def variable(self, name, shape, dtype, value=None):
if self._status == BEFORE_RUN:
assert shape is not None
if self._feed_spec is not None and value is None:
i = len(self._feed_dict)
range = self._feed_spec[i].get("range", None)
else:
range = None
feed_value = feeder.generate_random_data(shape,
dtype,
range=range,
value=value)
requires_grad = True if dtype in ["float16", "float32", "float64"
] else False
var = torch.tensor(feed_value,
requires_grad=requires_grad,
device=self._device)
if requires_grad:
var.retain_grad()
self._feed_dict[name] = var
if value is None:
self._generated_feed_values.append(feed_value)
else:
var = self._feed_dict[name]
return var
def variable(self, name, shape, dtype, value=None, stop_gradient=False):
assert shape is not None
if self._feed_spec is not None and value is None:
i = len(self._feed_dict)
range = self._feed_spec[i].get("range", None)
else:
range = None
feed_value = feeder.generate_random_data(
shape, dtype, range=range, value=value)
var = fluid.data(name=name, shape=shape, dtype=dtype, lod_level=0)
var.persistable = True
var.stop_gradient = stop_gradient
if value is None:
# When value is None, need to feed data to the variable.
self._feed_dict[var] = feed_value
return var
def variable(self, name, shape, dtype, value=None):
assert shape is not None
if self._feed_spec is not None and value is None:
i = len(self._feed_dict)
range = self._feed_spec[i].get("range", None)
else:
range = None
feed_value = feeder.generate_random_data(
shape, dtype, range=range, value=value)
if self._need_feed:
var = self.placeholder(name=name, shape=shape, dtype=dtype)
else:
var = tf.Variable(feed_value, name=name)
if value is None:
# When value is None, the variable is need to feed data.
self._feed_dict[var] = feed_value
return var
def variable(self, name, shape, dtype, value=None):
if self._status == BEFORE_RUN:
if self._feed_values is not None and value is None:
i = len(self._feed_dict)
feed_value = self._feed_values[i]
else:
assert shape is not None
if self._feed_spec is not None and value is None:
i = len(self._feed_dict)
range = self._feed_spec[i].get("range", None)
else:
range = None
feed_value = feeder.generate_random_data(shape,
dtype,
range=range,
value=value)
var = paddle.to_tensor(feed_value, stop_gradient=False)
self._feed_dict[name] = var
else:
var = self._feed_dict[name]
return var