pgjmblvch
/
mindspore2022
forked from huawei/mindspore2022
0
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

delete duplication_code, and fix the format and line too long problem.

This commit is contained in:
wangshuide2020 2021-05-21 20:24:26 +08:00
parent 0fe1b5d9be
commit 3551ee5a27
28 changed files with 198 additions and 232 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
mindspore/nn/layer/basic.py
View File

@ -786,9 +786,9 @@ class Unfold(Cell):
raise ValueError(f"For \'{prim_name}\' the format of {arg_name}s should be [1, {arg_name}_row, "
f"{arg_name}_col, 1], but got {arg_val}.")
if not isinstance(arg_val[1], int) or not isinstance(arg_val[2], int) or arg_val[1] < 1 or arg_val[2] < 1:
raise ValueError(f"For '{prim_name}' the {arg_name}_row and {arg_name}_col in {arg_name}s should be an "
f"positive integer number, but got {arg_name}_row is {arg_val[1]}, {arg_name}_col "
f"is {arg_val[2]}")
raise ValueError(f"For '{prim_name}' the {arg_name}_row and {arg_name}_col in {arg_name}s should be "
f"an positive integer number, but got {arg_name}_row is {arg_val[1]}, "
f"{arg_name}_col is {arg_val[2]}")
_check_tuple_or_list("ksize", ksizes, self.cls_name)
_check_tuple_or_list("stride", strides, self.cls_name)

23
mindspore/ops/_grad/grad_array_ops.py
View File

@ -18,7 +18,6 @@
import numpy as np
import mindspore as ms
from mindspore.ops import composite as C
from mindspore.common.tensor import Tensor
from .. import operations as P
from ..operations import _grad_ops as G
from ..operations import _inner_ops as inner
@ -30,6 +29,7 @@ from ..primitive import constexpr
from ... import context
from ...common import dtype as mstype
from ...common.tensor import RowTensor
from .._utils.utils import range_op, get_1d_shape
reduce_sum = P.ReduceSum()
unsorted_segment_sum = P.UnsortedSegmentSum()
@ -495,22 +495,6 @@ def get_bprop_sparse_gather_v2(self):
return bprop
@constexpr
def _range_op(start, limit, delta, dtype):
"""helper function for grad of Sort"""
output_tensor = Tensor(list(range(start, limit, delta)), dtype)
return output_tensor
@constexpr
def _get_1d_shape(in_shape):
"""helper function for grad of Sort"""
out_shape = 1
for i in in_shape:
out_shape *= i
return (out_shape,)
@constexpr
def _get_transposition(axis, rank):
"""helper function for grad of Sort"""
@ -557,13 +541,12 @@ def get_bprop_sort(self):
ind_2d = reshape_op(indices, (-1, ind_lastdim))
outer_dim = ind_2d.shape[0]
# [0, outterdim, 2*outerdim, ..., (k-1)*outerdim]
indices_dtype = dtype(indices)
range_flatten_index = _range_op(0, outer_dim * in_lastdim, in_lastdim, indices_dtype)
range_flatten_index = range_op(0, outer_dim * in_lastdim, in_lastdim, indices_dtype)
# expand_dims to (k, 1), then broadcast
ind = reshape_op(ind_2d + expand_dims(range_flatten_index, -1), (-1,))
x_shape_1d = _get_1d_shape(top_k_input_shape)
x_shape_1d = get_1d_shape(top_k_input_shape)
if transposition is not None:
dvalue = tranpose(dvalue, invert_permutation(transposition))

1
mindspore/ops/_grad/grad_math_ops.py
View File

@ -417,6 +417,7 @@ def get_bprop_xlogy(self):
return bprop
@bprop_getters.register(P.SquareSumAll)
def get_bprop_square_sum_all(self):
"""Grad definition for `Square` operation."""

