def __measure_parameterized(self, state, which_qubits, result_desired,
theta):
r"""进行参数化的测量。
Args:
state (ComplexVariable): 输入的量子态
which_qubits (list): 测量作用的量子比特编号
result_desired (list): 期望得到的测量结果，如 ``"0"``、``"1"`` 或者 ``["0", "1"]``
theta (Variable): 测量运算的参数
Returns:
ComplexVariable: 测量坍塌后的量子态
Variable：测量坍塌得到的概率
list: 测量得到的结果（0 或 1）
"""
n = self.get_qubit_number()
assert len(which_qubits) == len(result_desired), \
"the length of qubits wanted to be measured and the result desired should be same"
op_list = [fluid.dygraph.to_variable(np.eye(2, dtype=np.complex128))
] * n
for idx in range(0, len(which_qubits)):
i = which_qubits[idx]
ele = result_desired[idx]
if int(ele) == 0:
basis0 = fluid.dygraph.to_variable(
np.array([[1, 0], [0, 0]], dtype=np.complex128))
basis1 = fluid.dygraph.to_variable(
np.array([[0, 0], [0, 1]], dtype=np.complex128))
rho0 = elementwise_mul(basis0, cos(theta[idx]))
rho1 = elementwise_mul(basis1, sin(theta[idx]))
rho = elementwise_add(rho0, rho1)
op_list[i] = rho
elif int(ele) == 1:
# rho = diag(concat([cos(theta[idx]), sin(theta[idx])]))
# rho = ComplexVariable(rho, zeros((2, 2), dtype="float64"))
basis0 = fluid.dygraph.to_variable(
np.array([[1, 0], [0, 0]], dtype=np.complex128))
basis1 = fluid.dygraph.to_variable(
np.array([[0, 0], [0, 1]], dtype=np.complex128))
rho0 = elementwise_mul(basis0, sin(theta[idx]))
rho1 = elementwise_mul(basis1, cos(theta[idx]))
rho = elementwise_add(rho0, rho1)
op_list[i] = rho
else:
print("cannot recognize the results_desired.")
# rho = ComplexVariable(ones((2, 2), dtype="float64"), zeros((2, 2), dtype="float64"))
measure_operator = fluid.dygraph.to_variable(op_list[0])
if n > 1:
for idx in range(1, len(op_list)):
measure_operator = kron(measure_operator, op_list[idx])
state_measured = matmul(matmul(measure_operator, state),
dagger(measure_operator))
prob = trace(
matmul(matmul(dagger(measure_operator), measure_operator),
state)).real
state_measured = elementwise_div(state_measured, prob)
return state_measured, prob, result_desired
def test_sin(self):
program = Program()
with program_guard(program):
input = layers.data(name="input", shape=[16], dtype="float32")
out = layers.sin(input, name='sin')
self.assertIsNotNone(out)
print(str(program))
def compute_position_embedding(radians, speaker_position_rate):
"""Compute sin/cos interleaved matrix from the radians.
Arg:
radians (Variable): shape(n_vocab, embed_dim), dtype float32, the radians matrix.
speaker_position_rate (Variable): shape(B, ), speaker positioning rate.
Returns:
Variable: shape(B, n_vocab, embed_dim), the sin, cos interleaved matrix.
"""
_, embed_dim = radians.shape
batch_size = speaker_position_rate.shape[0]
scaled_radians = F.elementwise_mul(F.expand(F.unsqueeze(radians, [0]),
[batch_size, 1, 1]),
speaker_position_rate,
axis=0)
odd_mask = (np.arange(embed_dim) % 2).astype(np.float32)
odd_mask = dg.to_variable(odd_mask)
out = odd_mask * F.cos(scaled_radians) \
+ (1 - odd_mask) * F.sin(scaled_radians)
out = F.concat(
[F.zeros((batch_size, 1, embed_dim), radians.dtype), out[:, 1:, :]],
axis=1)
return out
def get_embedding(self, num_embeddings,
embedding_dim, padding_idx=None):
"""
Build sinusoidal embeddings.
This matches the implementation in tensor2tensor,
but differs slightly from the description
in Section 3.5 of "Attention Is All You Need".
