def suite():
test_modules = [
"tests.test_datasaver",
"tests.test_nas",
]
if is_torch_available():
test_modules.extend([
"tests.test_torch_data",
"tests.test_torch_model",
"tests.test_torch_optimizer",
])
if is_tf_available():
test_modules.extend([
"tests.test_tf_data",
"tests.test_tf_model",
"tests.test_tf_optimizer",
])
alltests = unittest.TestSuite()
for name in test_modules:
# Unexpectedly, the following code doesn't seem to work anymore
# in python 3
# exec('from %s import suite as test_suite' % name)
__import__(name)
test_suite = sys.modules[name].suite
alltests.addTest(test_suite())
return alltests
def _check_data_numpytype(self):
"""
Check data which is numpy array type
"""
if not isinstance(self._datas, dict):
raise TypeError("datas shoud be dict type.", self.__call__.__doc__)
bzs = 0
for name, array in self._datas.items():
self._check_attr_name(name=name)
if is_tf_available() and not isinstance(array, np.ndarray):
array = array.numpy()
if is_torch_available() and not isinstance(array, np.ndarray):
array = array.numpy()
assert isinstance(array,
np.ndarray), "array type is not `numpy.ndarray`"
if bzs:
if bzs != array.shape[0]:
raise ValueError(
"each datas array batch sizes are not same.")
else:
bzs = array.shape[0]
# This resape is made into a (B, *) shape.
# Shape is lowered to two contiguous dimensions, enabling IO operations to operate very quickly.
# https://www.slideshare.net/HDFEOS/caching-and-buffering-in-hdf5#25
if len(array.shape) == 1:
# this array is ground truth
array = array.reshape(-1, 1)
self._datas[name] = array.reshape(
-1, reduce(lambda x, y: x * y, array.shape[1:]))