def _test_real_serialization(self):
wannabe = TinyConvNet3D()
tik_torch = TikTorch(model=wannabe)
with tempdir() as d:
the_path = '{}/testfile.nn'.format(d)
tik_torch.serialize(to_path=the_path)
new_torch = TikTorch.unserialize(from_path=the_path)
self.assertIsInstance(new_torch, TikTorch)
def __init__(self,
tiktorch_net,
filename=None,
HALO_SIZE=32,
BATCH_SIZE=3):
"""
Args:
tiktorch_net (tiktorch): tiktorch object to be loaded into this
classifier object
filename (None, optional): Save file name for future reference
"""
self._filename = filename
if self._filename is None:
self._filename = ""
self.HALO_SIZE = HALO_SIZE
self.BATCH_SIZE = BATCH_SIZE
if tiktorch_net is None:
print(self._filename)
tiktorch_net = TikTorch.unserialize(self._filename)
# print (self._filename)
# assert tiktorch_net.return_hypercolumns == False
# print('blah')
self._tiktorch_net = tiktorch_net
self._opReorderAxes = OpReorderAxes(graph=Graph())
self._opReorderAxes.AxisOrder.setValue("zcyx")
def __init__(self, tiktorch_net, filename=None, HALO_SIZE=32, BATCH_SIZE=3):
"""
Args:
tiktorch_net (tiktorch): tiktorch object to be loaded into this
classifier object
filename (None, optional): Save file name for future reference
"""
self._filename = filename
if self._filename is None:
self._filename = ""
self.HALO_SIZE = HALO_SIZE
self.BATCH_SIZE = BATCH_SIZE
if tiktorch_net is None:
print(self._filename)
tiktorch_net = TikTorch.unserialize(self._filename)
# print (self._filename)
# assert tiktorch_net.return_hypercolumns == False
# print('blah')
self._tiktorch_net = tiktorch_net
self._opReorderAxes = OpReorderAxes(graph=Graph())
self._opReorderAxes.AxisOrder.setValue("zcyx")
def setUp(self):
# Build model and object
model = TinyConvNet3D()
tiktorch = TikTorch(model=model).configure(window_size=[3, 512, 512],
num_input_channels=1,
num_output_channels=1)
self.tiktorch = tiktorch
self.input_array = np.zeros(shape=(1, 3, 512, 512))
def _test_gpu_serialization(self):
wannabe = TinyConvNet3D()
tik_torch = TikTorch(model=wannabe)
tik_torch.cuda()
with tempdir() as d:
the_path = "{}/testfile.nn".format(d)
tik_torch.serialize(to_path=the_path)
new_torch = TikTorch.unserialize(from_path=the_path)
self.assertIsInstance(new_torch, TikTorch)
self.assertTrue(new_torch.is_cuda)
def create_and_train_pixelwise(self, feature_images, label_images, axistags=None, feature_names=None):
self._filename = PYTORCH_MODEL_FILE_PATH
logger.debug("Loading pytorch network from {}".format(self._filename))
# Save for future reference
# known_labels = numpy.sort(vigra.analysis.unique(y))
# TODO: check whether loaded network has the same number of classes as specified in ilastik!
self._loaded_pytorch_net = TikTorch.unserialize(self._filename)
logger.info(self.description)
# logger.info("OOB during training: {}".format( oob ))
return TikTorchLazyflowClassifier(self._loaded_pytorch_net, self._filename)
def deserialize_hdf5(cls, h5py_group):
# TODO: load from HDF5 instead of hard coded path!
logger.debug("Deserializing")
# HACK:
# filename = PYTORCH_MODEL_FILE_PATH
filename = h5py_group[cls.HDF5_GROUP_FILENAME]
logger.debug("Deserializing from {}".format(filename))
with tempfile.TemporaryFile() as f:
f.write(h5py_group["classifier"].value)
f.seek(0)
loaded_pytorch_net = TikTorch.unserialize(f)
return TikTorchLazyflowClassifier(loaded_pytorch_net, filename)
def deserialize_hdf5(cls, h5py_group):