23
mindspore/ops/_grad/grad_nn_ops.py
View File

@ -16,7 +16,6 @@
"""Define the grad rules of neural network related operations."""
import os
from mindspore.ops.primitive import constexpr
from mindspore.common.tensor import Tensor
from mindspore.ops.operations import nn_ops as nps
from .grad_base import bprop_getters
from .. import functional as F
@ -26,6 +25,7 @@ from ..composite.multitype_ops.zeros_like_impl import zeros_like
from ..operations import _grad_ops as G
from ..operations import _inner_ops as inner
from ... import context
from .._utils.utils import range_op, get_1d_shape
env_force_bprop_seq = os.getenv("ENV_FORCE_BPROP_SEQ")
@ -669,6 +669,7 @@ def get_bprop_fast_gelu_2(self):
return bprop
@bprop_getters.register(P.InstanceNorm)
def get_bprop_instance_norm(self):
"""Grad definition for `InstanceNorm` operation."""
@ -815,22 +816,6 @@ def get_bprop_onehot(self):
return bprop
@constexpr
def _range_op(start, limit, delta, dtype):
"""helper function for Grad TopK"""
output_tensor = Tensor(list(range(start, limit, delta)), dtype)
return output_tensor
@constexpr
def _get_1d_shape(in_shape):
"""helper function for Grad TopK"""
out_shape = 1
for i in in_shape:
out_shape *= i
return (out_shape,)
@bprop_getters.register(P.TopK)
def get_bprop_top_kv2(self):
"""Grad definition for `TopK` operation."""
@ -853,11 +838,11 @@ def get_bprop_top_kv2(self):
# [0, outterdim, 2*outerdim, ..., (k-1)*outerdim]
indices_dtype = dtype(indices)
range_flatten_index = _range_op(0, outerdim * in_lastdim, in_lastdim, indices_dtype)
range_flatten_index = range_op(0, outerdim * in_lastdim, in_lastdim, indices_dtype)
# expand_dims to (k, 1), then broadcast
ind = reshape_op(ind_2d + expand_dims(range_flatten_index, -1), (-1,))
in_shape_1d = _get_1d_shape(in_shape)
in_shape_1d = get_1d_shape(in_shape)
out_grad = reshape_op(
scatter(

2
mindspore/ops/_grad/grad_quant_ops.py
View File

@ -34,6 +34,7 @@ def get_bprop_fakequant_with_minmax(self):
return bprop
@bprop_getters.register(Q.FakeQuantWithMinMaxVars)
def get_bprop_fakequant_with_minmax_vars(self):
"""Generate bprop for FakeQuantWithMinMaxVars for Ascend"""
@ -184,6 +185,7 @@ def get_bprop_acts_ulq(self):
op = Q.ActsULQInputGrad()
op1 = Q.ActULQClampMinGrad()
op2 = Q.ActULQClampMaxGrad()
def bprop(x, clamp_min, clamp_max, out, dout):
dx = op(dout[0], out[1], out[2])
dx1 = op1(dout[0], out[1], out[3])

1
mindspore/ops/_register_for_op.py
View File

@ -23,6 +23,7 @@ class Registry(UserDict):
"""Registry class for registry functions for grad and vm_impl on Primitive."""
def register(self, prim):
"""register the function."""
def deco(fn):
"""Decorate the function."""
if isinstance(prim, str):

18
mindspore/ops/_utils/utils.py
View File

@ -15,9 +15,11 @@
"""utils for operator"""
from mindspore.common.tensor import Tensor
from ..._checkparam import Validator as validator
from ..._checkparam import Rel
from ...common import dtype as mstype
from ..primitive import constexpr
def get_broadcast_shape(x_shape, y_shape, prim_name):
@ -89,3 +91,19 @@ def get_concat_offset(x_shp, x_type, axis, prim_name):
else:
all_shp += v[axis]
return offset, all_shp, axis
@constexpr
def range_op(start, limit, delta, dtype):
"""helper function to get tensor in specified range."""
output_tensor = Tensor(list(range(start, limit, delta)), dtype)
return output_tensor
@constexpr
def get_1d_shape(in_shape):
"""helper function to get 1d shape."""
out_shape = 1
for i in in_shape:
out_shape *= i
return (out_shape,)