"""
half_dim = embedding_dim // 2
emb = layers.log(float(10000)) / (half_dim - -1)
emb = layers.exp(layers.arange(
start=0, end=half_dim, dtype='float32') * -emb)
# [num_embeddings, embedding_dim // 2]
emb = layers.unsqueeze(layers.arange(-num_embeddings // 2,
num_embeddings // 2, dtype='float32'), axis=1) *\
layers.unsqueeze(emb, axis=0)
emb = layers.concat([layers.sin(emb), layers.cos(emb)], dim=1)
# [num_embeddings, embedding_dim]
if embedding_dim % 2 == 1:
emb = layers.concat(
[emb, layers.zeros(shape=(num_embeddings, 1))], dim=1)
if padding_idx is not None:
emb[paddings_idx, :] = 0
self.origin_shift = num_embeddings // 2
return emb
def forward(self, tensor_list: NestedTensor):
x = tensor_list.tensors
mask = tensor_list.mask
assert mask is not None
bs, h, w = mask.shape
mask = mask.numpy()
not_mask = ~mask
not_mask = dg.to_variable(not_mask).astype('float32')
y_embed = L.cumsum(not_mask, axis=1) # [batch_size, h, w]
x_embed = L.cumsum(not_mask, axis=2) # [batch_size, h, w]
if self.normalize:
eps = 1e-6
y_embed = y_embed / (y_embed[:, -1:, :] + eps) * self.scale
x_embed = x_embed / (x_embed[:, :, -1:] + eps) * self.scale
dim_t = (np.arange(0, self.num_pos_feats, 1,
dtype="float32")) # [num_pos_feats]
dim_t = self.temperature**(2 * (dim_t // 2) / self.num_pos_feats
) # [num_pos_feats]
dim_t = dg.to_variable(dim_t)
x_embed = L.unsqueeze(x_embed, 3) # [batch_size, h, w, 1]
y_embed = L.unsqueeze(y_embed, 3) # [batch_size, h, w, 1]
pos_x = x_embed / dim_t # [batch_size, h, w, num_pos_feats]
pos_y = y_embed / dim_t # [batch_size, h, w, num_pos_feats]
pos_x_1 = L.sin(pos_x[:, :, :,
0::2]) # [batch_size, h, w, num_pos_feats / 2]
pos_x_2 = L.cos(pos_x[:, :, :,
1::2]) # [batch_size, h, w, num_pos_feats / 2]
pos_y_1 = L.sin(pos_y[:, :, :,
0::2]) # [batch_size, h, w, num_pos_feats / 2]
pos_y_2 = L.cos(pos_y[:, :, :,
1::2]) # [batch_size, h, w, num_pos_feats / 2]
pos_x = L.reshape(L.stack([pos_x_1, pos_x_2], axis=4),
(bs, h, w, -1)) # [batch_size, h, w, num_pos_feats]
pos_y = L.reshape(L.stack([pos_y_1, pos_y_2], axis=4),
(bs, h, w, -1)) # [batch_size, h, w, num_pos_feats]
pos = L.concat((pos_y, pos_x),
axis=3) # [batch_size, h, w, num_pos_feats * 2]
pos = L.transpose(pos,
perm=(0, 3, 1,
2)) # [batch_size, num_pos_feats * 2, h, w]
return pos
def rotation_y(theta):
"""
:param theta: must be a scale, shape [1], 'float32'
:return:
"""
cos_value = cos(theta/2)
sin_value = sin(theta/2)
ry_re = concat([cos_value, -sin_value, sin_value, cos_value], axis=0)
ry_im = pp_zeros([2, 2], "float32")
return ComplexVariable(reshape(ry_re, [2, 2]), ry_im)
def rotation_z(theta):
"""
:param theta: must be a scale, shape [1], 'float32'
:return:
"""
cos_value = cos(theta/2)
sin_value = sin(theta/2)
zero_pd = pp_zeros([1], "float32")
rz_re = concat([cos_value, zero_pd, zero_pd, cos_value], axis=0)
rz_im = concat([-sin_value, zero_pd, zero_pd, sin_value], axis=0)
return ComplexVariable(reshape(rz_re, [2, 2]), reshape(rz_im, [2, 2]))
def rotation_y(theta):
r"""生成单量子比特Y门
Args:
theta (Variable): Y门的旋转角度
Returns:
ComplexVariable: 单量子比特Y门的矩阵形式
"""
cos_value = cos(theta / 2)
sin_value = sin(theta / 2)
ry_re = concat([cos_value, -sin_value, sin_value, cos_value], axis=0)
ry_im = pp_zeros([2, 2], "float32")
return ComplexVariable(reshape(ry_re, [2, 2]), ry_im)
def rotation_z(theta):
r"""生成单量子比特Z门
Args:
theta (Variable): Z门的旋转角度
Returns:
ComplexVariable: 单量子比特Z门的矩阵形式
"""
cos_value = cos(theta / 2)
sin_value = sin(theta / 2)
zero_pd = pp_zeros([1], "float32")
rz_re = concat([cos_value, zero_pd, zero_pd, cos_value], axis=0)
rz_im = concat([-sin_value, zero_pd, zero_pd, sin_value], axis=0)
return ComplexVariable(reshape(rz_re, [2, 2]), reshape(rz_im, [2, 2]))
def compute_position_embedding_single_speaker(radians, speaker_position_rate):
"""Compute sin/cos interleaved matrix from the radians.
Arg:
radians (Variable): shape(n_vocab, embed_dim), dtype float32, the radians matrix.
speaker_position_rate (float or Variable): float or Variable of shape(1, ), speaker positioning rate.
Returns:
Variable: shape(n_vocab, embed_dim), the sin, cos interleaved matrix.
"""
_, embed_dim = radians.shape
scaled_radians = radians * speaker_position_rate
odd_mask = (np.arange(embed_dim) % 2).astype(np.float32)
odd_mask = dg.to_variable(odd_mask)
out = odd_mask * F.cos(scaled_radians) \
+ (1 - odd_mask) * F.sin(scaled_radians)
return out
def positional_encoding(tensor, start_index, omega):
"""
tensor: a reference tensor we use to get shape. actually only T and C are needed. Shape(B, T, C)
start_index: int, we can actually use start and length to specify them.
omega (B,): speaker position rates
return (B, T, C), position embedding
"""
dtype = omega.dtype
_, length, dimension = tensor.shape
index = F.range(start_index, start_index + length, 1, dtype=dtype)
channel = F.range(0, dimension, 2, dtype=dtype)
p = F.unsqueeze(omega, [1, 2]) \
* F.unsqueeze(index, [1]) \
/ (10000 ** (channel / float(dimension)))
encodings = F.concat([F.sin(p), F.cos(p)], axis=2)
return encodings