-
Notifications
You must be signed in to change notification settings - Fork 1
/
train.py
182 lines (144 loc) · 6.28 KB
/
train.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
import os
import argparse
import datetime
import cv2
import numpy as np
import sklearn.model_selection
from tensorflow import keras
from models.nvidia import NvidiaModel
from models.comma import CommaModel
from models.tiny import TinyModel
def training_generator(image_paths, steering_angles, preproc, batch_size):
"""
Generate training data.
"""
fourth_batch_size = batch_size // 4
slices_per_epoch = len(image_paths) // batch_size
datagen = keras.preprocessing.image.ImageDataGenerator(
rotation_range=10,
width_shift_range=0.1,
height_shift_range=0.1,
shear_range=0.1,
zoom_range=0.1,
channel_shift_range=.1,
fill_mode='nearest')
training_data_slice = 0
while True:
slice_begin = training_data_slice * fourth_batch_size
slice_end = slice_begin + fourth_batch_size
# original frames
orig_images = [preproc(cv2.imread(path)) for path in image_paths[slice_begin:slice_end]]
orig_steers = steering_angles[slice_begin:slice_end]
# original frames mirrored
orig_flip_images = [cv2.flip(x, 1) for x in orig_images]
orig_flip_steers = [-x for x in orig_steers]
# generated frames
gen = datagen.flow(
x=np.stack(orig_images, axis=0),
y=np.stack(orig_steers, axis=0),
batch_size=fourth_batch_size,
shuffle=False)
gen_images, gen_steers = next(gen)
# generated frames mirrored
gen_flip_images = [cv2.flip(x, 1) for x in gen_images]
gen_flip_steers = [-x for x in gen_steers]
images = np.concatenate((orig_images, orig_flip_images, gen_images, gen_flip_images), axis=0)
steers = np.concatenate((orig_steers, orig_flip_steers, gen_steers, gen_flip_steers), axis=0)
yield images, steers
training_data_slice = (training_data_slice + 1) % slices_per_epoch
def validation_generator(image_paths, steering_angles, preproc, batch_size):
"""
Generate cross validation data.
"""
while True:
indices = [np.random.choice(len(image_paths)) for x in range(batch_size)]
images = [preproc(cv2.imread(image_paths[i])) for i in indices]
steers = [steering_angles[i] for i in indices]
yield images, steers
def load_validation_data(image_paths, steering_angles, preproc):
"""
Load validation frames into memory.
"""
images = [preproc(cv2.imread(x)) for x in image_paths]
return np.array(images), np.array(steering_angles)
def read_datapoint(path):
"""
Read an image and it's metadata from disk.
"""
fname = os.path.basename(path)
fname = os.path.splitext(fname)[0]
elems = fname.split("_")
if len(elems) != 3:
raise Exception('Invalid data set: ' + path)
# return id, timestamp, steering value
return int(elems[0]), int(elems[1]), float(elems[2])
def load_data(data_dir, test_size):
"""
Load data from disk and split it into training and validation sets.
"""
X = []
y = []
sub_dirs = os.listdir(data_dir)
for sub_dir in sub_dirs:
sub_dir_path = os.path.join(data_dir, sub_dir)
img_files = os.listdir(sub_dir_path)
for img_file in img_files:
img_path = os.path.join(sub_dir_path, img_file)
_, _, steering = read_datapoint(img_path)
X.append(img_path)
y.append(steering)
return sklearn.model_selection.train_test_split(X, y, test_size=test_size, random_state=None)
def build_model(name, print_layer_summary=True):
if name == NvidiaModel.NAME:
model = NvidiaModel()
elif name == CommaModel.NAME:
model = CommaModel()
# FIXME: adapt training_generator() for 1-channel images of TinyModel
# elif name == TinyModel.NAME:
# model = TinyModel()
if print_layer_summary:
model.get().summary()
return model
def train_model(model, epochs, batch_size, X_train, X_valid, y_train, y_valid):
"""
Train the model.
"""
logging = keras.callbacks.TensorBoard(log_dir='logs')
checkpoint = keras.callbacks.ModelCheckpoint('model-' + model.NAME + '-{epoch:03d}-{val_loss:.4f}.h5',
monitor='val_loss',
save_best_only=True)
kmodel = model.get()
kmodel.compile(loss='mean_squared_error', optimizer=keras.optimizers.Adam(lr=1.0e-4))
print('{} model compiled'.format(datetime.datetime.now()))
# 3/4 of training data per batch is generated; we need 4x steps to get once
# through the whole (physical) training set.
steps_per_epoch = len(X_train) * 4 // batch_size
# Pre-loading all validation data in memory speeds up training 10-15%.
# Switch back to validation data generator if the data get's too large.
X_valid_data, y_valid_data = load_validation_data(X_valid, y_valid, model.preprocess)
kmodel.fit_generator(
training_generator(X_train, y_train, model.preprocess, batch_size),
steps_per_epoch=steps_per_epoch,
epochs=epochs,
validation_data=(X_valid_data, y_valid_data),
#validation_data=training_generator(X_valid, y_valid, model.preprocess, batch_size),
#validation_steps=len(X_valid) // batch_size,
#workers=4,
callbacks=[logging, checkpoint],
verbose=1)
def main():
parser = argparse.ArgumentParser()
parser.add_argument('-d', dest='data_dir', type=str, default='data', help='data directory')
parser.add_argument('-t', dest='test_size', type=float, default=0.2, help='test size fraction')
parser.add_argument('-n', dest='epochs', type=int, default=50, help='number of epochs')
parser.add_argument('-b', dest='batch_size', type=int, default=64, help='batch size')
parser.add_argument('-m', dest='model_name', type=str, default='nvidia', help='model architecture ({}, {}, or {})'.format(NvidiaModel.NAME, CommaModel.NAME, TinyModel.NAME))
args = parser.parse_args()
print('{} start'.format(datetime.datetime.now()))
data = load_data(args.data_dir, args.test_size)
model = build_model(args.model_name)
print('{} model built'.format(datetime.datetime.now()))
train_model(model, args.epochs, args.batch_size, *data)
print('{} done'.format(datetime.datetime.now()))
if __name__ == '__main__':
main()