第25章自动驾驶汽车实例–使汽车在两条白线间行驶

文章目录

步骤1：获取驾驶数据
步骤2：拆分数据
步骤3：增强训练数据
步骤4：创建驾驶模式
步骤5：训练模型
步骤6：评估性能
后续可进一步完善

本文将介绍如何使用 Keras + Tensorflow1.6，python为3.6 创建卷积神经网络，并对其进行训练，以使得车辆保持在两条白线之间,如下图：

import os
import urllib.request
import pickle
%matplotlib inline
import matplotlib
from matplotlib.pyplot import imshow

import os

import urllib.request

import pickle

%matplotlib inline

import matplotlib

from matplotlib.pyplot import imshow

下载数据

#downlaod driving data (450Mb) 
data_url = 'https://s3.amazonaws.com/donkey_resources/indoor_lanes.pkl'
file_path, headers = urllib.request.urlretrieve(data_url)
print(file_path)

#downlaod driving data (450Mb)

data_url = 'https://s3.amazonaws.com/donkey_resources/indoor_lanes.pkl'

file_path, headers = urllib.request.urlretrieve(data_url)

print(file_path)

备注：由于下载经常出现超时情况，大家可以先下载该文件，然后放到指定目录。

file_path='./autocar/indoor_lanes.pkl'

1	file_path='./autocar/indoor_lanes.pkl'

步骤1：获取驾驶数据

数据集由 ~7900 个图像和手动开车时收集的转向角组成。大约三分之二的图像与线之间的汽车。另外三分之一的车开始偏离航线，并且驶回线路之间。
数据集由 2 个 pickled 数组组成。X 是图像阵列，Y 是相应转向角度的阵列。

#抽取数据
with open(file_path, 'rb') as f:
    X, Y = pickle.load(f)

print('X.shape: ', X.shape)
print('Y.shape: ', Y.shape)
imshow(X[0])

#抽取数据

with open(file_path, 'rb') as f:

X, Y = pickle.load(f)

print('X.shape: ', X.shape)

print('Y.shape: ', Y.shape)

imshow(X[0])

打印结果
X.shape: (7892, 120, 160, 3)
Y.shape: (7892,)

步骤2：拆分数据

在这里，对数据进行洗牌（shuffle），并将数据分成三部分。训练数据将用于训练我们的驾驶模型，使用验证数据避免过度拟合模型，测试数据用于测试我们的模型。

import numpy as np
#用shuffle把数据打乱
def unison_shuffled_copies(X, Y):
    assert len(X) == len(Y)
    p = np.random.permutation(len(X))
    return X[p], Y[p]

shuffled_X, shuffled_Y = unison_shuffled_copies(X,Y)

len(shuffled_X)

import numpy as np

#用shuffle把数据打乱

def unison_shuffled_copies(X, Y):

assert len(X) == len(Y)

p = np.random.permutation(len(X))

return X[p], Y[p]

shuffled_X, shuffled_Y = unison_shuffled_copies(X,Y)

len(shuffled_X)

打印结果
7892

test_cutoff = int(len(X) * .8) # 80% of data used for training
val_cutoff = test_cutoff + int(len(X) * .1) # 10% of data used for validation and test data 

train_X, train_Y = shuffled_X[:test_cutoff], shuffled_Y[:test_cutoff]
val_X, val_Y = shuffled_X[test_cutoff:val_cutoff], shuffled_Y[test_cutoff:val_cutoff]
test_X, test_Y = shuffled_X[val_cutoff:], shuffled_Y[val_cutoff:]

len(train_X) + len(val_X) + len(test_X)

test_cutoff = int(len(X) * .8) # 80% of data used for training

val_cutoff = test_cutoff + int(len(X) * .1) # 10% of data used for validation and test data

train_X, train_Y = shuffled_X[:test_cutoff], shuffled_Y[:test_cutoff]

val_X, val_Y = shuffled_X[test_cutoff:val_cutoff], shuffled_Y[test_cutoff:val_cutoff]

test_X, test_Y = shuffled_X[val_cutoff:], shuffled_Y[val_cutoff:]

len(train_X) + len(val_X) + len(test_X)