6
mindspore/ops/composite/base.py
View File

@ -194,9 +194,9 @@ class GradOperation(GradOperation_):
sens_param (bool): Whether to append sensitivity (gradient with respect to output) as input.
If sens_param is False, a 'ones_like(outputs)' sensitivity will be attached automatically.
Default: False.
If the sensor_param is True, a sensitivity (gradient with respect to output) needs to be transferred through
the location parameter or key-value pair parameter. If the value is transferred through the key-value pair
parameter, the key must be sens.
If the sensor_param is True, a sensitivity (gradient with respect to output) needs to be transferred
through the location parameter or key-value pair parameter. If the value is transferred through
the key-value pair parameter, the key must be sens.
Returns:
The higher-order function which takes a function as argument and returns gradient function for it.

11
mindspore/ops/composite/math_ops.py
View File

@ -441,6 +441,7 @@ def _check_batch_size(x1_batch_size, x2_batch_size):
if x1_batch_size != x2_batch_size:
raise ValueError("Require both inputs with the same batch sizes.")
@constexpr
def _get_output_shape(batch_size, x1_ret, x2_ret):
"""
@ -543,20 +544,24 @@ def batch_dot(x1, x2, axes=None):
return final_result
@constexpr
def _check_same_type(dtype1, dtype2):
return dtype1 == dtype2
@constexpr
def _max(*args):
"""Returns the maximum value."""
return max(*args)
@constexpr
def _min(*args):
"""Returns the minimum value."""
return min(*args)
@constexpr
def _infer_shape_rem(shape1, shape2, ndim1, ndim2, transpose_b):
"""Infers the shape of the last two dimensions after performing matmul."""
@ -571,6 +576,7 @@ def _infer_shape_rem(shape1, shape2, ndim1, ndim2, transpose_b):
shape_rem.append(shape2[-1])
return tuple(shape_rem)
@constexpr
def _check_matmul_shapes(shape1, shape2):
"""Checks shape1 and shape2 are valid to perform matmul, and returns output shape after broadcasting."""
@ -588,6 +594,7 @@ def _check_matmul_shapes(shape1, shape2):
shape_out.appendleft(max_size)
return tuple(shape_out)
@constexpr
def _tile_size(shape, out_shape, ndim):
"""Returns tile_size such that shape*tile_size = out_shape"""
@ -597,22 +604,26 @@ def _tile_size(shape, out_shape, ndim):
size[idx] = j
return tuple(size)
@constexpr
def _check_need_broadcast(shape1, shape2):
"""Returns True if broadcast is necessary for batchmatmul."""
return shape1[:-2] != shape2[:-2]
def _expand(x, ndim):
"""Expand x to ndim from axis, which can be 0 or -1."""
while F.rank(x) < ndim:
x = F.expand_dims(x, 0)
return x
def _broadcast_to(x, shape_cur, shape_to, ndim_to):
"""Broadcasts x from shape_cur to shape_to."""
size = _tile_size(shape_cur, shape_to, ndim_to)
return F.tile(x, size)
def matmul(x1, x2, dtype=None):
"""
Returns the matrix product of two arrays.

6
mindspore/ops/composite/multitype_ops/_compile_utils.py
View File

@ -217,8 +217,10 @@ def _transform_indexing_tensor(broadcast_shape, final_shape, new_shape, item):
def _transform_ellipsis_to_slice(data, tuple_index, op_name):
"""Check if the tuple index len is longer than the data's dims and transform ellipsis in the indices
to several slice"""
"""
Check if the tuple index len is longer than the data's dims and transform ellipsis in the indices
to several slice.
"""
data_shape = F.shape(data)
data_rank = len(data_shape)
indexes_types = hyper_map(F.typeof, tuple_index)

