Skip to content

End2End Modelの作成

モデルと学習済みデータの保存

1
2
3
4
5
def model_save(model_json,model_h5):
    json_model = model.to_json()
    with open(model_json, "w") as f:
        f.write(json_model)
    model.save_weights(model_h5) 

model_save("model.json", "model.h5")

ここまでのソース

  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
import csv
from matplotlib import pyplot as plt
import cv2
import numpy as np
from sklearn.utils import shuffle
from sklearn.model_selection import train_test_split
from keras.models import *
from keras.layers import *
from keras.optimizers import Adam

def load_data():
    folder_path = ''
    center_img = []
    direction_handle = []

    f = open(folder_path + 'driving_log.csv', 'r')

    reader = csv.reader(f)
    header = next(reader)
    for row in reader:
        img_name = row[0]
        img_path = folder_path  +  img_name
        img = plt.imread(img_path)
        center_img.append(img)
        direction_handle.append(float(row[3]))
    f.close()

    return (center_img, direction_handle)

def draw_img(center_img, direction_handle, pos=0):
    g_rows = 3
    g_cols = 3
    size = g_rows * g_cols
    fig, axs = plt.subplots(ncols=g_rows, nrows=size/g_rows)

    for h in range(g_rows):
        for i in range(g_cols):
            axs[h][i].imshow(center_img[pos + i * h], cmap=plt.cm.gray_r,) 
            axs[h][i].set_title(str(direction_handle[pos + h * i]))
    plt.show()

def process_img(center_img):
    new_center = []
    for img in center_img:
        crop_img = img[60:140, 0:320, :]
        resize_img = cv2.resize((cv2.cvtColor(crop_img, cv2.COLOR_RGB2XYZ))[:,:,2],(40,10))
        new_center.append(resize_img)

    return new_center    

def grow_data(center_img, direction_handle, delta):
    new_center = []
    new_direction = []
    for i in range(len(center_img)):
        for j in range(3):
            new_center.append(center_img[i])
            if j == 0:
                new_direction.append(direction_handle[i])
            elif j == 1:
                new_direction.append(direction_handle[i] + float(delta))
            elif j == 2:
                new_direction.append(direction_handle[i] - float(delta))
    return (new_center,  new_direction)

def end2end_model():
    model = Sequential()
    model.add(Lambda(lambda x: x/127.5 - 1.,input_shape=(rows,cols,1)))    

    model.add(Convolution2D(8, 3, 3, init='normal',border_mode='valid'))
    model.add(Activation('relu'))
    model.add(MaxPooling2D((2,2),border_mode='valid'))

    model.add(Convolution2D(8, 3, 3,init='normal',border_mode='valid'))
    model.add(Activation('relu'))
    model.add(MaxPooling2D((2,2),border_mode='valid'))

    model.add(Dropout(0.2))
    model.add(Flatten())
    model.add(Dense(50))
    model.add(Activation('relu'))
    model.add(Dense(1))

    model.summary()

    return model

def model_save(model_json,model_h5):
    json_model = model.to_json()
    with open(model_json, "w") as f:
        f.write(json_model)
    model.save_weights(model_h5) 

if __name__ == '__main__':
    (center_img, direction_handle) = load_data()

    delta = 0.2
    (center_img, direction_handle) = grow_data(center_img, direction_handle, delta)

    center_data = process_img(center_img)

    draw_img(center_data, direction_handle, 50)

    center_data  = np.array(center_data).astype('float32')
    direction_handle = np.array(direction_handle).astype('float32')

    center_data, direction_handle = shuffle(center_data, direction_handle)

    train_center, test_center, train_direction, test_direction = train_test_split(center_data, direction_handle, random_state=0, test_size=0.1)

    rows = 10
    cols = 40
    train_center = train_center.reshape(train_center.shape[0], rows, cols, 1)
    test_center = test_center.reshape(test_center.shape[0], rows, cols, 1)

    batch = 128
    epoc = 10
    adam = Adam(lr=0.001, beta_1=0.9, beta_2=0.999, epsilon=1e-08, decay=0.0)

    model = end2end()
    model.compile(loss='mean_squared_error',optimizer='adam',metrics=['accuracy'])

    history = model.fit(train_center, train_direction,batch_size=batch, nb_epoch=epoc,verbose=1, validation_data=(test_center, test_direction))

    model_save("./model.json", "./model.h5")