forked from paul-o-alto/keras-resnet-sdc
-
Notifications
You must be signed in to change notification settings - Fork 0
/
model.py
491 lines (404 loc) · 16.3 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
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
import os
import csv
import time
import cv2
import scipy as sp
import numpy as np
import tensorflow as tf
flags = tf.app.flags
flags.FLAGS.CUDA_VISIBLE_DEVICES = ''
from augmentation import generate_train_from_PD_batch
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
from keras import backend as K
from keras.layers.core import Lambda
from keras.layers.advanced_activations import (
LeakyReLU,
PReLU,
ELU
)
from keras.callbacks import ModelCheckpoint
from keras.applications import VGG16, VGG19
from keras.applications.resnet50 import ResNet50
from keras.applications.inception_v3 import InceptionV3
from keras.applications.xception import Xception
from keras.optimizers import SGD, Adam, RMSprop
from keras.models import (
Model,
Sequential
)
from keras.layers import (
Input,
Activation,
merge,
Dense,
Flatten,
Dropout,
SpatialDropout2D,
Reshape,
ELU,
Conv2D,
GlobalAveragePooling2D
)
from keras.layers.convolutional import (
Convolution2D,
MaxPooling2D,
AveragePooling2D,
ZeroPadding2D
)
from keras.preprocessing.image import ImageDataGenerator
from keras.layers.normalization import BatchNormalization
from keras.utils.visualize_util import plot
from keras.wrappers.scikit_learn import KerasRegressor
from keras.regularizers import l2
from sklearn.preprocessing import StandardScaler
from sklearn.pipeline import Pipeline
from sklearn.model_selection import KFold, cross_val_score
from customlayers import convolution2Dgroup, crosschannelnormalization, \
splittensor, Softmax4D
from imagenet_tool import synset_to_id, id_to_synset,synset_to_dfs_ids
K.set_image_dim_ordering("tf")
DRIVING_TYPES = ['']
COURSES = ['']
# These two lists represent a tree: top level choose flat or inclines
# Second level choose mixed, corrective or inline
BASE_PATH = '/home/paul/workspace/keras-resnet-sdc/test_data'
MINI_BATCH_SIZE = 256
RESIZE_FACTOR = 0.5
EPOCHS = 25
#INPUT_SHAPE = (160*RESIZE_FACTOR, 320*RESIZE_FACTOR, 3)
INPUT_SHAPE = (64,64,3)
WEIGHTS = 'imagenet'
TRAINABLE = False # Determines if we train existing architectures end-to-end
NORMALIZE = False #True # Whether or not to normalize data before it enters the NN
DROPOUT = True
def sum_squared_error(y_true, y_pred):
return K.sum(K.square(y_pred - y_true), axis=0)
def tanh_scaled(x):
return 2*K.tanh(x)
def nvidia_model():
"""
This model is from the Nvidia "End-to-end" paper
"""
overall_activation = 'elu' #'linear' # DO NOT CHANGE! NEEDED IN ORDER TO AVOID SATURATION!
