def build(builder):
if blob_object.op_arg_parallel_attr.is_mirrored():
input_blob_def = input_blob_def_util.MirroredTensorDef(
ndarray.shape,
dtype=dtype_util.convert_numpy_dtype_to_oneflow_dtype(ndarray.dtype),
)
else:
input_blob_def = input_blob_def_util.FixedTensorDef(
ndarray.shape,
dtype=dtype_util.convert_numpy_dtype_to_oneflow_dtype(ndarray.dtype),
)
push_util.FeedValueToEagerBlob(blob_object, input_blob_def, ndarray)
def _FeedValueToVariable(var_blob: oneflow_api.EagerConsistentBlob,
value: ValueContainer) -> None:
"""
Feed the value of `value` to the variable `var_blob`
"""
assert isinstance(value, (EagerBlobTrait, FileBackendVariableBlob,
np.ndarray)), "Unknown value type: {}".format(
type(value).__name__)
if isinstance(value, FileBackendVariableBlob):
if not value.has_meta_info_:
value = FileBackendVariableBlob(value.var_dir_, var_blob.dtype,
var_blob.shape)
assert var_blob.shape == value.shape, "{} vs {}".format(
var_blob.shape, value.shape)
if isinstance(value, np.ndarray):
value_flow_dtype = dtype_util.convert_numpy_dtype_to_oneflow_dtype(
value.dtype)
else:
value_flow_dtype = value.dtype
assert var_blob.dtype == value_flow_dtype, "{} vs {}".format(
var_blob.dtype, value_flow_dtype)
for start, stop, slice in _ReadSlice(value):
slice_value_blob = _GetCpu0VariableBlobFromNumpy(slice, var_blob.dtype)
_LogicalSliceAssign(
var_blob,
slice_value_blob,
start,
stop,
)
def construct_tensor(
data,
dtype=None,
device=None,
requires_grad=False,
placement=None,
sbp=None,
is_consistent=False,
is_lazy=False,
):
if _is_scalar(data) or _input_args_is_data(data):
if (not _input_args_is_numpy(data) and dtype is None
and _input_dtype_is_float(data)):
dtype = flow.float32
data = np.array(data)
if dtype is None:
dtype = dtype_util.convert_numpy_dtype_to_oneflow_dtype(data.dtype)
return Tensor(
data,
dtype=dtype,
device=device,
requires_grad=requires_grad,
placement=placement,
sbp=sbp,
is_consistent=is_consistent,
is_lazy=is_lazy,
)
else:
raise TypeError("Construction error, invalid combination of arguments")
def build(builder):
blob_object = builder.MakeLazyRefBlobObject(op_name)
if blob_object.op_arg_blob_attr.is_tensor_list:
input_blob_def = input_blob_def_util.MirroredTensorListDef(
[x.shape for x in ndarray],
dtype=dtype_util.convert_numpy_dtype_to_oneflow_dtype(
ndarray.dtype),
)
elif blob_object.op_arg_parallel_attr.is_mirrored():
input_blob_def = input_blob_def_util.MirroredTensorDef(
ndarray.shape,
dtype=dtype_util.convert_numpy_dtype_to_oneflow_dtype(
ndarray.dtype),
)
else:
input_blob_def = input_blob_def_util.FixedTensorDef(
ndarray.shape,
dtype=dtype_util.convert_numpy_dtype_to_oneflow_dtype(
ndarray.dtype),
)
push_util.FeedValueToEagerBlob(blob_object, input_blob_def,
ndarray)
Yield()
def infer_oneflow_data_placeholder(
self, batch: Tuple[np.ndarray, ...] = None, optimizer_idx: int = 0
):
assert isinstance(batch, tuple), "model.NumpyDataModule must return a tuple."
data_placeholder_list = []
for item in batch:
assert isinstance(
item, np.ndarray
), "model.NumpyDataModule must return a tuple of numpy."
of_dtype = dtype_util.convert_numpy_dtype_to_oneflow_dtype(item.dtype)
numpy_placeholder = oneflow_typing.Numpy.Placeholder(
shape=item.shape, dtype=of_dtype
)
data_placeholder_list.append(numpy_placeholder)
return data_placeholder_list
def FeedValueToVariable(
var_blob: Union[oneflow._oneflow_internal.EagerConsistentBlob,
"oneflow.Tensor"],
value: ValueContainer,
scope_symbol_id: Optional[int],
) -> None:
"""
Feed the value of `value` to the variable `var_blob`
"""
assert isinstance(value,
(EagerBlobTrait, FileBackendVariableBlob, np.ndarray,
oneflow.Tensor)), "Unknown value type: {}".format(
type(value).__name__)
if isinstance(value, FileBackendVariableBlob):
if not value.has_meta_info_:
value = FileBackendVariableBlob(value.var_dir_, var_blob.dtype,
var_blob.shape)
assert var_blob.shape == value.shape, "{} vs {}".format(
var_blob.shape, value.shape)
if isinstance(value, np.ndarray):
value_flow_dtype = dtype_util.convert_numpy_dtype_to_oneflow_dtype(
value.dtype)
else:
value_flow_dtype = value.dtype
assert var_blob.dtype == value_flow_dtype, "{} vs {}".format(
var_blob.dtype, value_flow_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
for start, stop, slice in _ReadSlice(value):
slice_value_blob = _GetCpu0VariableBlobFromNumpy(slice, var_blob.dtype)
_LogicalSliceAssign(
var_blob_object,
slice_value_blob.blob_object,
start,
stop,
scope_symbol_id,
)