2
mindspore/ops/composite/multitype_ops/add_impl.py
View File

@ -268,6 +268,7 @@ def _add_umonad_umonad(x, y):
"""
return x
@_add_backward.register("IOMonad", "IOMonad")
def _add_iomonad_iomonad(x, y):
"""
@ -282,6 +283,7 @@ def _add_iomonad_iomonad(x, y):
"""
return x
@_add_backward.register("RowTensor", "Tensor")
def _add_rowtensor_tensor(x, y):
"""

1
mindspore/ops/composite/multitype_ops/greater_equal_impl.py
View File

@ -36,6 +36,7 @@ def _greater_equal_scala(x, y):
"""
return F.scalar_ge(x, y)
@greater_equal.register("Tensor", "Number")
@greater_equal.register("Number", "Tensor")
@greater_equal.register("Tensor", "Tensor")

1
mindspore/ops/composite/multitype_ops/greater_impl.py
View File

@ -36,6 +36,7 @@ def _greater_scalar(x, y):
"""
return F.scalar_gt(x, y)
@greater.register("Tensor", "Number")
@greater.register("Number", "Tensor")
@greater.register("Tensor", "Tensor")

1
mindspore/ops/composite/multitype_ops/less_equal_impl.py
View File

@ -36,6 +36,7 @@ def _less_equal_scala(x, y):
"""
return F.scalar_le(x, y)
@less_equal.register("Tensor", "Number")
@less_equal.register("Number", "Tensor")
@less_equal.register("Tensor", "Tensor")

1
mindspore/ops/composite/multitype_ops/less_impl.py
View File

@ -36,6 +36,7 @@ def _less_scala(x, y):
"""
return F.scalar_lt(x, y)
@less.register("Tensor", "Number")
@less.register("Number", "Tensor")
@less.register("Tensor", "Tensor")

1
mindspore/ops/composite/multitype_ops/uadd_impl.py
View File

@ -20,6 +20,7 @@ from mindspore.ops.composite import base
# using ".register" decorator
uadd = base.MultitypeFuncGraph("uadd", True)
@uadd.register("Tensor")
@uadd.register("Number")
def _uadd_scala(x):

3
mindspore/ops/op_info_register.py
View File

@ -382,7 +382,8 @@ class TBERegOp(RegOp):
def compute_cost(self, compute_cost):
"""
Define the calculation efficiency of operator, which refers to the value of the cost model in the tiling module.
Define the calculation efficiency of operator, which refers to the value of the cost model
in the tiling module.
Args:
compute_cost (int): Value of compute cost. Default: 10.

11
mindspore/ops/operations/__init__.py
View File

@ -31,9 +31,9 @@ from .array_ops import (Argmax, Argmin, Cast, Concat, Pack, Stack, Unpack, Unsta
ScatterNdAdd, ScatterNdSub, ScatterNonAliasingAdd, ReverseV2, Rint,
Squeeze, StridedSlice, Tile, TensorScatterUpdate, EditDistance, Sort,
Transpose, TruncatedNormal, TupleToArray, UnsortedSegmentMin, UnsortedSegmentMax,
UnsortedSegmentProd, UnsortedSegmentSum, SpaceToDepth, DepthToSpace, SpaceToBatch, BatchToSpace,
SpaceToBatchND, BatchToSpaceND, BroadcastTo, InplaceUpdate, ReverseSequence, EmbeddingLookup,
Unique, GatherD, Identity, Range)
UnsortedSegmentProd, UnsortedSegmentSum, SpaceToDepth, DepthToSpace, SpaceToBatch,
BatchToSpace, SpaceToBatchND, BatchToSpaceND, BroadcastTo, InplaceUpdate, ReverseSequence,
EmbeddingLookup, Unique, GatherD, Identity, Range)
from .comm_ops import (AllGather, AllReduce, _AlltoAll, AllSwap, ReduceScatter, Broadcast,
_MirrorOperator, _MirrorMiniStepOperator, _MiniStepAllGather, ReduceOp, _VirtualDataset,
_VirtualOutput, _VirtualDiv, _GetTensorSlice, _VirtualAdd,
@ -63,8 +63,9 @@ from .math_ops import (Abs, ACos, Asin, Asinh, AddN, AccumulateNV2, AssignAdd, A
from .random_ops import (RandomChoiceWithMask, StandardNormal, Gamma, Poisson, UniformInt, UniformReal,
RandomCategorical, StandardLaplace, Multinomial, UniformCandidateSampler,
LogUniformCandidateSampler)
from .nn_ops import (LSTM, SGD, Adam, FusedSparseAdam, FusedSparseLazyAdam, AdamNoUpdateParam, ApplyMomentum, BatchNorm,
BiasAdd, Conv2D, Conv3D, Conv3DTranspose,
from .nn_ops import (LSTM, SGD, Adam, FusedSparseAdam, FusedSparseLazyAdam, AdamNoUpdateParam, ApplyMomentum,
BatchNorm, BiasAdd, Conv2D, Conv3D, Conv3DTranspose,
DepthwiseConv2dNative,
DropoutDoMask, Dropout, Dropout2D, Dropout3D, DropoutGenMask, Flatten,
InstanceNorm, BNTrainingReduce, BNTrainingUpdate,

