def segment_normal_distribution_shift_flip_brightness_shadow():
data_set = DriveDataSet.from_csv(
"datasets/udacity-sample-track-1/driving_log.csv",
crop_images=True,
all_cameras_images=True,
filter_method=drive_record_filter_include_all)
# fine tune every part of training data so that make it meat std distrubtion
allocator = AngleSegmentRecordAllocator(
data_set,
AngleSegment((-1.5, -0.5), 10), # big sharp left
AngleSegment((-0.5, -0.25), 14), # sharp left
AngleSegment((-0.25, -0.249),
3), # sharp turn left (zero right camera)
AngleSegment((-0.249, -0.1), 10), # big turn left
AngleSegment((-0.1, 0), 11), # straight left
AngleSegment((0, 0.001), 4), # straight zero center camera
AngleSegment((0.001, 0.1), 11), # straight right
AngleSegment((0.1, 0.25), 10), # big turn right
AngleSegment((0.25, 0.251), 3), # sharp turn right (zero left camera)
AngleSegment((0.251, 0.5), 14), # sharp right
AngleSegment((0.5, 1.5), 10) # big sharp right
)
# a pipe line with shift -> flip -> brightness -> shadow augment processes
augment = pipe_line_generators(
shift_image_generator(angle_offset_pre_pixel=0.002), flip_generator,
brightness_image_generator(0.35), shadow_generator)
data_generator = DataGenerator(allocator.allocate, augment)
model = nvidia(input_shape=data_set.output_shape(), dropout=0.5)
Trainer(model,
learning_rate=0.0001,
epoch=45,
multi_process=use_multi_process,
custom_name=inspect.stack()[0][3]).fit_generator(
data_generator.generate(batch_size=256))
def raw_data_centre_left_right_image_crop():
# crop_images=True was the only difference
data_set = DriveDataSet.from_csv(
"datasets/udacity-sample-track-1/driving_log.csv",
crop_images=True,
all_cameras_images=True,
filter_method=drive_record_filter_include_all)
allocator = RecordRandomAllocator(data_set)
generator = image_itself
data_generator = DataGenerator(allocator.allocate, generator)
model = nvidia(input_shape=data_set.output_shape(), dropout=0.5)
Trainer(model,
learning_rate=0.0001,
epoch=10,
custom_name=inspect.stack()[0][3]).fit_generator(
data_generator.generate(batch_size=128))
def raw_data_centre_left_right_crop_shift_flip():
data_set = DriveDataSet.from_csv(
"datasets/udacity-sample-track-1/driving_log.csv",
crop_images=True,
all_cameras_images=True,
filter_method=drive_record_filter_include_all)
allocator = RecordRandomAllocator(data_set)
# shift_image_generator was the only difference
generator = pipe_line_generators(
shift_image_generator(angle_offset_pre_pixel=0.002), flip_generator)
data_generator = DataGenerator(allocator.allocate, generator)
model = nvidia(input_shape=data_set.output_shape(), dropout=0.5)
Trainer(model,
learning_rate=0.0001,
epoch=20,
multi_process=use_multi_process,
custom_name=inspect.stack()[0][3]).fit_generator(
data_generator.generate(batch_size=128))
def raw_data_centre_image_no_dropout():
# Create DriveDataSet from csv file, you can specify crop image, using all cameras and which data will included in
data_set = DriveDataSet.from_csv(
"datasets/udacity-sample-track-1/driving_log.csv",
crop_images=False,
all_cameras_images=False,
filter_method=drive_record_filter_include_all)
# What the data distribution will be, below example just randomly return data from data set, so that the
# distribution will be same with what original data set have
allocator = RecordRandomAllocator(data_set)
# what's the data augment pipe line have, this have no pipe line, just the image itself
augment = image_itself
# connect allocator and augment together
data_generator = DataGenerator(allocator.allocate, augment)
# create the model
model = nvidia(input_shape=data_set.output_shape(), dropout=0.0)
# put everthing together, start a real Keras training process with fit_generator
Trainer(model,
learning_rate=0.0001,
epoch=10,
custom_name=inspect.stack()[0][3]).fit_generator(
data_generator.generate(batch_size=128))
def segment_left_centre_right():
data_set = DriveDataSet.from_csv(
"datasets/udacity-sample-track-1/driving_log.csv",
crop_images=True,
all_cameras_images=True,
filter_method=drive_record_filter_include_all)
allocator = AngleTypeWithZeroRecordAllocator(
data_set,
left_percentage=20,
right_percentage=20,
zero_percentage=8,
zero_left_percentage=6,
zero_right_percentage=6,
left_right_image_offset_angle=0.25)
generator = pipe_line_generators(
shift_image_generator(angle_offset_pre_pixel=0.002), flip_generator,
brightness_image_generator(0.25), shadow_generator)
data_generator = DataGenerator(allocator.allocate, generator)
model = nvidia(input_shape=data_set.output_shape(), dropout=0.5)
Trainer(model, learning_rate=0.0001,
epoch=10).fit_generator(data_generator.generate(batch_size=128))
for batch_size in BATCH_SIZE:
training_data_generator = KerasGenerator(training_data,
batch_size,
INPUT_SHAPE,
preprocessor,
augmenter=augmenter)
validation_data_generator = KerasGenerator(validation_data, batch_size,
INPUT_SHAPE, preprocessor)
for lr in LR:
model_details = '{}-{}.{}'.format(batch_size, lr, model_name)
model = nvidia(INPUT_SHAPE)
optimizer = Adam(lr=lr)
model.compile(optimizer=optimizer, loss=LOSS)
progbar_logger = ProgbarLogger(count_mode='steps')
model_checkpoint = ModelCheckpoint(
'../models/' + model_details +
'.{epoch:02d}-{val_loss:.2f}.hdf5')
early_stopping = EarlyStopping(patience=PATIENCE)
callbacks = [progbar_logger, model_checkpoint, early_stopping]
history = model.fit_generator(
training_data_generator,
validation_data=validation_data_generator,
callbacks=callbacks,