打印结果
7892

步骤3：增强训练数据

为了加倍我们的训练数据并防止转向偏差，我们翻转每个图像和转向角并将其添加到数据集中。还有其他的方法来增加使用翻译和假阴影驾驶数据，但我没有使用这些自动驾驶仪。

X_flipped = np.array([np.fliplr(i) for i in train_X])
Y_flipped = np.array([-i for i in train_Y])
train_X = np.concatenate([train_X, X_flipped])
train_Y = np.concatenate([train_Y, Y_flipped])
len(train_X)

X_flipped = np.array([np.fliplr(i) for i in train_X])

Y_flipped = np.array([-i for i in train_Y])

train_X = np.concatenate([train_X, X_flipped])

train_Y = np.concatenate([train_Y, Y_flipped])

len(train_X)

打印结果
12626

步骤4：创建驾驶模式

这种驾驶模式将是一个端到端的神经网络，接受图像阵列作为输入，并输出-90（左）和90（右）之间的转向角。要做到这一点，我们将使用一个完全连接图层的3层卷积网络。该模型基于 Otavio 的 Carputer，但不产生油门值输出，不使用过去的转向值作为模型的输入，并且使用较少的卷积层。

from tensorflow.python.keras.models import Model, load_model
from tensorflow.python.keras.layers import Input, Convolution2D, MaxPooling2D, Activation, Dropout, Flatten, Dense
from tensorflow.python.keras import backend as K

from tensorflow.python.keras.models import Model, load_model

from tensorflow.python.keras.layers import Input, Convolution2D, MaxPooling2D, Activation, Dropout, Flatten, Dense

from tensorflow.python.keras import backend as K

使用 TensorFlow 后端。
查看目前image数据格式，是tensorflow格式（channels_last）还是theano格式（channels_first）

K.image_data_format()

1	K.image_data_format()

'channels_last'

如果不是tensorflow格式，可以通过K.set_image_data_format('channels_last')进行修改，或修改~/.keras/keras.json文件。前者只对当前会话有效，后者修改，将永久有效。

img_in = Input(shape=(120, 160, 3), name='img_in')
angle_in = Input(shape=(1,), name='angle_in')

x = Convolution2D(8, kernel_size=(3, 3),strides=(1,1))(img_in)
x = Activation('relu')(x)
x = MaxPooling2D(pool_size=(2, 2))(x)

x = Convolution2D(16, kernel_size=(3, 3),strides=(1,1))(x)
x = Activation('relu')(x)
x = MaxPooling2D(pool_size=(2, 2))(x)

x = Convolution2D(32, kernel_size=(3, 3),strides=(1,1))(x)
x = Activation('relu')(x)
x = MaxPooling2D(pool_size=(2, 2))(x)

merged = Flatten()(x)

x = Dense(256)(merged)
x = Activation('linear')(x)
x = Dropout(.2)(x)

angle_out = Dense(1, name='angle_out')(x)

model = Model(inputs=[img_in], outputs=[angle_out])
model.compile(optimizer='adam', loss='mean_squared_error')
model.summary()

img_in = Input(shape=(120, 160, 3), name='img_in')

angle_in = Input(shape=(1,), name='angle_in')

x = Convolution2D(8, kernel_size=(3, 3),strides=(1,1))(img_in)

x = Activation('relu')(x)

x = MaxPooling2D(pool_size=(2, 2))(x)

x = Convolution2D(16, kernel_size=(3, 3),strides=(1,1))(x)

x = Activation('relu')(x)

x = MaxPooling2D(pool_size=(2, 2))(x)

x = Convolution2D(32, kernel_size=(3, 3),strides=(1,1))(x)

x = Activation('relu')(x)

x = MaxPooling2D(pool_size=(2, 2))(x)

merged = Flatten()(x)

x = Dense(256)(merged)

x = Activation('linear')(x)

x = Dropout(.2)(x)

angle_out = Dense(1, name='angle_out')(x)

model = Model(inputs=[img_in], outputs=[angle_out])

model.compile(optimizer='adam', loss='mean_squared_error')

model.summary()

