1
import random
2
import numpy as np
3
from keras.utils import to_categorical
4
from keras.preprocessing import image
5
from os import listdir
6
from os.path import isdir, join
7

8

9
def load_data(container_path='datasets', folders=['A', 'B', 'C'],
10
              size=2000, test_split=0.2, seed=0):
11
    """
12
    Loads sign language dataset.
13
    """
14
    
15
    filenames, labels = [], []
16

17
    for label, folder in enumerate(folders):
18
        folder_path = join(container_path, folder)
19
        images = [join(folder_path, d)
20
                     for d in sorted(listdir(folder_path))]
21
        labels.extend(len(images) * [label])
22
        filenames.extend(images)
23
    
24
    random.seed(seed)
25
    data = list(zip(filenames, labels))
26
    random.shuffle(data)
27
    data = data[:size]
28
    filenames, labels = zip(*data)
29

30
    
31
    # Get the images
32
    x = paths_to_tensor(filenames).astype('float32')/255
33
    # Store the one-hot targets
34
    y = to_categorical(np.array(labels))  # Use to_categorical directly
35

36
    x_train = np.array(x[:int(len(x) * (1 - test_split))])
37
    y_train = np.array(y[:int(len(x) * (1 - test_split))])
38
    x_test = np.array(x[int(len(x) * (1 - test_split)):])
39
    y_test = np.array(y[int(len(x) * (1 - test_split)):])
40

41
    return (x_train, y_train), (x_test, y_test)
42

43

44
def path_to_tensor(img_path, size):
45
    # loads RGB image as PIL.Image.Image type
46
    img = image.load_img(img_path, target_size=(size, size))
47
    # convert PIL.Image.Image type to 3D tensor
48
    x = image.img_to_array(img)
49
    # convert 3D tensor to 4D tensor 
50
    return np.expand_dims(x, axis=0)
51

52
def paths_to_tensor(img_paths, size=50):
53
    list_of_tensors = [path_to_tensor(img_path, size) for img_path in img_paths]
54
    return np.vstack(list_of_tensors)
55

56