def create_tensor(value, dtype, shape):
if framework._in_eager_mode_:
out = _C_ops.final_state_full(shape, value, dtype,
framework._current_expected_place())
else:
out = _varbase_creator(dtype=dtype)
out = _C_ops.fill_constant(out, 'dtype', dtype, 'shape', shape,
'value', value, 'force_cpu', False)
out.stop_gradient = True
return out
def __call__(self, var, block=None):
"""Initialize the input tensor with constant.
Args:
var(Tensor): Tensor that needs to be initialized.
block(Block, optional): The block in which initialization ops
should be added. Used in static graph only, default None.
Returns:
The initialization op
"""
block = self._check_block(block)
assert (isinstance(var, framework.Variable)
or isinstance(var, framework.EagerParamBase))
assert isinstance(block, framework.Block)
if framework._non_static_mode():
_C_ops.fill_constant(var, 'value', float(self._value),
'force_cpu', self._force_cpu, 'dtype',
int(var.dtype), 'str_value',
str(float(self._value)), 'shape', var.shape)
return None
else:
# fill constant should set the "str_value" to preserve precision
op = block.append_op(type="fill_constant",
outputs={"Out": var},
attrs={
"shape": var.shape,
"dtype": int(var.dtype),
"value": float(self._value),
'str_value': str(float(self._value)),
'force_cpu': self._force_cpu
},
stop_gradient=True)
var.op = op
return op
def create_tensor(value, dtype, shape):
out = _varbase_creator(dtype=dtype)
out = _C_ops.fill_constant(out, 'dtype', dtype, 'shape', shape,
'value', value, 'force_cpu', False)
out.stop_gradient = True
return out
def __call__(self, var, block=None):
"""Initialize the input tensor with dirac initializer.
Args:
var(Tensor): Tensor that needs to be initialized.
block(Block, optional): The block in which initialization ops
should be added. Used in static graph only, default None.
Returns:
The most critical OP(scatter) in this initializer, which contains 7~8 ops in total.
"""
block = self._check_block(block)
assert isinstance(var, framework.Parameter)
assert isinstance(block, framework.Block)
check_variable_and_dtype(var, "Out",
['float16', 'bfloat16', 'float32', 'float64'],
'Dirac')
assert len(var.shape) in [
3, 4, 5
], "Only Tensor with 3/4/5 dimensions can be initialized by Dirac"
assert (var.shape[0] % self._groups
) == 0, "Tensor 0-dimension must be divisible by groups"
if var.dtype != VarDesc.VarType.FP32:
out_var = block.create_var(name=unique_name.generate(".".join(
['dirac', var.name, 'tmp'])),
shape=var.shape,
dtype=VarDesc.VarType.FP32,
type=VarDesc.VarType.LOD_TENSOR,
persistable=False)
else:
out_var = var
op = None
if framework.in_dygraph_mode():
with fluid.dygraph.no_grad():
_C_ops.fill_constant(out_var, 'value', float(0), 'force_cpu',
False,
'dtype', out_var.dtype, 'str_value',
str(float(0)), 'shape', out_var.shape)
else:
block.append_op(type='fill_constant',
inputs={},
outputs={'Out': out_var},
attrs={
'value': float(0),
'dtype': out_var.dtype,
'shape': out_var.shape,
},
stop_gradient=True)
origin_shape = var.shape
num_per_group = origin_shape[0] // self._groups
min_shape = min(num_per_group, origin_shape[1])
idx_list = []
value_list = []
strides = []
prod = 1
for dim in reversed(origin_shape):
strides.insert(0, prod)
prod *= dim
for i in range(self._groups):
for j in range(min_shape):
value_list.append(1.0)
offset = 0
for (k, stride) in enumerate(strides):
if (k == 0):
offset += (j + i * num_per_group) * stride
elif (k == 1):
offset += j * stride
else:
offset += origin_shape[k] // 2 * stride
idx_list.append(offset)
if framework.in_dygraph_mode():
with fluid.dygraph.no_grad():
tmp_out, _ = _C_ops.reshape2(out_var, None, 'shape', [-1])
tmp_out._share_underline_tensor_to(out_var)
else:
x_shape = block.create_var(name=unique_name.generate(".".join(
[out_var.name, "XShape"])),
dtype=out_var.dtype,
shape=out_var.shape,
type=VarDesc.VarType.LOD_TENSOR,
persistable=False,
stop_gradient=True)
block.append_op(type="reshape2",
inputs={"X": out_var},
attrs={'shape': [-1]},
outputs={
"Out": out_var,
"XShape": x_shape
},
stop_gradient=True)
index_tensor = block.create_var(
name=unique_name.generate('scatter_index'),
persistable=False,
stop_gradient=True)
if framework.in_dygraph_mode():
with fluid.dygraph.no_grad():
tmp_tensor = framework._varbase_creator()
_C_ops.assign_value(tmp_tensor, 'shape', [len(idx_list)],
'dtype', VarDesc.VarType.INT64,
'int64_values', idx_list)
tmp_tensor._share_underline_tensor_to(index_tensor)
else:
block.append_op(type='assign_value',
outputs={'Out': index_tensor},
attrs={
'dtype': VarDesc.VarType.INT64,
'shape': [len(idx_list)],
'int64_values': idx_list
},
stop_gradient=True)
value_tensor = block.create_var(
name=unique_name.generate('scatter_value'),
persistable=False,
stop_gradient=True)
if framework.in_dygraph_mode():
with fluid.dygraph.no_grad():
tmp_tensor = framework._varbase_creator()
_C_ops.assign_value(tmp_tensor, 'shape', [len(value_list)],
'dtype', VarDesc.VarType.FP32,
'fp32_values', value_list)
tmp_tensor._share_underline_tensor_to(value_tensor)
else:
block.append_op(type='assign_value',
outputs={'Out': value_tensor},
attrs={
'dtype': VarDesc.VarType.FP32,
'shape': [len(value_list)],
'fp32_values': value_list
},
stop_gradient=True)
if framework.in_dygraph_mode():
with fluid.dygraph.no_grad():
tmp_out = _C_ops.final_state_scatter(out_var, index_tensor,
value_tensor, True)
tmp_out._share_underline_tensor_to(out_var)
tmp_reshape_out, _ = _C_ops.reshape2(out_var, None, 'shape',
origin_shape)
tmp_reshape_out._share_underline_tensor_to(out_var)
if var.dtype != VarDesc.VarType.FP32:
tmp_cast_out = _C_ops.cast(out_var, 'in_dtype',
out_var.dtype, 'out_dtype',
var.dtype)
tmp_cast_out._share_underline_tensor_to(var)
else:
op = block.append_op(type="scatter",
inputs={
"X": out_var,
"Ids": index_tensor,
"Updates": value_tensor
},
attrs={'overwrite': True},
outputs={"Out": out_var},
stop_gradient=True)
x_shape = block.create_var(name=unique_name.generate(".".join(
[out_var.name, "XShape"])),
dtype=out_var.dtype,
shape=out_var.shape,
type=VarDesc.VarType.LOD_TENSOR,
persistable=False,
stop_gradient=True)
block.append_op(type="reshape2",
inputs={"X": out_var},
attrs={'shape': origin_shape},
outputs={
"Out": out_var,
"XShape": x_shape
},
stop_gradient=True)
if var.dtype != VarDesc.VarType.FP32:
block.append_op(type="cast",
inputs={"X": out_var},
outputs={"Out": var},
attrs={
"in_dtype": out_var.dtype,
"out_dtype": var.dtype
},
stop_gradient=True)
if not in_dynamic_mode():
var.op = op
return op