Awesome

create_tfrecords

A simpler way of preparing large-scale image dataset by generalizing functions from TensorFlow-slim.

Requirements

1. Python 2.7.x
2. TensorFlow >= 0.12

NOTE: If you want to run this program on Python 3, clone and run git checkout python-3.0 for the Python 3 branch instead.

Usage

$python create_tfrecord.py --dataset_dir=/path/to/dataset/ --tfrecord_filename=dataset_name

#Example: python create_tfrecord.py --dataset_dir=/path/to/flowers --tfrecord_filename=flowers
#Note that the dataset_dir should be the folder that contains the root directory and not the root directory itself.

Arguments

Required arguments:

• dataset_dir (string): The directory to your dataset that is arranged in a structured way where your subdirectories keep classes of your images.

For example:

flowers\
    flower_photos\
        tulips\
            ....jpg
            ....jpg
            ....jpg
        sunflowers\
            ....jpg
        roses\
            ....jpg
        dandelion\
            ....jpg
        daisy\
            ....jpg

  Note: Your dataset_dir should be /path/to/flowers and not /path/to/flowers/flowers_photos

• tfrecord_filename (string): The output name of your TFRecord files.

Optional Arguments

• validation_size (float): The proportion of the dataset to be used for evaluation.

• num_shards (int): The number of shards to split your TFRecord files into.

• random_seed (int): The random seed number for repeatability.

Complete Guide

For a complete guide, please visit here.