def test_angle_segment_shift_image(self):
data_set = self.create_real_dataset(
filter_method=drive_record_filter_include_all)
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),
1), # sharp turn left (zero right camera)
AngleSegment((-0.249, -0.1), 12), # big turn left
AngleSegment((-0.1, 0), 12), # straight left
AngleSegment((0, 0.001), 2), # straight zero center camera
AngleSegment((0.001, 0.1), 12), # straight right
AngleSegment((0.1, 0.25), 12), # big turn right
AngleSegment((0.25, 0.251),
1), # 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
generator = pipe_line_generators(
shift_image_generator(angle_offset_pre_pixel=0.0015),
flip_random_generator, brightness_image_generator(0.35),
shadow_generator)
self._angle_distribution(
"angle_distribution_generator_angle_segment_shift_image",
100,
256,
allocator=allocator.allocate,
generator=generator)
def segment_normal_distribution_shift_flip_brightness_shadow_reg():
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_with_regularizer(input_shape=data_set.output_shape(),
dropout=0.2)
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 test_angle_distribution_generator_45_10_45_pipe_line(self):
data_set = self.create_real_dataset(
filter_method=drive_record_filter_exclude_duplicated_small_angles)
allocator = AngleTypeWithZeroRecordAllocator(data_set, 20, 20, 15, 15,
15, 0.25)
generator = pipe_line_generators(
shift_image_generator(angle_offset_pre_pixel=0.002),
flip_generator, brightness_image_generator(0.25))
self._angle_distribution(
"angle_distribution_generator_exclude_duplicated_small_angles_40_20_40_pipe_line",
100,
256,
allocator=allocator.allocate,
angle_offset_pre_pixel=0.006,
generator=generator)
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 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))
def _angle_distribution(self,
name,
batches,
batch_size,
allocator,
angle_offset_pre_pixel=0.002,
generator=None):
if generator is None:
generator = pipe_line_random_generators(
image_itself,
shift_image_generator(
angle_offset_pre_pixel=angle_offset_pre_pixel),
flip_generator)
data_generator = DataGenerator(allocator, generator)
angles = np.array([])
for index in range(batches):
print("batch {} / {}".format(index, batches))
_, _angles = next(data_generator.generate(batch_size=batch_size))
angles = np.append(angles, _angles)
plt = Plot.angle_distribution(angles)
plt.savefig("angle/{}.jpg".format(name))