spinenet/dataloader.py at master · DanielWeidinger/spinenet · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
import tensorflow as tf
import numpy as np
import matplotlib.image as ilib
import matplotlib.pyplot as plt
import os
from sklearn.externals import joblib
import cv2

from augmentation import color, flip, rotate, zoom

class DataLoader():

    def __init__(self, params):
        self.params = params

        if not (os.path.exists(params.save_path)):
            self.preprocess()

        self.load_data()


    def preprocess(self):
        data = []
        label = []

        for path, _, files in os.walk(self.params.data_path):
            for file in files:
                ex = os.path.splitext(file)[1]
                if(ex != '.txt' and ex != '.sh'):
                    fpath = path + '/' + file

                    try:
                        pic = ilib.imread(fpath)#[..., :3]
                        if pic.shape[-1] == 4:
                            pic = pic[..., 1:]
                        pic = (pic/127.5) - 1 #scale into [1, -1] range
                        pic = tf.image.resize(pic, (self.params.img_size, self.params.img_size))
                        data.append(pic)
                        label.append(self.to_one_hot(path.split('/')[-1]))
                    except:
                        print("File corrupted: " + fpath)


        with open(self.params.save_path,"wb") as f:
            joblib.dump([data, label], f, protocol=2)

    def load_data(self):

        print('Start loading')

        with open(self.params.save_path,"rb") as f:
           [self.raw_data, self.raw_label] = joblib.load(f)

        dataset = tf.data.Dataset.from_tensor_slices((
            self.raw_data,
            self.raw_label
        )).shuffle(self.params.shuffle_buffer_size)
        all_data = len(self.raw_data)

        train_size = int(all_data * (100 - self.params.val_percentage)/100)

        val_size = all_data - train_size

        self.train = dataset.take(train_size)
        dataset.skip(train_size)

        # Add augmentations
        augmentations = [flip, color, zoom] #no rotation

        # Add the augmentations to the dataset
        for f in augmentations:
            # Apply the augmentation, run 4 jobs in parallel.
            self.train = self.train.map(f, num_parallel_calls=4)

        #TODO lower augmentaion chance

        self.train = self.train.map(lambda x, y: (tf.clip_by_value(x, 0, 1), y))

        self.val = dataset.take(val_size)

        print(f'Dataset Loaded. Features: {all_data}')
        print(f'Train {train_size}. Val: {val_size}')

    def get_train(self):
        return self.train.batch(self.params.batch_size).prefetch(tf.data.experimental.AUTOTUNE)

    def get_val(self):
        return self.val.batch(self.params.batch_size)

    def get_raw_data(self):
        return tf.data.Dataset.from_tensor_slices((
            self.raw_data,
            self.raw_label
        )).shuffle(self.params.shuffle_buffer_size)

    def to_one_hot(self, label):
        result = np.zeros(14)

        poses = get_poses()

        result[poses.get(label)] = 1

        return result

def get_poses():
    return {
            'bridge': 0,
            'camel': 1,
            'chair': 2,
            'chaturanga_dandasana': 3,
            'cobra': 4,
            'cow': 5,
            'dog': 6,
            'half_moon': 7,
            'plank': 8,
            'tree': 9,
            'triangle': 10,
            'warrior_I': 11,
            'warrior_II': 12,
            'warrior_III': 13
        }

def get_Idx():
    return {
            0: 'bridge',
            1: 'camel',
            2: 'chair',
            3: 'chaturanga_dandasana',
            4: 'cobra',
            5: 'cow',
            6: 'dog',
            7: 'half_moon',
            8: 'plank',
            9: 'tree',
            10: 'triangle',
            11: 'warrior_I',
            12: 'warrior_II',
            13: 'warrior_III'
        }