Skip to content
This repository has been archived by the owner on Jul 10, 2022. It is now read-only.

Latest commit

 

History

History
154 lines (123 loc) · 7.36 KB

detection_training_guide.md

File metadata and controls

154 lines (123 loc) · 7.36 KB
Raw

Training with the TensorFlow Object Detection API

Introduction

This guide describes how the TensorFlow Object Detection API (TFODA) needs to be used to train a detection model for TensorFlow. The guide has been written for Ubuntu 18.04 LTS and might not work on Windows or other Linux distributions. The guide assumes you use Conda environments.

Preconditions

  • TensorFlow 1.15 installed
  • Pycocotools 2.0.0 installed
  • Dataset has been prepared
  • CUDA and cuDNN installed

Preparations

Download and setup Object Detection API

TFODA is part of TensorFlow models repository. It can be found on GitHub.

  1. Clone the repository into a location on the target host
  2. Navigate to the models/research/ directory
  3. Run setup.py to get the required dependencies for TFODA

You might need to compile protobufs and install the object_detection package as described in the official installation guide and the Install step of the official demo on the TF models repository. To do so run the following commands from models/research/:

protoc object_detection/protos/*.proto --python_out=.

pip install .

Datasets

Before following this guide the necessary datasets have to be loaded.

To train the detection model three different datasets are used in conjunction.
The necessary datasets can be found in the directory 1_data_preprocessing.

This package contains the data preparation instructions if you would like to train your own models.
Please note that merge_datasets_detection.ipynb has to be only executed after the other notebooks have already been followed.

After executing the data preprocessing script all files can be found under your defined folder name.

Set $PYTHONPATH

Python needs to know where it can find its required dependencies. One of these dependencies is slim that is provided with the TF models repository. For this you need to edit the $PYTHONPATH environment variable.

From the command line run:

export PYTHONPATH=:'[PathToTFODAslim]':$PYTHONPATH

where [PathToTFODAslim] is the absolute path of the slim/ directory located in models/research/slim/ within the TF models repository. Make sure you do not forget to add :$PYTHONPATH to the command or you will overwrite your current $PYTHONPATH variable instead of appending values to it.

You can check if the values have been added to the environment variable like follows:

echo $PYTHONPATH

TFRecord generation

TFODA needs information how the training and validation data is structured. This information is provided in TFRecord format.

Instructions on how to generate these records can be found in the generate_tfrecords.ipynb Jupyter Notebook. This notebook needs to be placed in the models/research/ directory within the TF models repository.

  1. Place generate_tfrecords.ipynb in models/research/
  2. Start JupyterLab in that directory: jupyter notebook
  3. Open JupyterLab. By default it can be found here: http://localhost:8888
  4. Open generate_tfrecords.ipynb and follow the instructions to generate the records
  5. Place the generated records train.record and val.record like described in the folder structure

Folder structure

The folder structure needs to match the paths set in generate_tfrecords.ipynb.

  1. Place the images for training in detection_training/images_train/ and the images for validation in detection_training/images_val/.
  2. Place all other files in detection_training/
detection_training/  
├── images_train/  
│   ├── train_image1.jpg  
│   ├── train_image2.jpg  
│   ├── train_image3.jpg
│   └── ... 
├── images_val/  
│   ├── val_image1.jpg  
│   ├── val_image2.jpg  
│   ├── val_image3.jpg
│   └── ...  
├── output/
├── train.record
├── val.record  
├── labels_train.csv
├── labels_val.csv
└── ssd_mobilenet_v2.config

Prepare config file

The training configuration must be provided in form of a ssd_mobilenet_v2.config file. Various parameters like the model architecture, image size or the learning rate can be set. You can take a provided config file or have a look at the TFODA sample config files.

You can find the .config file used for this work in the current folder named as ssd_mobilenet_v2.

Start the training

The start command needs to be executed from the models/research/ directory in the TF models repository. The paths need to match the ones defined in generate_tfrecords.ipynb. The number of training steps can be adjusted with num_train_steps.

python3 object_detection/model_main.py \
    --pipeline_config_path=/home/jetbot/Documents/detection_training/ssd_mobilenet_v2.config \
    --model_dir=/home/jetbot/Documents/detection_training/output \
    --num_train_steps=249999 \
    --sample_1_of_n_eval_examples=1 \
    --alsologtostderr

This will start to generate checkpoints in the detection_training/output/ directory. Checkpoints will allow you to stop and restart the training from that checkpoint with the command above. Once the defined number of training steps have been reached the training cannot be restarted again unless the number of steps will be increased.

Observe the training

The training progress can be observed with Tensorboard.

  1. Navigate to the detection_training/output/ directory
  2. Start Tensorboard: tensorboard --logdir=./
  3. Open http://localhost:6006/

Export model graph

Once the training is finished you can find the generated checkpoints in detection_training/output/. To export the model you can run the command below from the models/research/ directory within the TF models repository.

python3 object_detection/export_inference_graph.py \
    --input_type=image_tensor \
    --pipeline_config_path=/home/jetbot/Documents/detection_training/ssd_mobilenet_v2.config \
    --trained_checkpoint_prefix=/home/jetbot/Documents/detection_training/output/model.ckpt-249999 \
    --output_directory=/home/jetbot/Documents/detection_training/output/export

This will generate a saved_model/saved_model.pb and a frozen_inference_graph.pb in the export directory.

Your model is now ready to be deployed.

References