files/'
for f in *.xml; do echo "Processing $f file.." && sed -i 's/png/jpg/g' $f ; done
for f in *.xml; do cat $f|grep "jpg"; done
for f in *.xml; do cat $f|grep "png"; done
for f in *.xml; do mv $f IBM_$f; done 

ls -1 *.jpg | head -n 1 | xargs -n 1 bash -c 'IN="$0"; IFS='.' read -ra NAMES <<< "$IN"; ls -1 ${NAMES[0]}*'
ls -1 eight*.jpg | head -n 1 | xargs -n 1 bash -c 'IN="$0"; IFS='.' read -ra NAMES <<< "$IN"; cp ${NAMES[0]}* ../train'

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

setup.md

setup.md

Setup Tensorflow Objection Detection API

Clone tensorflow/models repo

Create virtualenv

Build Object Detection Python module & compile the protobufs

Install dependencies

Test if object_detection api is working.

Training Custom Object Detector

Download SSD Model

Prepare the Datasets using labelImg

Convert the xml files to csv files

Generate tf records from datasets

Setup the training projects with relevant files

Train the model

Create saved_model for exporting the model.

Test the model

Files

setup.md

Latest commit

History

setup.md

File metadata and controls

Setup Tensorflow Objection Detection API

Clone tensorflow/models repo

Create virtualenv

Build Object Detection Python module & compile the protobufs

Install dependencies

Test if object_detection api is working.

Training Custom Object Detector

Download SSD Model

Prepare the Datasets using labelImg

Convert the xml files to csv files

Generate tf records from datasets

Setup the training projects with relevant files

Train the model

Create saved_model for exporting the model.

Test the model