def var(x, axis=None, unbiased=True, keepdim=False, name=None):
"""
Computes the variance of ``x`` along ``axis`` .
Args:
x (Tensor): The input Tensor with data type float32, float64.
axis (int|list|tuple, optional): The axis along which to perform
variance calculations. ``axis`` should be int, list(int) or
tuple(int). If ``axis`` is a list/tuple of dimension(s), variance
is calculated along all element(s) of ``axis`` . ``axis`` or
element(s) of ``axis`` should be in range [-D, D), where D is the
dimensions of ``x`` . If ``axis`` or element(s) of ``axis`` is less
than 0, it works the same way as :math:`axis + D` . If ``axis`` is
None, variance is calculated over all elements of ``x``. Default
is None.
unbiased (bool, optional): Whether to use the unbiased estimation. If
``unbiased`` is True, the divisor used in the computation is
:math:`N - 1`, where :math:`N` represents the number of elements
along ``axis`` , otherwise the divisor is :math:`N`. Default is True.
keepdim (bool, optional): Whether to reserve the reduced dimension(s)
in the output Tensor. If ``keepdim`` is True, the dimensions of
the output Tensor is the same as ``x`` except in the reduced
dimensions(it is of size 1 in this case). Otherwise, the shape of
the output Tensor is squeezed in ``axis`` . Default is False.
name (str, optional): Name for the operation (optional, default is None).
For more information, please refer to :ref:`api_guide_Name`.
Returns:
Tensor, results of variance along ``axis`` of ``x``, with the same data
type as ``x``.
Examples:
.. code-block:: python
import paddle
import numpy as np
paddle.disable_static()
x = np.array([[1.0, 2.0, 3.0], [1.0, 4.0, 5.0]])
x = paddle.to_tensor(x)
out1 = paddle.var(x)
# [2.66666667]
out2 = paddle.var(x, axis=1)
# [1. 4.33333333]
"""
if not in_dygraph_mode():
check_variable_and_dtype(x, 'x', ['float32', 'float64'], 'var')
u = mean(x, axis, True, name)
out = paddle.sum((x - u)**2, axis, keepdim=keepdim, name=name)
n = paddle.cast(paddle.numel(x), x.dtype) \
/ paddle.cast(paddle.numel(out), x.dtype)
if unbiased:
one_const = paddle.ones([1], x.dtype)
n = where(n > one_const, n - 1., one_const)
out /= n
return out
def test_numel_imperative(self):
paddle.disable_static(paddle.CPUPlace())
input_1 = np.random.random([2, 1, 4, 5]).astype("int32")
input_2 = np.random.random([1, 4, 5]).astype("int32")
x_1 = paddle.to_tensor(input_1)
x_2 = paddle.to_tensor(input_2)
out_1 = paddle.numel(x_1)
out_2 = paddle.numel(x_2)
assert (np.array_equal(out_1.numpy().item(0), np.size(input_1)))
assert (np.array_equal(out_2.numpy().item(0), np.size(input_2)))
paddle.enable_static()
def var(x, axis=None, unbiased=True, keepdim=False, name=None):
u = paddle.mean(x, axis, True, name)
out = paddle.sum((x - u) * (x - u), axis, keepdim=keepdim, name=name)
n = paddle.cast(paddle.numel(x), x.dtype) \
/ paddle.cast(paddle.numel(out), x.dtype)
if unbiased:
one_const = paddle.ones([1], x.dtype)
n = paddle.where(n > one_const, n - 1., one_const)
out /= n
return out
def test_numel_static(self):
main_program = fluid.Program()
startup_program = fluid.Program()
with fluid.program_guard(main_program, startup_program):
shape1 = [2, 1, 4, 5]
shape2 = [1, 4, 5]
x_1 = paddle.fluid.data(shape=shape1, dtype='int32', name='x_1')
x_2 = paddle.fluid.data(shape=shape2, dtype='int32', name='x_2')
input_1 = np.random.random(shape1).astype("int32")
input_2 = np.random.random(shape2).astype("int32")
out_1 = paddle.numel(x_1)
out_2 = paddle.numel(x_2)
exe = paddle.static.Executor(place=paddle.CPUPlace())
res_1, res_2 = exe.run(feed={
"x_1": input_1,
"x_2": input_2,
},
fetch_list=[out_1, out_2])
assert (np.array_equal(
res_1,
np.array([np.size(input_1)]).astype("int64")))
assert (np.array_equal(
res_2,
np.array([np.size(input_2)]).astype("int64")))
def _calculate_fan_in_and_fan_out(tensor):