41
mindspore/ops/operations/_grad_ops.py
View File

@ -148,7 +148,8 @@ class RsqrtGrad(PrimitiveWithInfer):
def infer_dtype(self, x_dtype, dout_dtype):
args = {"x": x_dtype, "dout": dout_dtype}
validator.check_tensors_dtypes_same_and_valid(args, [mstype.float16, mstype.float32, mstype.int32, mstype.int8],
validator.check_tensors_dtypes_same_and_valid(args,
[mstype.float16, mstype.float32, mstype.int32, mstype.int8],
self.name)
return x_dtype
@ -971,6 +972,7 @@ def _get_max_pool3d_grad_pads_by_pad_mode(input_shape, kernel_size, strides, pad
pads = pads_d + pads_h + pads_w
return pads
class MaxPool3DGrad(PrimitiveWithInfer):
"""Gradients of the max pool3d operation."""
@ -981,7 +983,8 @@ class MaxPool3DGrad(PrimitiveWithInfer):
validator.check_value_type('pad_mode', pad_mode, [str], self.name)
self.format = validator.check_string(data_format, ['NCDHW'], 'format', self.name)
self.pad_mode = validator.check_string(pad_mode.upper(), ['VALID', 'SAME'], 'pad_mode', self.name)
self.kernel_size = _check_3d_int_or_tuple("kernel_size", kernel_size, self.name, allow_five=True, ret_five=True)
self.kernel_size = _check_3d_int_or_tuple("kernel_size", kernel_size, self.name,
allow_five=True, ret_five=True)
self.add_prim_attr("kernel_size", self.kernel_size)
self.strides = _check_3d_int_or_tuple("strides", strides, self.name, allow_five=True, ret_five=True)
self.add_prim_attr("strides", self.strides)
@ -1006,7 +1009,8 @@ class MaxPool3DGradGrad(PrimitiveWithInfer):
validator.check_value_type('pad_mode', pad_mode, [str], self.name)
self.format = validator.check_string(data_format, ['NCDHW'], 'format', self.name)
self.pad_mode = validator.check_string(pad_mode.upper(), ['VALID', 'SAME'], 'pad_mode', self.name)
self.kernel_size = _check_3d_int_or_tuple("kernel_size", kernel_size, self.name, allow_five=True, ret_five=True)
self.kernel_size = _check_3d_int_or_tuple("kernel_size", kernel_size, self.name,
allow_five=True, ret_five=True)
self.add_prim_attr("kernel_size", self.kernel_size)
self.strides = _check_3d_int_or_tuple("strides", strides, self.name, allow_five=True, ret_five=True)
self.add_prim_attr("strides", self.strides)
@ -1248,12 +1252,10 @@ class LSTMGradData(PrimitiveWithInfer):
validator.check_int(dhy_shape[0], self.num_layers * self.num_directions, Rel.EQ, "h_shape[0]", self.name)
validator.check_equal_int(dhy_shape[2], self.hidden_size, "h_shape[2]", self.name)
# dy: (seq_len, batch_size, hidden_size * num_directions)
validator.check_equal_int(len(dy_shape), 3, "dy_shape", self.name)