打印结果
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
img_in (InputLayer) (None, 120, 160, 3) 0
_________________________________________________________________
conv2d_1 (Conv2D) (None, 118, 158, 8) 224
_________________________________________________________________
activation_1 (Activation) (None, 118, 158, 8) 0
_________________________________________________________________
max_pooling2d_1 (MaxPooling2 (None, 59, 79, 8) 0
_________________________________________________________________
conv2d_2 (Conv2D) (None, 57, 77, 16) 1168
_________________________________________________________________
activation_2 (Activation) (None, 57, 77, 16) 0
_________________________________________________________________
max_pooling2d_2 (MaxPooling2 (None, 28, 38, 16) 0
_________________________________________________________________
conv2d_3 (Conv2D) (None, 26, 36, 32) 4640
_________________________________________________________________
activation_3 (Activation) (None, 26, 36, 32) 0
_________________________________________________________________
max_pooling2d_3 (MaxPooling2 (None, 13, 18, 32) 0
_________________________________________________________________
flatten_1 (Flatten) (None, 7488) 0
_________________________________________________________________
dense_1 (Dense) (None, 256) 1917184
_________________________________________________________________
activation_4 (Activation) (None, 256) 0
_________________________________________________________________
dropout_1 (Dropout) (None, 256) 0
_________________________________________________________________
angle_out (Dense) (None, 1) 257
=================================================================
Total params: 1,923,473
Trainable params: 1,923,473
Non-trainable params: 0

步骤5：训练模型

我们已经学会了很难的方法，即使这一切都是完美的，如果你没有正确地训练，你的自动驾驶仪将无法工作。我遇到的最大的问题是过度适应模型，以至于在很少的情况下都不能正常工作。这里是 2 个 Keras回调，将节省您的时间。警告 - 如果仅使用CPU，则需要较长时间，这里我采用GPU进行训练

import os
from tensorflow.python.keras import callbacks

model_path = os.path.expanduser('~/best_autopilot.hdf5')

#Save the model after each epoch if the validation loss improved.
save_best = callbacks.ModelCheckpoint(model_path, monitor='val_loss', verbose=1, 
                                     save_best_only=True, mode='min')

#stop training if the validation loss doesn't improve for 5 consecutive epochs.
early_stop = callbacks.EarlyStopping(monitor='val_loss', min_delta=0, patience=5, 
                                     verbose=0, mode='auto')

callbacks_list = [save_best, early_stop]

import os

from tensorflow.python.keras import callbacks

model_path = os.path.expanduser('~/best_autopilot.hdf5')

#Save the model after each epoch if the validation loss improved.

save_best = callbacks.ModelCheckpoint(model_path, monitor='val_loss', verbose=1,

save_best_only=True, mode='min')

#stop training if the validation loss doesn't improve for 5 consecutive epochs.

early_stop = callbacks.EarlyStopping(monitor='val_loss', min_delta=0, patience=5,

verbose=0, mode='auto')

callbacks_list = [save_best, early_stop]

因时间关系，这里只进行 10 次迭代（epochs），训练时间2分钟左右，比较快。

model.fit(train_X, train_Y, batch_size=64, epochs=10, validation_data=(val_X, val_Y), callbacks=callbacks_list)

1	model.fit(train_X, train_Y, batch_size=64, epochs=10, validation_data=(val_X, val_Y), callbacks=callbacks_list)

运行结果
Train on 12626 samples, validate on 789 samples
Epoch 1/10
12608/12626 [============================>.]12608/12626 [============================>.] - ETA: 0s - loss: 41288.4184
Epoch 00001: val_loss improved from inf to 656.89127, saving model to /home/wumg/best_autopilot.hdf5
12626/12626 [==============================]12626/12626 [==============================] - 22s 2ms/step - loss: 41230.2161 - val_loss: 656.8913

Epoch 2/10
12608/12626 [============================>.]12608/12626 [============================>.] - ETA: 0s - loss: 550.5565
Epoch 00002: val_loss improved from 656.89127 to 543.14232, saving model to /home/wumg/best_autopilot.hdf5
12626/12626 [==============================]12626/12626 [==============================] - 15s 1ms/step - loss: 550.1622 - val_loss: 543.1423

Epoch 3/10
12608/12626 [============================>.]12608/12626 [============================>.] - ETA: 0s - loss: 476.8377
Epoch 00003: val_loss improved from 543.14232 to 417.43844, saving model to /home/wumg/best_autopilot.hdf5
12626/12626 [==============================]12626/12626 [==============================] - 15s 1ms/step - loss: 476.6054 - val_loss: 417.4384

Epoch 4/10
12608/12626 [============================>.]12608/12626 [============================>.] - ETA: 0s - loss: 420.1108
Epoch 00004: val_loss improved from 417.43844 to 416.97928, saving model to /home/wumg/best_autopilot.hdf5
12626/12626 [==============================]12626/12626 [==============================] - 15s 1ms/step - loss: 420.0584 - val_loss: 416.9793