dimensions = len(tensor.shape)
if dimensions < 2:
raise ValueError(
"Fan in and fan out can not be computed for tensor with fewer than 2 dimensions"
)
num_input_fmaps = tensor.shape[1]
num_output_fmaps = tensor.shape[0]
receptive_field_size = 1
if len(tensor.shape) > 2:
receptive_field_size = paddle.numel(tensor[0][0])
fan_in = num_input_fmaps * receptive_field_size
fan_out = num_output_fmaps * receptive_field_size
return fan_in, fan_out
def _split_activation(tensor):
global _hcg
mp_degree = _hcg.get_model_parallel_world_size()
mp_rank = _hcg.get_model_parallel_rank()
if mp_degree < 2:
return tensor
tensor_numel = paddle.numel(tensor)
assert tensor_numel != 0, "can't recompute zero element"
assert tensor_numel % mp_degree == 0, "The capacity of the activation () cannot be divisible by mp_degree()".format(
tensor_numel, mp_degree)
# use inplace operation to save memory
data = tensor.flatten_()
part_size = tensor_numel // mp_degree
start = part_size * mp_rank
end = start + part_size
return data[start:end]
def forward(self, hidden, target, keep_order=False):
assert (hidden.shape[0] == target.shape[0])
if self.num_clusters == 0:
logit = self._compute_logits(hidden, self.out_layers_weight[0],
self.out_layers_bias[0],
self.out_projs[0])
nll = -paddle.log(F.softmax(logit, axis=-1))
idx = paddle.concat(
[
paddle.arange(0, nll.shape[0]).unsqueeze([1]),
target.unsqueeze(1)
],
axis=1)
nll = paddle.gather_nd(nll, idx)
else:
weights, biases = [], []
for i in range(len(self.cutoffs)):
if self.div_val == 1:
l_idx, r_idx = self.cutoff_ends[i], self.cutoff_ends[i + 1]
weight_i = self.out_layers_weight[0][l_idx:r_idx]
bias_i = self.out_layers_bias[0][l_idx:r_idx]
else:
weight_i = self.out_layers_weight[i]
bias_i = self.out_layers_bias[i]
if i == 0:
weight_i = paddle.concat(
[weight_i, self.cluster_weight], axis=0)
bias_i = paddle.concat([bias_i, self.cluster_bias], axis=0)
weights.append(weight_i)
biases.append(bias_i)
head_weight, head_bias, head_proj = weights[0], biases[
0], self.out_projs[0]
head_logit = self._compute_logits(hidden, head_weight, head_bias,
head_proj)
head_logprob = paddle.log(F.softmax(head_logit, axis=-1))
nll = paddle.zeros_like(target, dtype=hidden.dtype)
offset = 0
cutoff_values = [0] + self.cutoffs
for i in range(len(cutoff_values) - 1):
l_idx, r_idx = cutoff_values[i], cutoff_values[i + 1]
mask_i = paddle.cast(
target >= l_idx,
dtype=paddle.get_default_dtype()) * paddle.cast(
target < r_idx, dtype="int64")
indices_i = paddle.nonzero(mask_i).squeeze([1])
if paddle.numel(indices_i) == 0:
continue
target_i = paddle.gather(target, indices_i, axis=0) - l_idx
head_logprob_i = paddle.gather(head_logprob, indices_i, axis=0)
if i == 0:
target_i_idx = paddle.concat(
[
paddle.arange(0, head_logprob_i.shape[0]).unsqueeze(
[1]), target_i.unsqueeze([1])
],
axis=1)
logprob_i = head_logprob_i.gather_nd(target_i_idx)
else:
weight_i, bias_i, proj_i = weights[i], biases[
i], self.out_projs[i].weight if self.out_projs[
i] is not None else None
hidden_i = paddle.gather(hidden, indices_i, axis=0)
tail_logit_i = self._compute_logits(hidden_i, weight_i,
bias_i, proj_i)
tail_logprob_i = paddle.log(
F.softmax(
tail_logit_i, axis=-1))
target_i_idx = paddle.concat(
[
paddle.arange(0, tail_logprob_i.shape[0]).unsqueeze(
[1]), target_i.unsqueeze([1])
],
axis=1)
logprob_i = tail_logprob_i.gather_nd(target_i_idx)
logprob_i = head_logprob_i[:, -i] + logprob_i
if self.keep_order or keep_order:
nll = paddle.scatter(nll, indices_i, -logprob_i)
else:
index = paddle.arange(offset, offset + logprob_i.shape[0],
1)
nll = paddle.scatter(nll, index, -logprob_i)
offset += logprob_i.shape[0]
return nll
def forward(self, inputs):
"""
forward
"""
x = paddle.numel(inputs)
return x
def test_x_type():
shape = [1, 4, 5]
input_1 = np.random.random(shape).astype("int32")
out_1 = paddle.numel(input_1)
