A tool for converting CIFAR-10 and CIFAR-100 datasest into PNG images with additional preprocessing options such as grayscaling.
The tools provided are compatible with CIFAR-10 and CIFAR-100 datasets which contains 32x32 images that are a subset of the 80 Million Tiny Images dataset.
CIFAR-10 dataset contains 60,000 32x32 colour images in 10 classes, with 6,000 images per class. There are 50,000 training images and 10,000 test images. CIFAR-100 is similar to CIFAR-10, except it has 100 classes containing 600 images each. There are 500 training images and 100 testing images per class.
Benchmarks for the CIFAR-10 and CIFAR-100 datasets, and others can be found here.
The training and test datasets are provided in files that contain a 'pickled' object produced with cPickle.
The script loads the datasets into Numpy arrays, and the features are separated from the labels. The data is then converted to images into training and testing directories. The format for the filename is as follows: TYPE_RANDOM_LABEL_LABELCOUNT.png
TYPE: Indicates dataset type, could be eithertrainortestRANDOM: Short randomly generated UUID-style characters e.g.7daa28LABEL: The groundtruth label for the image (between 0-9)LABELCOUNT: The count for how many times a label was seen to easily
This format is useful for quickly extracting information about the dataset and target labels from the filename, while ensuring that each image's filename is unique.
- Python 2.7+
Install the Python dependencies using pip: pip install -r REQUIREMENTS.txt
The training data in CIFAR-10 comes in 5 different 'batch' files, while the testing data comes in a single file. Before starting, ensure that you have the data_batch_1, data_batch_2, data_batch_3, data_batch_4, data_batch_5 and test_batch provided here. The script accepts a folder path as the input directory containing the necessary files. The script also assumes the output directory exists so ensure that you have a designated output directory for the preprocessed images, it will not be created automatically.
To preprocess the training set, use the following:
python src/cifar10.py --dataset train --input_folder /path/to/pickled/files --output_path /path/to/output/training
To preprocess the test set, use the following:
python src/cifar10.py --dataset test --input_folder /path/to/pickled/files --output_path /path/to/output/testing
Note: We assume that the filenames are kept intact from the original dataset. If they have been renamed, the constants can be easily changed inside src/cifar10.py to the appropriate filename.
Unlike CIFAR-10, the CIFAR-100 dataset comes in a single file for the training set and a sinlge file for the test set. Before starting, ensure that you have the train and test sets provided here. The script accepts a folder path as the input directory containing the necessary files. As above, the script assumes the output directory exists so ensure that you have a designated output directory for the preprocessed images.
To preprocess the training set, use the following:
python src/cifar100.py --dataset train --input_file /path/to/train --output_path /path/to/output/training
To preprocess the test set, use the following:
python src/cifar100.py --dataset train --input_file /path/to/test --output_path /path/to/output/testing
The original images are coloured, you may optionally pass the --grayscale parameter to convert the images to grayscale.
You may optionally pass the --logging info parameter to display the progress of the script, which looks like this:
...
[utils.py:120 - preprocess() - INFO] Step #6000: saved train_b9c487_42_65.png
[utils.py:120 - preprocess() - INFO] Step #7000: saved train_8961fc_14_71.png
[utils.py:120 - preprocess() - INFO] Step #8000: saved train_a300ef_91_76.png
[utils.py:120 - preprocess() - INFO] Step #9000: saved train_23da3c_42_93.png
[utils.py:120 - preprocess() - INFO] Step #10000: saved train_686758_48_103.png
[utils.py:120 - preprocess() - INFO] Step #11000: saved train_56e075_64_130.png
[utils.py:120 - preprocess() - INFO] Step #12000: saved train_fa3998_53_121.png
[utils.py:120 - preprocess() - INFO] Step #13000: saved train_378111_17_105.png
...
Learning Multiple Layers of Features from Tiny Images, Alex Krizhevsky, 2009.