def init_by_initializer_conf(
var_blob: EagerBlobTrait,
initializer_conf: initializer_conf_util.InitializerConf,
sync_between_multi_machine: bool,
random_seed=0,
):
initializer = initializer_util.GetInitializer(initializer_conf,
random_seed, var_blob.shape)
# initializer is None if and only if the initializer_conf is empty_initializer
if initializer is None:
return
def GenerateValueAndAssign(var_blob, start_nd_idx, stop_nd_idx):
np_dtype = np.dtype(
dtype_util.convert_oneflow_dtype_to_numpy_dtype(var_blob.dtype))
length = _ElemCnt(np.array(stop_nd_idx) - np.array(start_nd_idx))
vals = (np.array(initializer(length)).astype(np_dtype).reshape(
np.array(stop_nd_idx) - np.array(start_nd_idx)))
slice_value_blob = _GetCpu0VariableBlobFromNumpy(vals, var_blob.dtype)
_LogicalSliceAssign(
var_blob,
slice_value_blob,
start_nd_idx,
stop_nd_idx,
)
# we just want to run f on every slice without caring about the return value
for _ in _ForEachSlice(var_blob, GenerateValueAndAssign):
pass
if sync_between_multi_machine:
oneflow_api.eager.Sync()
def _init_eager_local_tensor_by_initializer_conf(eager_local_tensor,
initializer_conf,
random_seed=0):
shape = tuple(eager_local_tensor.shape)
initializer = initializer_util.GetInitializer(initializer_conf,
random_seed, shape)
# initializer is None if and only if the initializer_conf is empty_initializer
if initializer is None:
return
_copy_from_numpy_to_eager_local_tensor(
eager_local_tensor,
check_point_v2.generate_values_by_initializer(
initializer, shape, eager_local_tensor.dtype),
)
def Init() -> None:
oneflow.sync_default_session()
sess = session_ctx.GetDefaultSession()
for op_name, var_blob in GetAllVariables().items():
var_conf = sess.OpAttribute4InterfaceOpName(
op_name).op_conf.variable_conf
if not (var_conf.HasField("initializer")
or var_conf.HasField("initialize_with_snapshot")):
continue
if var_conf.HasField("initialize_with_snapshot"):
initialize_with_snapshot_conf = var_conf.initialize_with_snapshot
if initialize_with_snapshot_conf.HasField("key"):
snapshot_key = op_name
else:
snapshot_key = initialize_with_snapshot_conf.key
var_dir = os.path.dirname(
os.path.join(
initialize_with_snapshot_conf.path,
snapshot_key,
))
LoadVariables({op_name: GetCheckpoint(var_dir)})
continue
g = initializer_util.GetInitializer(var_conf.initializer,
var_conf.random_seed,
var_blob.shape)
def GenerateValueAndAssign(var_blob, start_nd_idx, stop_nd_idx):
np_dtype = np.dtype(
dtype_util.convert_oneflow_dtype_to_numpy_dtype(
var_blob.dtype))
length = _ElemCnt(np.array(stop_nd_idx) - np.array(start_nd_idx))
vals = (np.array(g(length)).astype(np_dtype).reshape(
np.array(stop_nd_idx) - np.array(start_nd_idx)))
slice_value_blob = _GetCpu0VariableBlobFromNumpy(
vals, var_blob.dtype)
_LogicalSliceAssign(
var_blob,
slice_value_blob,
start_nd_idx,
stop_nd_idx,
)
# we just want to run f on every slice without caring about the return value
for _ in _ForEachSlice(var_blob, GenerateValueAndAssign):
pass
oneflow_api.eager.Sync()
def init_by_initializer_conf(
var_blob: Union[EagerBlobTrait, "oneflow.Tensor"],
initializer_conf: initializer_conf_util.InitializerConf,
sync_between_multi_machine: bool,
scope_symbol_id: Optional[int],
random_seed: int = 0,
):
initializer = initializer_util.GetInitializer(initializer_conf,
random_seed, var_blob.shape)
# initializer is None if and only if the initializer_conf is empty_initializer
if initializer is None:
return
def GenerateValueAndAssign(var_blob, start_nd_idx, stop_nd_idx):
np_dtype = np.dtype(
dtype_util.convert_oneflow_dtype_to_numpy_dtype(var_blob.dtype))
length = _ElemCnt(np.array(stop_nd_idx) - np.array(start_nd_idx))
vals = (np.array(initializer(length)).astype(np_dtype).reshape(
np.array(stop_nd_idx) - np.array(start_nd_idx)))
if isinstance(var_blob, oneflow.Tensor):
var_blob_object = var_blob._blob_object
else:
assert isinstance(var_blob, EagerBlobTrait)
var_blob_object = var_blob.blob_object
slice_value_blob = _GetCpu0VariableBlobFromNumpy(vals, var_blob.dtype)
_LogicalSliceAssign(
var_blob_object,
slice_value_blob.blob_object,
start_nd_idx,
stop_nd_idx,
scope_symbol_id,
)
# we just want to run f on every slice without caring about the return value
for _ in _ForEachSlice(var_blob, GenerateValueAndAssign):
pass
if sync_between_multi_machine:
oneflow_api.eager.single_client.Sync()
def init_by_initializer_conf(
var_blob: Union[EagerBlobTrait, "oneflow.Tensor"],
initializer_conf: initializer_conf_util.InitializerConf,
sync_between_multi_machine: bool,
scope_symbol_id: Optional[int],
random_seed: int = 0,
):
initializer = initializer_util.GetInitializer(initializer_conf,
random_seed, var_blob.shape)
# initializer is None if and only if the initializer_conf is empty_initializer
if initializer is None:
return
def GenerateValueAndAssign(var_blob, start_nd_idx, stop_nd_idx):
shape = np.array(stop_nd_idx) - np.array(start_nd_idx)
vals = generate_values_by_initializer(initializer, shape,
var_blob.dtype)
if isinstance(var_blob, oneflow.Tensor):
raise ValueError("Tensor object arguments are not supported")
else:
assert isinstance(var_blob, EagerBlobTrait)
var_blob_object = var_blob.blob_object
slice_value_blob = _GetCpu0VariableBlobFromNumpy(vals, var_blob.dtype)
_LogicalSliceAssign(
var_blob_object,
slice_value_blob.blob_object,
start_nd_idx,
stop_nd_idx,
scope_symbol_id,
)
# we just want to run f on every slice without caring about the return value
for _ in _ForEachSlice(var_blob, GenerateValueAndAssign):
pass
if sync_between_multi_machine:
oneflow._oneflow_internal.eager.single_client.Sync()
