def test_more_map_operations(self):
transformed = TubSequence.build_pipeline(
self.sequence,
x_transform=lambda record: record.underlying['user/angle'],
y_transform=lambda record: record.underlying['user/throttle'])
transformed_2 = TubSequence.build_pipeline(
self.sequence,
x_transform=lambda record: record.underlying['user/angle'] * 2,
y_transform=lambda record: record.underlying['user/throttle'] * 2)
transformed_3 = TubSequence.map_pipeline(
x_transform=lambda x: x,
y_transform=lambda y: y,
pipeline=transformed_2
)
self.assertEqual(len(transformed), size)
self.assertEqual(len(transformed_2), size)
self.assertEqual(len(transformed_3), size)
transformed_list = list(transformed)
transformed_list_2 = list(transformed_3)
index = np.random.randint(0, 9)
x1, y1 = transformed_list[index]
x2, y2 = transformed_list_2[index]
self.assertAlmostEqual(x1 * 2, x2)
self.assertAlmostEqual(y1 * 2, y2)
def test_iterator_consistency(self):
extract = TubSequence.build_pipeline(
self.sequence,
x_transform=lambda record: record.underlying['user/angle'],
y_transform=lambda record: record.underlying['user/throttle'])
# iterate twice through half the data
r1 = list()
r2 = list()
for r in r1, r2:
iterator = iter(extract)
for i in range(size // 2):
r.append(next(iterator))
self.assertEqual(r1, r2)
# now transform and iterate through pipeline twice to see iterator
# doesn't exhaust
transformed = TubSequence.map_pipeline(
x_transform=lambda x: 2 * x,
y_transform=lambda y: 3 * y,
pipeline=extract)
l1 = list(transformed)
l2 = list(transformed)
self.assertEqual(l1, l2)
for e, t in zip(extract, transformed):
ex, ey = e
tx, ty = t
self.assertAlmostEqual(2 * ex, tx)
self.assertAlmostEqual(3 * ey, ty)
def __init__(self, model: KerasPilot, config: Config,
records: List[TubRecord], is_train: bool) -> None:
self.model = model
self.config = config
self.sequence = TubSequence(records)
self.batch_size = self.config.BATCH_SIZE
self.is_train = is_train
self.augmentation = ImageAugmentation(config)
self.pipeline = self._create_pipeline()
class BatchSequence(object):
"""
The idea is to have a shallow sequence with types that can hydrate
themselves to np.ndarray initially and later into the types required by
tf.data (i.e. dictionaries or np.ndarrays).
"""
def __init__(self, model: KerasPilot, config: Config,
records: List[TubRecord], is_train: bool) -> None:
self.model = model
self.config = config
self.sequence = TubSequence(records)
self.batch_size = self.config.BATCH_SIZE
self.is_train = is_train
self.augmentation = ImageAugmentation(config, 'AUGMENTATIONS')
self.transformation = ImageAugmentation(config, 'TRANSFORMATIONS')
self.pipeline = self._create_pipeline()
def __len__(self) -> int:
return math.ceil(len(self.pipeline) / self.batch_size)
def image_processor(self, img_arr):
""" Transformes the images and augments if in training. Then
normalizes it. """
img_arr = self.transformation.run(img_arr)
if self.is_train:
img_arr = self.augmentation.run(img_arr)
norm_img = normalize_image(img_arr)
return norm_img
def _create_pipeline(self) -> TfmIterator:
""" This can be overridden if more complicated pipelines are
required """
# 1. Initialise TubRecord -> x, y transformations
def get_x(record: TubRecord) -> Dict[str, Union[float, np.ndarray]]:
""" Extracting x from record for training"""
out_tuple = self.model.x_transform_and_process(
record, self.image_processor)
# convert tuple to dictionary which is understood by tf.data
out_dict = self.model.x_translate(out_tuple)
return out_dict
def get_y(record: TubRecord) -> Dict[str, Union[float, np.ndarray]]:
""" Extracting y from record for training """
y0 = self.model.y_transform(record)
y1 = self.model.y_translate(y0)
return y1
# 2. Build pipeline using the transformations
pipeline = self.sequence.build_pipeline(x_transform=get_x,
y_transform=get_y)
return pipeline
def create_tf_data(self) -> tf.data.Dataset:
""" Assembles the tf data pipeline """
dataset = tf.data.Dataset.from_generator(
generator=lambda: self.pipeline,
output_types=self.model.output_types(),
output_shapes=self.model.output_shapes())
return dataset.repeat().batch(self.batch_size)
class BatchSequence(object):
"""
The idea is to have a shallow sequence with types that can hydrate
themselves to np.ndarray initially and later into the types required by
tf.data (i.e. dictionaries or np.ndarrays).