Epoch 5/10
12608/12626 [============================>.]12608/12626 [============================>.] - ETA: 0s - loss: 399.4089
Epoch 00005: val_loss did not improve
12626/12626 [==============================]12626/12626 [==============================] - 15s 1ms/step - loss: 399.0109 - val_loss: 420.7120

Epoch 6/10
12608/12626 [============================>.]12608/12626 [============================>.] - ETA: 0s - loss: 373.5299
Epoch 00006: val_loss improved from 416.97928 to 411.31800, saving model to /home/wumg/best_autopilot.hdf5
12626/12626 [==============================]12626/12626 [==============================] - 15s 1ms/step - loss: 373.7999 - val_loss: 411.3180

Epoch 7/10
12608/12626 [============================>.]12608/12626 [============================>.] - ETA: 0s - loss: 356.3707
Epoch 00007: val_loss improved from 411.31800 to 384.65238, saving model to /home/wumg/best_autopilot.hdf5
12626/12626 [==============================]12626/12626 [==============================] - 15s 1ms/step - loss: 356.4031 - val_loss: 384.6524

Epoch 8/10
12608/12626 [============================>.]12608/12626 [============================>.] - ETA: 0s - loss: 338.6780
Epoch 00008: val_loss did not improve
12626/12626 [==============================]12626/12626 [==============================] - 15s 1ms/step - loss: 338.4380 - val_loss: 435.1508

Epoch 9/10
12608/12626 [============================>.]12608/12626 [============================>.] - ETA: 0s - loss: 325.1715
Epoch 00009: val_loss improved from 384.65238 to 348.49219, saving model to /home/wumg/best_autopilot.hdf5
12626/12626 [==============================]12626/12626 [==============================] - 15s 1ms/step - loss: 325.0649 - val_loss: 348.4922

Epoch 10/10
12608/12626 [============================>.]12608/12626 [============================>.] - ETA: 0s - loss: 304.1439
Epoch 00010: val_loss did not improve
12626/12626 [==============================]12626/12626 [==============================] - 15s 1ms/step - loss: 304.2825 - val_loss: 349.3760

步骤6：评估性能

我们可以通过绘制预测值和实际值来检查我们的模型预测是否合理。第一个图表显示我们的测试数据中存在一个学习关系（在训练期间,部分测试数据我们的模型没有注意到）。

import pandas as pd

model = load_model(model_path)
test_P = model.predict(test_X)
test_P = test_P.reshape((test_P.shape[0],)) 


df = pd.DataFrame({'predicted':test_P, 'actual':test_Y})
ax = df.plot.scatter('predicted', 'actual')

import pandas as pd

model = load_model(model_path)

test_P = model.predict(test_X)

test_P = test_P.reshape((test_P.shape[0],))

df = pd.DataFrame({'predicted':test_P, 'actual':test_Y})

ax = df.plot.scatter('predicted', 'actual')

第二张图，使用包含训练数据的非混洗（unshuffled）数据，来显示预测角度紧跟实际转向角度。

P = model.predict(X[:700])
#predict outputs nested arrays so we need to reshape to plot.
P = P.reshape((P.shape[0],)) 

ax = pd.DataFrame({'predicted':P, 'actual':Y[:700]}).plot()
ax.set_ylabel("steering angle")

P = model.predict(X[:700])

#predict outputs nested arrays so we need to reshape to plot.

P = P.reshape((P.shape[0],))

ax = pd.DataFrame({'predicted':P, 'actual':Y[:700]}).plot()

ax.set_ylabel("steering angle")

后续可进一步完善

改善模型，这个模型是纯粹（navie）的，因为它不使用过去的值来帮助预测未来。我们可以通过将过去的转向角度作为模型的输入来进行试验，添加一个递归层，或者只是改变卷积层的结构。
添加更多数据，随着我们添加更多驾驶数据，此模型将会得到改进。预测油门，输出目前自动驾驶仪只能转向并保持恒定的速度。一个更复杂的模型将加速在直路上，并在路缘之前放缓。

Python技术交流与分享

分享技术平台

步骤1：获取驾驶数据

步骤2：拆分数据

步骤3：增强训练数据

步骤4：创建驾驶模式

步骤5：训练模型

步骤6：评估性能

后续可进一步完善