model = Sequential()
if NORMALIZE:
model.add(Lambda(lambda x: x/127.5 - 1,
input_shape=INPUT_SHAPE))
model.add(Convolution2D(24, 5, 5, subsample=(2,2),
input_shape=INPUT_SHAPE,
border_mode='valid', dim_ordering='tf'))
model.add(Activation(overall_activation))
model.add(Convolution2D(36, 5, 5, border_mode='valid', subsample=(2,2)))
model.add(Activation(overall_activation))
model.add(Convolution2D(48, 5, 5, border_mode='valid', subsample=(2,2)))
model.add(Activation(overall_activation))
model.add(Convolution2D(64, 3, 3, border_mode='valid'))
model.add(Activation(overall_activation))
model.add(Convolution2D(64, 3, 3, border_mode='valid'))
model.add(Activation(overall_activation))
model.add(Flatten())
model.add(Dense(1164, init="normal"))
if DROPOUT: model.add(Dropout(0.1))
model.add(Activation(overall_activation))
model.add(Dense(100, init="normal"))
if DROPOUT: model.add(Dropout(0.1))
model.add(Activation(overall_activation))
model.add(Dense(50, init="normal"))
if DROPOUT: model.add(Dropout(0.1))
model.add(Activation(overall_activation))
model.add(Dense(10, init="normal"))
if DROPOUT: model.add(Dropout(0.1))
model.add(Activation(overall_activation))
model.add(Dense(1))
model.add(Activation('linear'))
return model
def vgg16_model():
input_image = Input(shape=INPUT_SHAPE)
n_input = Lambda(lambda input_image: input_image/127.5 - 1,
input_shape=INPUT_SHAPE)
base_model = VGG16(weights=WEIGHTS,
input_tensor=input_image,
include_top=False)
if not TRAINABLE:
for layer in base_model.layers[:-6]:
layer.trainable = False
x = base_model.output
x = Flatten()(x) #GlobalAveragePooling2D()(x)
x = Dense(4096, activation="relu")(x) #, W_regularizer=l2(0.01))(x)
x = Dense(4096, activation="relu")(x) #, W_regularizer=l2(0.01))(x)
x = Dense(1, activation="linear")(x)
model = Model(input=input_image, output=x)
return model
def vgg19_model():
input_image = Input(shape=INPUT_SHAPE)
base_model = VGG19(weights=WEIGHTS,
input_tensor=input_image, include_top=False)
if not TRAINABLE:
for layer in base_model.layers[:-6]:
layer.trainable = False
x = base_model.output
x = Flatten()(x) #GlobalAveragePooling2D()(x)
x = Dense(4096, activation="relu")(x)
x = Dense(4096, activation="relu")(x)
x = Dense(1, activation="linear")(x)
return Model(input=input_image, output=x)
def resnet_model():
input_image = Input(shape=INPUT_SHAPE)
base_model = ResNet50(input_tensor=input_image,
weights=WEIGHTS,
include_top=False)
if not TRAINABLE:
for layer in base_model.layers:
layer.trainable = False
x = base_model.output
x = GlobalAveragePooling2D()(x)
x = Dense(100, activation='relu', name='fc1000')(x)
pred = Dense(1, activation='relu')(x)
model = Model(input=base_model.input, output=pred)
return model
def xception_model():
base_model = Xception(weights='imagenet', include_top=False)
for layer in base_model.layers:
layer.trainable = False
x = base_model.output
x = GlobalAveragePooling2D()(x)
x = Dense(1164, init="normal", activation='elu')(x)
x = Dense(100, init="normal", activation='elu')(x)
x = Dense(50, init="normal", activation='elu')(x)
x = Dense(10, init="normal", activation='elu')(x)
pred = Dense(1)(x)
model = Model(input=base_model.input, output=pred)
return model
def inception_model():
base_model = InceptionV3(weights=WEIGHTS, include_top=False)
if not TRAINABLE:
for layer in base_model.layers:
layer.trainable = False
x = base_model.output
x = GlobalAveragePooling2D()(x)
x = Dense(1024, activation='relu')(x)
x = Dense(200, activation='relu')(x)
pred = Dense(1)(x)
model = Model(input=base_model.input, output=pred)
return model
def comma_ai_model():
model = Sequential()
overall_activation = 'elu'
if NORMALIZE:
model.add(Lambda(lambda x: x/127.5 - 1,
input_shape=INPUT_SHAPE))
model.add(Conv2D(16, 8, 8, subsample=(4, 4), border_mode="same",
input_shape=INPUT_SHAPE))
model.add(Activation(overall_activation))
model.add(Conv2D(32, 5, 5, subsample=(2, 2), border_mode="same"))
model.add(Activation(overall_activation))
model.add(Conv2D(64, 5, 5, subsample=(2, 2), border_mode="same"))
model.add(Flatten())
if DROPOUT: model.add(Dropout(0.2))
model.add(Activation(overall_activation))
model.add(Dense(512))
if DROPOUT: model.add(Dropout(0.5))
model.add(Activation(overall_activation))
