/tf_detection_api

Implement Tensorflow Object Detection API on custom dataset in one bash command.

Primary LanguagePythonApache License 2.0Apache-2.0

Implement Tensorflow Object Detection API on Custom Dataset

Ubuntu16 Anaconda3 Python37 tf15 required tf20 not supported
GitHub Workflow Status (branch) GitHub issues GitHub closed issues

cat and dog

Tensorflow Object Detection API provides a collection of detection models pre-trained on the COCO dataset, the Kitti dataset, the Open Images dataset,etc. These models can be useful for out-of-the-box inference or initializing your models when training on custom datasets.

Typically, the following steps are required to train on a custom dataset:

  1. Installation.
    • Anaconda python=3.7 (optional)
    • tensorflow 1.15
    • Required python packages (i.e., requirements.txt)
    • Clone Tensorflow model garden
    • Download and install protobuf
  2. Prepare custom dataset
    • Prepares images (.jpg format)
    • Make annotations (.xml format) for each image
    • Combine annotation .xml files into a .csv file for train and test set
    • Create label map file .pbtxt
    • Generate tf records from such datasets
  3. Download pre-trained models
  4. Configure training pipeline (e.g., edit pipeline.config file)
  5. Prepare options and run ```python object_detection/model_main.py```

This repos aims to process all of the required steps mentioned above in one bash command

✔️ Requirements:

  • Anaconda3
  • Python 3.7

⚡ Usage

bash main.sh --config_file {PATH_TO_CONFIG_FILE} --install

❓ Help

main.sh [--install] [--config_file str] [--env str]

Implement tensorflow (tf-gpu 1.15.0) Object Detection API on custom Dataset.

Requirements: Anaconda3, python 3.7

Args:
    -h|--help           Show this help message
    --install           Install conda env, pip requirements, and tf models API.
                        Default: False
    --config_file  str  Path to config file.
                        Default: /tf_detection_api/config.json
    --env          str  Conda environment name.
                        Default: tf1Detection

📄 Config file

The contents of the config .json file is as belows:

{
    "image_dir": "PATH_TO_IMAGES_DIRECTORY",
    "anno_dir": "PATH_TO_ANNOTATIONS_DIRECTORY",
    "id":[1,2],
    "name":["cat","dog"],
    "pretrained_model": "ssd_mobilenet_v1_coco",
    "train_ratio":0.75
}

  • image_dir: absolute path to the local directory that contains all of the images (both train and test sets). If you plan to download images from google search, Selenium could be a good choice to automate this process. Detailed isntructions can be found in [4]

  • anno_dir: absolute path to annotations directory that contains .xml annotaiton files. labelImg is a nice tool to generate the annotation .xml file from input images. Details can be found in [5]

  • id: list of category IDs from your custom dataset. 📝 Note that the id starts from 1 because id=0 is used for background as default.

  • name: list of category names from custom dataset

  • pretrained_model: pre-trained model name from tensorflow model garden. Full list of pretrained models

model zoo

  • train_ratio: Ratio to split train and test sets. Default: 0.75

📋 References

[1] Tensorflow model garden installation https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/installation.md

[2] Setup for custom dataset https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/using_your_own_dataset.md

[3] Run the traning job https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/running_locally.md

[4] Search and Download image from google with Python and Selenium https://towardsdatascience.com/image-scraping-with-python-a96feda8af2d

[5] Label images with labelImg tool https://github.com/tzutalin/labelImg

Licence

License