# TODO: load from HDF5 instead of hard coded path!
logger.debug("Deserializing")
# HACK:
# filename = PYTORCH_MODEL_FILE_PATH
filename = h5py_group[cls.HDF5_GROUP_FILENAME]
logger.debug("Deserializing from {}".format(filename))
with tempfile.TemporaryFile() as f:
f.write(h5py_group["classifier"].value)
f.seek(0)
loaded_pytorch_net = TikTorch.unserialize(f)
return TikTorchLazyflowClassifier(loaded_pytorch_net, filename)
def test_less_output_channels(self):
model = TinyConvNet3D(num_input_channels=3, num_output_channels=2)
tiktorch = TikTorch(model=model).configure(window_size=[3, 512, 512],
num_input_channels=3,
num_output_channels=2)
self.tiktorch = tiktorch
self.input_array = np.zeros(shape=(3, 3, 512, 512))
# Forward and check
output = self.tiktorch.forward([self.input_array])
# noinspection PyTypeChecker
self.assertEqual(len(output), 1)
# noinspection PyUnresolvedReferences
self.assertIsInstance(output[0], np.ndarray)
# noinspection PyUnresolvedReferences
self.assertEqual(output[0].shape, (2, 3, 512, 512))
def setup_class(cls):
cls.graph = Graph()
cls.model = TinyConvNet2D(num_input_channels=1, num_output_channels=2)
cls.tiktorch_net = TikTorch(model=cls.model)
# HACK: Window size catered to settings in tiktorchLazyflowClassifier :/
cls.tiktorch_net.configure(window_size=[192, 192],
num_input_channels=1,
num_output_channels=2)
# HACK: this is also catered to the hardcoded settings in tiktorchLazyflowClassifier
cls.data = numpy.arange(3 * 192 * 192).astype(numpy.uint8).reshape(
(3, 1, 192, 192))
cls.tmp_dir = tempfile.mkdtemp()
cls.h5_file = h5py.File(os.path.join(cls.tmp_dir, "h5_file.h5"),
mode="a")
cls.classifier_group = cls.h5_file.create_group("classifier")
def create_and_train_pixelwise(self,
feature_images,
label_images,
axistags=None,
feature_names=None):
self._filename = PYTORCH_MODEL_FILE_PATH
logger.debug("Loading pytorch network from {}".format(self._filename))
# Save for future reference
# known_labels = numpy.sort(vigra.analysis.unique(y))
# TODO: check whether loaded network has the same number of classes as specified in ilastik!
self._loaded_pytorch_net = TikTorch.unserialize(self._filename)
logger.info(self.description)
# logger.info("OOB during training: {}".format( oob ))
return TikTorchLazyflowClassifier(self._loaded_pytorch_net,
self._filename)
def setUp(self):
TikTorch.read_config = lambda self: self
tiktorch = TikTorch(build_directory=".")
if self.MODEL == "HEAVY":
tiktorch._model = nn.Sequential(
nn.Conv2d(1, 512, 9, padding=4),
nn.ELU(),
nn.Conv2d(512, 512, 9, padding=4),
nn.ELU(),
nn.Conv2d(512, 512, 9, padding=4),
nn.ELU(),
nn.Conv2d(512, 512, 9, padding=4),
nn.ELU(),
nn.Conv2d(512, 1, 9, padding=4),
)
elif self.MODEL == "LITE":
tiktorch._model = nn.Conv2d(1, 1, 1)
tiktorch._config = {"input_shape": [1, 512, 512], "dynamic_input_shape": "(32 * (nH + 1), 32 * (nW + 1))"}
tiktorch._set_handler(tiktorch._model)
self.tiktorch = tiktorch
def setUp(self):
TikTorch.read_config = lambda self: self
tiktorch = TikTorch(build_directory='.')
if self.MODEL == 'HEAVY':
tiktorch._model = nn.Sequential(
nn.Conv2d(1, 512, 9, padding=4),
nn.ELU(),
nn.Conv2d(512, 512, 9, padding=4),
nn.ELU(),
nn.Conv2d(512, 512, 9, padding=4),
nn.ELU(),
nn.Conv2d(512, 512, 9, padding=4),
nn.ELU(),
nn.Conv2d(512, 1, 9, padding=4),
)
elif self.MODEL == 'LITE':
tiktorch._model = nn.Conv2d(1, 1, 1)
tiktorch._config = {
'input_shape': [1, 512, 512],
'dynamic_input_shape': '(32 * (nH + 1), 32 * (nW + 1))'
}
tiktorch._set_handler(tiktorch._model)
self.tiktorch = tiktorch
