mindspore-Ecosystem
/
mindspore
21
0
Fork 27
Code Issues Pull Requests 6 Projects Releases Wiki Activity

delete pack description about num and add raise error 

SparseApplyAdagrad example format

PReLU and grad state 1-d not supported

atan2 bprop dx dy revert to original shapes
This commit is contained in:
zhaozhenlong 2020-04-29 20:58:06 +08:00
parent d17d61265d
commit 4694c979a5
5 changed files with 52 additions and 28 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
mindspore/ops/_grad/grad_math_ops.py
View File

@ -788,7 +788,7 @@ def get_bprop_atan2(self):
def bprop(x, y, out, dout):
tmp = dout / (square(x) + square(y))
dx = tmp * y
dy = tmp * (-x)
return (dx, dy)
bc_dx = tmp * y
bc_dy = tmp * (-x)
return binop_grad_common(x, y, bc_dx, bc_dy)
return bprop

5
mindspore/ops/operations/_grad_ops.py
View File

@ -669,6 +669,9 @@ class PReLUGrad(PrimitiveWithInfer):
r"""
Gradients of PReLU operation.
Note:
1-dimensional input_x is not supported.
Inputs:
- **y_backprop** (Tensor) - Representing the backprop of the next layer.
- **input_x** (Tensor) - Should be the input `input_x` of forward operator PRelu.
@ -683,6 +686,8 @@ class PReLUGrad(PrimitiveWithInfer):
pass
def infer_shape(self, y_backprop_shape, A_shape, w_shape):
if len(A_shape) == 1:
raise ValueError(f'For \'{self.name}\' input_x rank 1 is not supported.')
return y_backprop_shape, w_shape
def infer_dtype(self, y_backprop_dtype, A_dtype, w_dtype):

17
mindspore/ops/operations/array_ops.py
View File

@ -1308,8 +1308,8 @@ class Concat(PrimitiveWithInfer):
def _get_pack_shape(x_shape, x_type, axis, prim_name):
"""for pack output shape"""
validator.check_value_type("shape", x_shape, [tuple, list], prim_name)
validator.check_integer("len of input_x shape", len(x_shape), 0, Rel.GT, prim_name)
validator.check_subclass("shape0", x_type[0], mstype.tensor, prim_name)
validator.check_integer("len of input_x", len(x_shape), 1, Rel.GT, prim_name)
validator.check_subclass("input_x[0]", x_type[0], mstype.tensor, prim_name)
validator.check_integer("len of input_x0 shape", len(x_shape[0]), 0, Rel.GT, prim_name)
rank_base = len(x_shape[0])
N = len(x_shape)
@ -1320,7 +1320,7 @@ def _get_pack_shape(x_shape, x_type, axis, prim_name):
for i in range(1, N):
v = x_shape[i]
validator.check('len of x_shape[%d]' % i, len(v), 'len of rank_base', rank_base, Rel.EQ, prim_name)
validator.check('x_type[%d]' % i, x_type[i], 'base', x_type[0], Rel.EQ, prim_name)
validator.check('x_type[%d]' % i, x_type[i], 'base', x_type[0], Rel.EQ, prim_name, TypeError)
for j in range(rank_base):
if v[j] != x_shape[0][j]:
raise ValueError(f"For \'{prim_name}\' element {i} shape in input can not pack with first element")
@ -1346,6 +1346,12 @@ class Pack(PrimitiveWithInfer):
Outputs:
Tensor. A packed Tensor with the same type as `input_x`.
Raises:
TypeError: If the data types of elements in input_x are not the same.
ValueError: If length of input_x is not greater than 1;
or if axis is out of the range [-(R+1), R+1);
or if the shapes of elements in input_x are not the same.
Examples:
>>> data1 = Tensor(np.array([0, 1]).astype(np.float32))
>>> data2 = Tensor(np.array([2, 3]).astype(np.float32))
@ -1386,8 +1392,6 @@ class Unpack(PrimitiveWithInfer):
Args:
axis (int): Dimension along which to pack. Default: 0.
Negative values wrap around. The range is [-R, R).
num (int): The number of tensors to be unpacked to. Default : "None".
If `num` is not specified, it is inferred from the shape of `input_x`.
Inputs:
- **input_x** (Tensor) - The shape is :math:`(x_1, x_2, ..., x_R)`.
@ -1397,8 +1401,7 @@ class Unpack(PrimitiveWithInfer):
A tuple of Tensors, the shape of each objects is same.
Raises:
ValueError: If axis is out of the range [-len(input_x.shape()), len(input_x.shape())),
or if len(input_x.shape[axis]) not equal to num.
ValueError: If axis is out of the range [-len(input_x.shape()), len(input_x.shape())).
Examples:
>>> unpack = P.Unpack()

