Tensorflow_TFRecords_Estimator_Pipeline

A pipeline/template for

• Converting dataset to TFRecords.
• Training and evaluating multi-class image classifier using custom tensorflow estimator.

Requirements

Tensorflow >= 1.4.0

Setup Environment

# Virtual environment (optional)
sudo apt install -y virtualenv

# Tensorflow (optional)
sudo apt-get install python3-pip python3-dev python-virtualenv # for Python 3.n
virtualenv --system-site-packages -p python3 tensorflow170_py35_gpu # for Python 3.n with GPU
source tensorflow170_py35_gpu/bin/activate
easy_install -U pip
pip3 install --upgrade tensorflow-gpu # for Python 3.n and GPU

# Dependencies
pip install matplotlib
pip install bunch
pip install pudb
pip install tqdm

Dataset

Download knifey-spoony dataset

cd scripts
./download_dataset_knifey_spoony.sh

Train and evaluate

./run.sh

For image classification on new dataset

• Place the new dataset inside datasets folder. Images of each class should be in be in different folder.

Example:

datasets
  knifey_spoony_vanilla
    train
      forky
      knifey
      spoony
    test
      forky
      knifey
      spoony

• Modify configs/config_knifey_spoony.json "labels", "dataset_path_train" and "dataset_path_test" fields.

• Modify models/model_knifey_spoony.py model_fn() as per requirement.

• ./run.sh

Acknowledgement

https://github.com/Hvass-Labs/TensorFlow-Tutorials/blob/master/18_TFRecords_Dataset_API.ipynb

https://github.com/MrGemy95/Tensorflow-Project-Template