validator.check_equal_int(dy_shape[1], dhy_shape[1], "dy[1]", self.name)
validator.check_int(dy_shape[2], self.hidden_size * self.num_directions, Rel.EQ, "dy[2]", self.name)
# (seq_len, batch_size, input_size)
dx_shape = (y_shape[0], y_shape[1], self.input_size)
dhx_shape = dhy_shape
dcx_shape = dcy_shape
@ -1332,12 +1334,10 @@ class LSTMGrad(PrimitiveWithInfer):
validator.check_int(dhy_shape[0], self.num_layers * self.num_directions, Rel.EQ, "h_shape[0]", self.name)
validator.check_equal_int(dhy_shape[2], self.hidden_size, "h_shape[2]", self.name)
# dy: (seq_len, batch_size, hidden_size * num_directions)
validator.check_equal_int(len(dy_shape), 3, "dy_shape", self.name)
validator.check_equal_int(dy_shape[1], dhy_shape[1], "dy[1]", self.name)
validator.check_int(dy_shape[2], self.hidden_size * self.num_directions, Rel.EQ, "dy[2]", self.name)
# (seq_len, batch_size, input_size)
dx_shape = (y_shape[0], y_shape[1], self.input_size)
dhx_shape = dhy_shape
dcx_shape = dcy_shape
@ -1561,16 +1561,16 @@ class PReLUGrad(PrimitiveWithInfer):
def __init__(self):
pass
def infer_shape(self, y_backprop_shape, A_shape, w_shape):
if len(A_shape) == 1:
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):
def infer_dtype(self, y_backprop_dtype, a_dtype, w_dtype):
tuple(map(partial(validator.check_tensor_dtype_valid,
valid_dtypes=(mstype.float16, mstype.float32), prim_name=self.name),
('y_backprop', "input_x", "weight"),
(y_backprop_dtype, A_dtype, w_dtype)))
(y_backprop_dtype, a_dtype, w_dtype)))
return y_backprop_dtype, w_dtype
@ -1746,8 +1746,8 @@ class SigmoidGrad(PrimitiveWithInfer):
return out
class HSigmoidGrad(PrimitiveWithInfer):
"""Gets the gradient of HSigmoid operation."""
class _ActivationGrad(PrimitiveWithInfer):
"""_ActivationGrad base class."""
@prim_attr_register
def __init__(self):
@ -1763,21 +1763,12 @@ class HSigmoidGrad(PrimitiveWithInfer):
return x_dtype
class HSwishGrad(PrimitiveWithInfer):
class HSwishGrad(_ActivationGrad):
"""Gets the gradient of HSwish operation."""
@prim_attr_register
def __init__(self):
self.init_prim_io_names(inputs=['y_grad', 'x'], outputs=['output'])
def infer_shape(self, y_grad_shape, x_shape):
return x_shape
def infer_dtype(self, y_grad_dtype, x_dtype):
valid_dtypes = (mstype.float16, mstype.float32)
validator.check_tensor_dtype_valid("y_grad", y_grad_dtype, valid_dtypes, self.name)
validator.check_tensor_dtype_valid("x", x_dtype, valid_dtypes, self.name)
return x_dtype
class HSigmoidGrad(_ActivationGrad):
"""Gets the gradient of HSigmoid operation."""
class SigmoidCrossEntropyWithLogitsGrad(PrimitiveWithInfer):