model.add(Dense(1))
return model
def alexnet_model(weights_path=None, regression=True):
if regression:
input_ = Input(shape=INPUT_SHAPE)
model = Sequential()
model.add(Convolution2D(64, 11, 11, subsample=(4, 4),
input_shape=INPUT_SHAPE, border_mode='valid', dim_ordering='tf'))
model.add(BatchNormalization())
model.add(Activation('relu'))
#model.add(MaxPooling2D(pool_size=(3, 3)))
model.add(Convolution2D(128, 7, 7, subsample=(3,3), border_mode='valid'))
model.add(BatchNormalization())
model.add(Activation('relu'))
#model.add(MaxPooling2D(pool_size=(3, 3)))
model.add(Convolution2D(192, 3, 3, subsample=(3,3), border_mode='valid'))
model.add(BatchNormalization())
model.add(Activation('relu'))
#model.add(MaxPooling2D(pool_size=(3, 3)))
model.add(Convolution2D(256, 3, 3, subsample=(3,3), border_mode='valid'))
model.add(BatchNormalization())
model.add(Activation('relu'))
#model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Flatten())
model.add(Dense(8192, init='normal'))
model.add(BatchNormalization())
model.add(Activation('relu'))
model.add(Dense(4096, init='normal'))
model.add(BatchNormalization())
model.add(Activation('relu'))
model.add(Dense(4096, init='normal'))
#model.add(BatchNormalization(1000))
model.add(Dense(1)) #Activation('softmax'))
return model
def load_batch(course_data, course_name, batches_so_far, batch_size, test=False):
driving_type = np.random.choice(DRIVING_TYPES)
driving_type_data = course_data[driving_type]
exhausted = False
if not test:
#coin = 0 #np.random.randint(0,2)
# Choose which side of the data to grab (0=beginning, 1=ending)
# Older vs. newer
if True: #if coin == 0:
s_index = batches_so_far*batch_size
if s_index + batch_size*2 > len(driving_type_data):
s_index = s_index % len(driving_type_data)
exhausted = True
sub_list = driving_type_data[s_index:s_index+batch_size]
#print(sub_list)
#else:
# sub_list = driving_type_data[-1:-1-batch_size*2]
data, labels = build_batch(batch_size, sub_list, course_name, driving_type)
else:
batch_size = int(0.3*batch_size)
sub_list = driving_type_data[-1-batch_size:-1] #len(course_data)-1-batch_size:]
data, labels = build_batch(batch_size, sub_list, course_name, driving_type)
print("Fetching batch of size %s for the %s course with driving type %s" % (batch_size,
course_name, driving_type))
return exhausted, data, labels
def build_batch(batch_size, sub_list, course_name, driving_type):
labels, data = [], []
current_size = 1
# Randomly sample from that subset
stop = False
while current_size < batch_size:
row = sub_list[current_size]
if "," not in row:
print ("Invalid csv row, no comma(s)")
break # Only thing that works
values = row.split(",")
if len(values) != 7:
print ("Invalid csv line, values != 7")
continue
filename_full, _, _, label, _, _, _ = values
filename_partial = filename_full.split("/")[-1]
# above is a hack, because of how Udacity simulator works
# Randomly choose between mixed, corrective, or inline driving sets
tmp_path = os.path.join(BASE_PATH, course_name, driving_type,
"IMG", filename_partial)
tmp_img = cv2.imread(tmp_path)
tmp_img = cv2.cvtColor(tmp_img, cv2.COLOR_RGB2YUV)
if RESIZE_FACTOR < 1:
tmp_img = cv2.resize(tmp_img, (0, 0), fx=64/320, fy=64/160)
data.append(tmp_img)
labels.append([label])
current_size += 1
data = np.stack(data, axis=0)
labels = np.stack(labels, axis=0)
return data, labels
def load_data(csv_lists, batches_so_far=0, batch_size=256, test=False):
# Randomly choose between the courses
course_name = np.random.choice(COURSES)
course_data = csv_lists[course_name]
exhausted, X_train, y_train = load_batch(course_data, course_name, batches_so_far,
batch_size=batch_size)
y_train = np.reshape(y_train, (len(y_train), 1))
if test:
_, X_test, y_test = load_batch(course_data, course_name, batches_so_far,
batch_size=batch_size, test=test)
y_test = np.reshape(y_test, (len(y_test), 1))
else:
X_test = None
y_test = None
return exhausted, (X_train, y_train), (X_test, y_test)
def main():
mini_batch_size = MINI_BATCH_SIZE
model = 'nvda'
if model == 'nvda':
model = nvidia_model()
elif model == 'alex':
model = alexnet_model()
elif model == 'comma':
model = comma_ai_model()
elif model == 'resnet':
model = resnet_model() # WORST OF THEM ALL!
