Create a dataset of images to use for learning

Introduction

――The total size of the learning image, test image, and padded image this time is about 1.4GB. ――It will take some time to load these images when the learning program is implemented. —— Also, when you run a learning program in another environment, transfer time will occur. ――In addition, it takes time to resize the image and convert from color to gray. ――By creating a dataset that has been resized and converted to gray in advance, we were able to make it about 50MB. --The complete source is here.

Library

--Like last time, I'm using Numpy`` Pillow.

Setting

--The following settings have been added. --The dataset created this time is saved in DATASETS_PATH. --ʻIMG_ROWS ʻIMG_COLS is a resizing of the image size. This time, we will resize it to the size of 28 x 28. --The image size is also referred to in the training model of the post-process.

config.py

DATASETS_PATH = os.path.join(DATA_PATH, 'datasets')

IMG_ROWS, IMG_COLS = 28, 28

Creating a file list

--Create a file list of training images, test images, and padded images. --query is given CLASSES in sequence. --ʻAugmentThe argument is a flag for the availability of padded images. --Last time, the padded image was created for eachquery`` 6000`. If it is not enough, it is an error.

def make_filesets(augment):
    """Creating a file set."""

    filesets = {'train': dict(), 'test': dict(), 'augment': dict()}

    for query in CLASSES:

        train_path = os.path.join(TRAIN_PATH, query)
        test_path = os.path.join(TEST_PATH, query)
        augment_path = os.path.join(AUGMENT_PATH, query)

        if not os.path.isdir(train_path):
            print('no train path: {}'.format(train_path))
            return None
        if not os.path.isdir(test_path):
            print('no test path: {}'.format(test_path))
            return None
        if not os.path.isdir(augment_path):
            print('no augment path: {}'.format(augment_path))
            return None

        train_files = glob.glob(os.path.join(train_path, '*.jpeg'))
        train_files.sort()
        filesets['train'][query] = train_files

        test_files = glob.glob(os.path.join(test_path, '*.jpeg'))
        test_files.sort()
        filesets['test'][query] = test_files

        augment_files = glob.glob(os.path.join(augment_path, '*.jpeg'))
        random.shuffle(augment_files)
        filesets['augment'][query] = augment_files

        if augment and len(augment_files) < AUGMENT_NUM:
            print('less augment num: {}, path: {}'.format(len(augment_files), augment_path))
            return None

    return filesets

Image loading function

--Processes the image based on the full path of the file. --It will be resized according to the configuration file. --Originally, in OpenCV Haar Cascades, it was saved without resizing. Resizing in a later process is more convenient to try in various sizes. ―― LANCZOS takes time, but it resizes with good quality. The default is NEAREST. Speed is prioritized over quality. --Reference: https://pillow.readthedocs.io/en/4.0.x/handbook/concepts.html#filters --Then, convert to grayscale and then to ʻuint8`.

def read_image(filename):
    """Image loading, resizing, gray conversion."""

    image = Image.open(filename)
    image = image.resize((IMG_ROWS, IMG_COLS), Image.LANCZOS)
    image = image.convert('L')
    image = np.array(image, dtype=np.uint8)

    return image

Creating a dataset

--Prepare an array of training images, training labels, test images, and test labels.

def make_datasets(augment, filesets):
    """Creating a dataset."""

    train_images = []
    train_labels = []
    test_images = []
    test_labels = []

--query is given CLASSES in sequence. --Num is given a sequential label. ――For example, at the beginning of CLASSES, the Abe Oto label is like 0. --Determine whether to use the inflated image with ʻaugment. If you want to use it, set only the number described in ʻAUGMENT_NUM to train_files. --The tqdm is also used to read each image. The processing progress is displayed, which is easy to understand. --Give read_image the file path of the image to read the resized and grayscaled image. --At the same time, give a label.

    for num, query in enumerate(CLASSES):
        print('create dataset: {}'.format(query))

        if augment:
            train_files = filesets['augment'][query][:AUGMENT_NUM]
        else:
            train_files = filesets['train'][query]
        test_files = filesets['test'][query]

        for train_file in tqdm.tqdm(train_files, desc='create train', leave=False):
            train_images.append(read_image(train_file))
            train_labels.append(num)
        for test_file in tqdm.tqdm(test_files, desc='create test', leave=False):
            test_images.append(read_image(test_file))
            test_labels.append(num)

--Collect training images, training labels, test images, and test labels as a data set. --DATASET_PATH`` CLASSES ʻIMG_ROWS ʻIMG_COLS Decide the file name of the dataset based on whether or not the padded image is used.

    datasets = ((np.array(train_images), (np.array(train_labels))), (np.array(test_images), (np.array(test_labels))))

    datasets_path = os.path.join(DATASETS_PATH, ','.join(CLASSES))
    os.makedirs(datasets_path, exist_ok=True)
    train_num = AUGMENT_NUM if augment else 0
    datasets_file = os.path.join(datasets_path, '{}x{}-{}.pickle'.format(IMG_ROWS, IMG_COLS, train_num))
    with open(datasets_file, 'wb') as fout:
        pickle.dump(datasets, fout)
    print('save datasets: {}'.format(datasets_file))

――The use of inflated images is switched by the following options.

$ python save_datasets.py

$ python save_datasets.py --augment

--The pickle data set is as follows. --In the original case, a pickle file of train`` test totaling about 148MB to 3.2MB --For padded images, pickle files from ʻaugment`` testtotaling about1433MB to 46MB`

$ du -d1 -h .
115M	./train
 33M	./test
 51M	./datasets
1.4G	./augment


$ ls
3.2M 12 15 23:22 28x28-0.pickle
46M 12 15 22:24 28x28-6000.pickle

in conclusion

――We have created a data set that resized and grayscaled the image data so that it can be easily used from the learning program. ――You can create various data sets by changing the number of inflated images and multiple sizes, and use them while switching by file name. ――Next time, I plan to create a part that reads the dataset from the learning program.