6
mindspore/ops/operations/_inner_ops.py
View File

@ -65,9 +65,9 @@ class ExtractImagePatches(PrimitiveWithInfer):
raise ValueError(f"For \'{prim_name}\' the format of {arg_name}s should be [1, {arg_name}_row, "
f"{arg_name}_col, 1], but got {arg_val}.")
if not isinstance(arg_val[2], int) or not isinstance(arg_val[3], int) or arg_val[2] < 1 or arg_val[3] < 1:
raise ValueError(f"For '{prim_name}' the {arg_name}_row and {arg_name}_col in {arg_name}s should be an "
f"positive integer number, but got {arg_name}_row is {arg_val[2]}, {arg_name}_col "
f"is {arg_val[3]}")
raise ValueError(f"For '{prim_name}' the {arg_name}_row and {arg_name}_col in {arg_name}s should be "
f"an positive integer number, but got {arg_name}_row is {arg_val[2]}, "
f"{arg_name}_col is {arg_val[3]}")
_check_tuple_or_list("ksize", ksizes, self.name)
_check_tuple_or_list("stride", strides, self.name)

6
mindspore/ops/operations/_quant_ops.py
View File

@ -1279,7 +1279,8 @@ class BatchNormFold2(PrimitiveWithInfer):
validator.check("batch_std shape", batch_std_shape, "running_std shape", running_std_shape, Rel.EQ, self.name)
validator.check("batch_std shape", batch_std_shape, "batch_mean shape", batch_mean_shape, Rel.EQ, self.name)
validator.check("batch_std shape", batch_std_shape, "beta shape", beta_shape, Rel.EQ, self.name)
validator.check("batch_std shape", batch_std_shape, "running_mean shape", running_mean_shape, Rel.EQ, self.name)
validator.check("batch_std shape", batch_std_shape, "running_mean shape", running_mean_shape,
Rel.EQ, self.name)
validator.check("batch_std shape", batch_std_shape, "batch_mean shape", gamma_shape, Rel.EQ, self.name)
validator.check("batch_std_shape[0]", batch_std_shape[0], "x_shape channel size", x_shape[self.channel_axis],
Rel.EQ, self.name)
@ -1327,7 +1328,8 @@ class BatchNormFold2Grad(PrimitiveWithInfer):
running_std_shape, running_mean_shape, global_step_shape):
validator.check("batch_std shape", batch_std_shape, "batch_mean shape", batch_mean_shape, Rel.EQ, self.name)
validator.check("batch_std shape", batch_std_shape, "running_std shape", running_std_shape, Rel.EQ, self.name)
validator.check("batch_std shape", batch_std_shape, "running_mean shape", running_mean_shape, Rel.EQ, self.name)
validator.check("batch_std shape", batch_std_shape, "running_mean shape", running_mean_shape,
Rel.EQ, self.name)
validator.check("batch_std shape", batch_std_shape, "gamma shape", gamma_shape, Rel.EQ, self.name)
validator.check("batch_std size", batch_std_shape[0], "dout channel size", dout_shape[self.channel_axis],
Rel.EQ, self.name)