elif model == 'inception':
model = inception_model()
elif model == 'xception':
model = xception_model()
elif model == 'vgg16':
model = vgg16_model()
elif model == 'vgg19':
model = vgg19_model()
plot(model, show_shapes=True, to_file='model.png')
model.compile(loss='mse', #metrics=['accuracy'],
optimizer='adam')
model.summary()
seed = 7
np.random.seed(seed)
# autosave best Model and load any previous weights
model_file = "./model.h5"
checkpointer = ModelCheckpoint(model_file,
verbose = 1, save_best_only = True)
if os.path.isfile(model_file):
model.load_weights(model_file)
model_json = model.to_json()
with open("model.json", "w") as json_file:
json_file.write(model_json)
# this will do preprocessing and realtime data augmentation
datagen = ImageDataGenerator(
rescale=1,
featurewise_center = False, # set input mean to 0 over the dataset
samplewise_center = False, # set each sample mean to 0
featurewise_std_normalization = False, # divide inputs by std of the dataset
samplewise_std_normalization = False, # divide each input by its std
zca_whitening = False, # apply ZCA whitening
rotation_range = 0, # randomly rotate images in the range (degrees, 0 to 180)
horizontal_flip = False, # randomly flip images
vertical_flip = False) # randomly flip images
print("Start training...")
csv_lists = {}
for course in COURSES:
course_dict = {}
type_dict = {}
for driving_type in DRIVING_TYPES:
f = open(os.path.join(BASE_PATH, course, driving_type, 'driving_log.csv'))
csv_list = f.read().split("\n")
type_dict[driving_type] = csv_list
f.close()
csv_lists[course] = type_dict
for key in csv_lists.keys():
print(key)
for driving_type in csv_lists[key].keys():
print(" %s" % driving_type)
print("fitting the model on the batches generated by datagen.flow()")
exhausted = False # Means we have gone through all of the training set
batches = 0
exhausted, (X_train, y_train), (X_test, y_test) = load_data(csv_lists,
batches_so_far=batches,
test=True)
epoch = 0
exhaust_batch_amount = 0
val_size = 1
print("Starting first epoch")
print(csv_list[0])
csv_list = csv_list[1:] # Remove keys
print(csv_list[0])
try:
while epoch < EPOCHS:
batches += 1
train_r_generator = generate_train_from_PD_batch(csv_list, mini_batch_size)
nb_vals = np.round(len(csv_list)/val_size) - 1
model.fit_generator(#datagen.flow(X_train, y_train,
train_r_generator,
#batch_size=mini_batch_size,
samples_per_epoch=20000,
#samples_per_epoch=len(X_train),
#validation_split=0.33,
nb_epoch=1, # on subsample
verbose=1,
validation_data=(X_test, y_test),
callbacks=[checkpointer]
)
#exhausted, (X_train, y_train), _ = load_data(csv_lists, batch_size=mini_batch_size,
# batches_so_far=batches)
epoch += 1
print("End epoch %s on training set" % epoch)
except KeyboardInterrupt:
print("Saving most recent weights before halting...")
model.save_weights(model_file)
if __name__ == '__main__':
main()