def numpy(self) -> np.ndarray:
if not self.has_meta_info_:
raise RuntimeError("This variable does not have meta info")
return np.fromfile(
self.file_path,
dtype=dtype_util.convert_oneflow_dtype_to_numpy_dtype(self.dtype),
).reshape(self.shape)
def run_job(test_case, device=None, in_type=None, shape=None):
assert shape is not None
flow.clear_default_session()
func_config = flow.FunctionConfig()
@flow.global_function(type="train", function_config=func_config)
def FuseBnAddReluJob(x: oft.Numpy.Placeholder(
shape, dtype=in_type)) -> oft.Numpy:
addend = flow.constant_like(x, 2)
with flow.scope.placement(device, "0:0-0"):
x = (flow.get_variable(
"x1",
shape=shape,
dtype=in_type,
initializer=flow.random_uniform_initializer(minval=-10,
maxval=10),
trainable=True,
) + x)
loss = flow.nn.relu(_batch_norm(x, last=False) + addend) + 1
flow.optimizer.SGD(flow.optimizer.PiecewiseConstantScheduler(
[], [0.0001]),
momentum=0).minimize(loss)
return loss
np_in_type = dtype_util.convert_oneflow_dtype_to_numpy_dtype(in_type)
x = (np.random.rand(*shape) * 10).astype(np_in_type)
FuseBnAddReluJob(x)
def run_fuse_cast_scale_mlir(test_case,
device=None,
in_type=None,
out_type=None,
shape=None):
flow.clear_default_session()
func_config = flow.FunctionConfig()
@flow.global_function(function_config=func_config)
def FuseCastScaleJob(x: oft.Numpy.Placeholder(
shape, dtype=in_type)) -> Tuple[oft.Numpy, oft.Numpy]:
with flow.scope.placement(device, "0:0-0"):
scale = flow.get_variable(
"scale",
shape=(1, ),
dtype=out_type,
initializer=flow.random_uniform_initializer(),
trainable=False,
)
loss = flow.cast(x, dtype=out_type) * scale
return (loss, scale)
np_in_type = dtype_util.convert_oneflow_dtype_to_numpy_dtype(in_type)
x = (np.random.rand(*shape) * 10).astype(np_in_type)
ret = FuseCastScaleJob(x)
(loss, scale) = ret
test_case.assertTrue(np.allclose(loss, x * scale))
def run_job(test_case, device=None, in_type=None, shape=None):
assert shape is not None
flow.clear_default_session()
func_config = flow.FunctionConfig()
@flow.global_function(function_config=func_config)
def FuseTrilScaleJob(x: oft.Numpy.Placeholder(
shape, dtype=in_type)) -> oft.Numpy:
with flow.scope.placement(device, "0:0-0"):
scale = 3.0
loss = flow.math.tril(x * scale)
loss += flow.math.tril(x) * scale
return loss
np_in_type = dtype_util.convert_oneflow_dtype_to_numpy_dtype(in_type)
x = (np.random.rand(*shape) * 10).astype(np_in_type)
FuseTrilScaleJob(x)
def __init__(
self,
var_dir: str,
dtype: Optional[oneflow.dtype] = None,
shape: Optional[Sequence[int]] = None,
):
data_path = os.path.join(var_dir, DATA_FILENAME)
if not os.path.isfile(data_path):
raise FileNotFoundError()
self.var_dir_ = var_dir
meta_info_path = os.path.join(self.var_dir_, META_INFO_FILENAME)
if os.path.exists(meta_info_path):
meta_info = variable_meta_info_pb.VariableMetaInfo()
with open(meta_info_path) as f:
text_format.Parse(f.read(), meta_info)
self.has_meta_info_ = True
else:
self.has_meta_info_ = False
if self.has_meta_info_:
assert dtype is None and shape is None
self.shape_ = tuple(meta_info.shape.dim)
self.dtype_ = dtype_util.convert_proto_dtype_to_oneflow_dtype(
meta_info.data_type)
elif shape is not None and dtype is not None:
self.shape_ = shape
self.dtype_ = dtype
self.has_meta_info_ = True
elif shape is not None or dtype is not None:
raise RuntimeError(
"both or neither of shape and dtype should be None")
else:
pass
if self.has_meta_info_:
itemsize = np.dtype(
dtype_util.convert_oneflow_dtype_to_numpy_dtype(
self.dtype_)).itemsize
assert os.path.getsize(data_path) == np.prod(
self.shape).item() * itemsize
def generate_values_by_initializer(initializer, shape, dtype):
np_dtype = np.dtype(dtype_util.convert_oneflow_dtype_to_numpy_dtype(dtype))
length = _elem_cnt(shape)
return np.array(initializer(length)).astype(np_dtype).reshape(shape)