24
mindspore/ops/operations/nn_ops.py
View File

@ -2087,6 +2087,9 @@ class PReLU(PrimitiveWithInfer):
where :math:`x_i` is an element of an channel of the input.
Note:
1-dimensional input_x is not supported.
Inputs:
- **input_x** (Tensor) - Float tensor, representing the output of the preview layer.
- **weight** (Tensor) - Float Tensor, w > 0, there is only two shapes are legitimate,
@ -2106,14 +2109,13 @@ class PReLU(PrimitiveWithInfer):
input_x_dim = len(input_x_shape)
weight_dim = len(weight_shape)
if input_x_dim == 1:
raise ValueError(f'For \'{self.name}\' input_x rank 1 is not supported.')
if weight_dim != 1:
raise ValueError(f'For \'{self.name}\' weight_dim must be 1, while weight_dim is {weight_dim}.')
if input_x_dim == 1 and weight_shape[0] != 1:
raise ValueError(f'For \'{self.name}\' when input_x_dim is 1, weight_shape[0] must be 1, '
f'while weight_shape[0] is {weight_shape[0]}.')
if input_x_dim != 1 and weight_shape[0] != input_x_shape[1] and weight_shape[0] != 1:
if weight_shape[0] != input_x_shape[1] and weight_shape[0] != 1:
raise ValueError(f'For \'{self.name}\' channel of input_x and weight must be matched,'
f' while channel of input_x is {input_x_shape[1]},'
f' weight_shape[0] is {weight_shape[0]}.')
@ -2556,12 +2558,12 @@ class SparseApplyAdagrad(PrimitiveWithInfer):
Tensor, has the same shape and type as `var`.
Examples:
var = Tensor(np.random.random((3, 3)), mindspore.float32)
accum = Tensor(np.random.random((3, 3)), mindspore.float32)
grad = Tensor(np.random.random((3, 3)), mindspore.float32)
indices = Tensor(np.ones((3,), np.int32))
sparse_apply_ada_grad = P.SparseApplyAdagrad(0.5)
sparse_apply_ada_grad(var, accum, grad, indices)
>>> var = Tensor(np.random.random((3, 3)), mindspore.float32)
>>> accum = Tensor(np.random.random((3, 3)), mindspore.float32)
>>> grad = Tensor(np.random.random((3, 3)), mindspore.float32)
>>> indices = Tensor(np.ones((3,), np.int32))
>>> sparse_apply_ada_grad = P.SparseApplyAdagrad(0.5)
>>> sparse_apply_ada_grad(var, accum, grad, indices)
"""
@prim_attr_register

28
tests/ut/python/ops/test_ops.py
View File

@ -1033,11 +1033,6 @@ test_case_array_ops = [
'desc_bprop':[[3, 2, 3, 3]],
}),
('Pack_2', {
'block': NetForPackInput(P.Pack()),
'desc_inputs':[[2, 2]],
'desc_bprop':[[1, 2, 2]],
}),
('Pack_3', {
'block': NetForPackInput(P.Pack()),
'desc_inputs':[[128, 128], [128, 128]],
'desc_bprop':[[2, 128, 128]],
@ -1052,16 +1047,26 @@ test_case_array_ops = [
'desc_inputs':[Tensor(np.array([[1, 1, 1]], np.float32))],
'desc_bprop':[[1], [1], [1]],
}),
('Diag', {
('Diag_1', {
'block': P.Diag(),
'desc_inputs': [[4]],
'desc_bprop': [[4, 4]],
}),
('DiagPart', {
('Diag_2', {
'block': P.Diag(),
'desc_inputs': [[4, 4]],
'desc_bprop': [[4, 4, 4, 4]],
}),
('DiagPart_1', {
'block': P.DiagPart(),
'desc_inputs': [[4, 4]],
'desc_bprop': [[4]],
}),
('DiagPart_2', {
'block': P.DiagPart(),
'desc_inputs': [[4, 4, 4, 4]],
'desc_bprop': [[4, 4]],
}),
('SpaceToBatch_1', {
'block': P.SpaceToBatch(2, [[0, 0], [0, 0]]),
'desc_inputs': [[1, 3, 2, 2]],
@ -1200,6 +1205,15 @@ raise_set = [
Tensor(np.ones((2, 2), np.int32)),
Tensor(np.ones((2,), np.float32))),
'desc_bprop': [[2, 3]]}),
('Pack', {
'block': (NetForPackInput(P.Pack()), {'exception': ValueError}),
'desc_inputs':[[2, 2]],
'desc_bprop':[[1, 2, 2]]}),
('PReLU', {
'block': (P.PReLU(), {'exception': ValueError}),
'desc_inputs':[[2], [1]],
'desc_bprop':[[1]]}),
]