-
Notifications
You must be signed in to change notification settings - Fork 0
/
model.py
219 lines (164 loc) · 6.53 KB
/
model.py
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
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
"""
Model definition
"""
import json
import csv
import numpy as np
import tensorflow as tf
import pandas
import cv2
from keras.layers.core import Activation, Flatten, Dense, Dropout, Lambda
from keras.layers.convolutional import Convolution2D, Cropping2D
from keras.models import Sequential
from keras.optimizers import Adam
from keras.utils.visualize_util import plot
from sklearn.model_selection import train_test_split
from sklearn.utils import shuffle
tf.python.control_flow_ops = tf
# Custom utilities and constants
from utils import IMAGE_SHAPE
from utils import flip, load_img, normalize, show_img, translate, perturb_brightness, add_shadow, do_flip
from utils import show_img, plot_angle_dist, plot_loss
# HYPER PARAMETERS
STEERING_OFFSET = 0.20 # Angle offset used for left/right camera images.
EPOCHS = 12 # Epochs used for training.
BATCH_SIZE = 256 # Batch sizes for training.
DATA_PATH = './data/' # Location of csv file containing training dataset metadata.
LEARNING_RATE = 0.0001 # Adam's optimizer learning rate.
DROPOUT = 0.5 # Model's dropout percentage.
# LOADING METADATA
def data_path(p):
"""
Prefixes correctly paths to look for training files.
"""
return DATA_PATH + p.strip()
def load_img_from_csv_line(line):
"""
Extracts an image for a given csv line. Randomly chooses
between left, center or right image.
Returns a tuple of image and respective steering angle.
"""
steering_angle = float(line["steering"])
side = np.random.randint(3)
if side == 0:
img_path = line["left"]
steering_angle += STEERING_OFFSET
elif side == 1:
img_path = line["center"]
else:
img_path = line["right"]
steering_angle -= STEERING_OFFSET
return (load_img(data_path(img_path)), steering_angle)
def process_csv_line(line):
"""
Given a csv line, it loads the image and performs some modifications
with the purpose of equalizing steering angles distribution.
"""
img, angle = load_img_from_csv_line(line)
# image perturbations
img, angle = translate(img , angle)
img = perturb_brightness(img)
img = add_shadow(img)
if do_flip():
img, angle = flip(img, angle)
return img, angle
# data augmentation
def generate_data(metadata, batch_size = 32, bias = 1.0):
"""
"""
samples = len(metadata)
shuffle(metadata)
batch_images = np.zeros( (batch_size, ) + IMAGE_SHAPE )
batch_angles = np.zeros( batch_size )
while 1:
for batch_id in range(batch_size):
csv_line_id = np.random.randint(samples)
csv_line = metadata.iloc[csv_line_id]
keep_sample = False
while not keep_sample:
img, angle = process_csv_line(csv_line)
threshold = np.random.uniform()
if (abs(angle) + bias) >= threshold:
keep_sample = True
batch_images[batch_id] = img
batch_angles[batch_id] = angle
angles.append(angle)
yield shuffle(np.array(batch_images), np.array(batch_angles))
def build_model():
"""
Defining the network architecture.
"""
activation = 'relu'
model = Sequential()
model.add(Cropping2D(cropping=((50, 30), (0, 0)), input_shape=IMAGE_SHAPE))
model.add(Lambda(normalize))
model.add(Convolution2D(24, 5, 5, border_mode='valid', subsample=(2, 2), activation=activation))
model.add(Convolution2D(36, 5, 5, border_mode='valid', subsample=(2, 2), activation=activation))
model.add(Convolution2D(48, 5, 5, border_mode='valid', subsample=(2, 2), activation=activation))
model.add(Convolution2D(64, 3, 3, border_mode='valid', subsample=(1, 1), activation=activation))
model.add(Convolution2D(64, 3, 3, border_mode='valid', subsample=(1, 1), activation=activation))
model.add(Flatten())
model.add(Dense(100, activation=activation))
model.add(Dropout(DROPOUT))
model.add(Dense(50, activation=activation))
model.add(Dropout(DROPOUT))
model.add(Dense(10, activation=activation))
model.add(Dropout(DROPOUT))
model.add(Dense(1))
print(model.summary())
return model
def persist_model(model, plot=False):
"""
We take a model and persist it on disk.
This method produces two files: model.json and model.h5
Optionally, image model.png will be produced with the network
architecture.
"""
model_json = model.to_json()
with open('model.json', 'w') as modelFile:
json.dump(model.to_json(), modelFile)
modelFile.close()
model.save('model.h5')
if plot:
plot(model, to_file='model.png', show_shapes=True, show_layer_names=False)
angles = []
def train(model, metadata_train, metadata_val):
"""
Performs the training of the given model, using
training metadata to produce training data and
metadata validation to produce validation data.
A tuple of lists is returned. Each list represents the loss
at each epoch for training set and validation set respectively.
"""
adam = Adam(lr=LEARNING_RATE, beta_1=0.9, beta_2=0.999, epsilon=1e-08, decay=0.0)
model.compile(optimizer=adam, loss='mse', metrics=['mean_squared_error'])
train_samples = BATCH_SIZE * 40
val_samples = BATCH_SIZE * 4
loss_hist = []
val_loss_hist = []
for it in range(EPOCHS):
bias = 1.0/(it + 1)
epoch_angles = []
training_results = model.fit_generator(
generate_data(metadata_train, batch_size = BATCH_SIZE, bias=bias),
validation_data=generate_data(metadata_val, batch_size = BATCH_SIZE),
nb_val_samples=val_samples,
samples_per_epoch=train_samples,
nb_epoch=1,
verbose=1)
loss_hist.extend(training_results.history['loss'])
val_loss_hist.extend(training_results.history['val_loss'])
plot_angle_dist(angles, "Steering angle distribution")
return loss_hist, val_loss_hist
if __name__ == '__main__':
metadata = pandas.read_csv(data_path("driving_log.csv"), sep=',')
# spliting metadata for training and validation
metadata_train, metadata_val = train_test_split(metadata, test_size=0.2)
# create model and train it.
model = build_model()
loss, val_loss = train(model, metadata_train, metadata_val)
persist_model(model)
# generate image containing plot of training and validation loss.
plot_loss(loss, val_loss)
# to avoid issues explained here: https://github.com/tensorflow/tensorflow/issues/3388
import gc; gc.collect()