"""
def __init__(self, model: KerasPilot, config: Config,
records: List[TubRecord], is_train: bool) -> None:
self.model = model
self.config = config
self.sequence = TubSequence(records)
self.batch_size = self.config.BATCH_SIZE
self.is_train = is_train
self.augmentation = ImageAugmentation(config)
self.pipeline = self._create_pipeline()
def __len__(self) -> int:
return math.ceil(len(self.pipeline) / self.batch_size)
def _create_pipeline(self) -> TfmIterator:
""" This can be overridden if more complicated pipelines are
required """
# 1. Initialise TubRecord -> x, y transformations
def get_x(record: TubRecord) -> Dict[str, Union[float, np.ndarray]]:
""" Extracting x from record for training"""
# this transforms the record into x for training the model to x,y
x0 = self.model.x_transform(record)
# for multiple input tensors the return value here is a tuple
# where the image is in first slot otherwise x0 is the image
x1 = x0[0] if isinstance(x0, tuple) else x0
x1 = np.squeeze(x1)
# apply augmentation to training data only
x2 = self.augmentation.augment(x1) if self.is_train else x1
# normalise image, assume other input data comes already normalised
x3 = normalize_image(x2)
# fill normalised image back into tuple if necessary
x4 = (x3, ) + x0[1:] if isinstance(x0, tuple) else x3
# convert tuple to dictionary which is understood by tf.data
x5 = self.model.x_translate(x4)
return x5
def get_y(record: TubRecord) -> Dict[str, Union[float, np.ndarray]]:
""" Extracting y from record for training """
y0 = self.model.y_transform(record)
y1 = self.model.y_translate(y0)
return y1
# 2. Build pipeline using the transformations
pipeline = self.sequence.build_pipeline(x_transform=get_x,
y_transform=get_y)
return pipeline
def create_tf_data(self) -> tf.data.Dataset:
""" Assembles the tf data pipeline """
dataset = tf.data.Dataset.from_generator(
generator=lambda: self.pipeline,
output_types=self.model.output_types(),
output_shapes=self.model.output_shapes())
return dataset.repeat().batch(self.batch_size)
def __init__(self, config, records: List[TubRecord], transform=None):
"""Create a PyTorch Tub Dataset
Args:
config (object): the configuration information
records (List[TubRecord]): a list of tub records
transform (function, optional): a transform to apply to the data
"""
self.config = config
# Handle the transforms
if transform:
self.transform = transform
else:
self.transform = get_default_transform()
self.sequence = TubSequence(records)
self.pipeline = self._create_pipeline()
class TorchTubDataset(IterableDataset):
'''
Loads the dataset, and creates a train/test split.
'''
def __init__(self, config, records: List[TubRecord], transform=None):
"""Create a PyTorch Tub Dataset
Args:
config (object): the configuration information
records (List[TubRecord]): a list of tub records
transform (function, optional): a transform to apply to the data
"""
self.config = config
# Handle the transforms
if transform:
self.transform = transform
else:
self.transform = get_default_transform()
self.sequence = TubSequence(records)
self.pipeline = self._create_pipeline()
def _create_pipeline(self):
""" This can be overridden if more complicated pipelines are
required """
def y_transform(record: TubRecord):
angle: float = record.underlying['user/angle']
throttle: float = record.underlying['user/throttle']
predictions = torch.tensor([angle, throttle], dtype=torch.float)
# Normalize to be between [0, 1]
# angle and throttle are originally between [-1, 1]
predictions = (predictions + 1) / 2
return predictions
def x_transform(record: TubRecord):
# Loads the result of Image.open()
img_arr = record.image(cached=True, as_nparray=False)
return self.transform(img_arr)
# Build pipeline using the transformations
pipeline = self.sequence.build_pipeline(x_transform=x_transform,
y_transform=y_transform)
return pipeline
def __iter__(self):
return iter(self.pipeline)
def setUp(self):
records = random_records(size=size)
self.sequence = TubSequence(records=records)