131
mindspore/ops/operations/_thor_ops.py
View File

@ -1,4 +1,4 @@
# Copyright 2020 Huawei Technologies Co., Ltd
# Copyright 2020-2021 Huawei Technologies Co., Ltd
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
@ -18,6 +18,7 @@ import math
from ..primitive import prim_attr_register, PrimitiveWithInfer
from ...common import dtype as mstype
from ..._checkparam import Validator as validator
from ..operations.nn_ops import _check_positive_int_or_tuple
from ..._checkparam import Rel
__all__ = ["CusBatchMatMul",
@ -34,37 +35,6 @@ __all__ = ["CusBatchMatMul",
]
def _check_positive_int_or_tuple(arg_name, arg_value, prim_name, allow_four=False, ret_four=False):
"""
Checks whether an argument is a positive int or tuple with 2 or 4(when allow_four is True) positive int elements.
"""
def _raise_message():
raise ValueError(f"For '{prim_name}' attr '{arg_name}' should be an positive int number or a tuple of two "
f"{'or four ' if allow_four else ''}positive int numbers, but got {arg_value}")
def _get_return_value():
if isinstance(arg_value, int):
ret = (1, 1, arg_value, arg_value) if ret_four else (arg_value, arg_value)
elif len(arg_value) == 2:
ret = (1, 1, arg_value[0], arg_value[1]) if ret_four else arg_value
elif len(arg_value) == 4:
if not allow_four:
_raise_message()
ret = arg_value if ret_four else (arg_value[2], arg_value[3])
else:
_raise_message()
return ret
validator.check_value_type(arg_name, arg_value, (int, tuple), prim_name)
ret_value = _get_return_value()
for item in ret_value:
if isinstance(item, int) and item > 0:
continue
_raise_message()
return ret_value
class CusBatchMatMul(PrimitiveWithInfer):
"""
Multiplies matrix `a` by matrix `b` in batch.
@ -205,7 +175,6 @@ class CusImg2Col(PrimitiveWithInfer):
bs, c, h, w = data1_shape
_, stride_h, stride_w, _ = self.strides
_, k_w, k_h, _ = self.ksizes
# assert m == n
c0 = 16
c1 = c // 16
if c1 == 0:
@ -483,7 +452,7 @@ class CusMatMulCubeFraczLeftCast(PrimitiveWithInfer):
class Im2Col(PrimitiveWithInfer):
"""
extracts image pathes from image.
extracts image paths from image.
The rank of input_x1 must be `4`, data_format is "NCHW".
@ -574,7 +543,8 @@ class UpdateThorGradient(PrimitiveWithInfer):
- **input_x1** (Tensor) - The first input is the diag part of the cov matrix of feature map.
Supported dtype [float32].
- **input_x2** (Tensor) - The second input is the corresponding 1st-order grad. Supported dtype [float32].
- **input_x3** (Tensor) - The third input is the diag part of the cov matrix of dout. Supported dtype [float32].
- **input_x3** (Tensor) - The third input is the diag part of the cov matrix of dout.
Supported dtype [float32].
Outputs:
Tensor, the shape is the same as the shape of input_x2, it will be used to update the weights.
@ -608,71 +578,50 @@ class UpdateThorGradient(PrimitiveWithInfer):
return x2_dtype
class Cholesky(PrimitiveWithInfer):
class _Cholesky(PrimitiveWithInfer):
"""
Inner API for _Cholesky base class.
"""
@prim_attr_register
def __init__(self, split_dim=0):
self.init_prim_io_names(inputs=['x1'], outputs=['y'])
self.split_dim = split_dim
self.add_prim_attr('split_dim', self.split_dim)
def infer_shape(self, x1_shape):
if self.split_dim != 0:
assert len(x1_shape) == 2
height = x1_shape[0]
width = x1_shape[1]
assert height == width
if height <= self.split_dim:
out_shape = [1, height, width]
else:
batch = height // self.split_dim
if height != batch * self.split_dim:
batch += 1
out_shape = [batch, self.split_dim, self.split_dim]
else:
out_shape = x1_shape
return out_shape
def infer_dtype(self, x1_dtype):
validator.check_tensor_dtype_valid('x1', x1_dtype, [mstype.float32], self.name)
return x1_dtype
class Cholesky(_Cholesky):
"""
Inner API for positive-definite matrix Cholesky decomposition GPU backend.
"""
@prim_attr_register
def __init__(self, split_dim=0):
self.init_prim_io_names(inputs=['x1'], outputs=['y'])
self.split_dim = split_dim
self.add_prim_attr('split_dim', self.split_dim)
def infer_shape(self, x1_shape):
if self.split_dim != 0:
assert len(x1_shape) == 2
height = x1_shape[0]
width = x1_shape[1]
assert height == width
if height <= self.split_dim:
out_shape = [1, height, width]
else:
batch = height // self.split_dim
if height != batch * self.split_dim:
batch += 1
out_shape = [batch, self.split_dim, self.split_dim]
else:
out_shape = x1_shape
return out_shape
def infer_dtype(self, x1_dtype):
validator.check_tensor_dtype_valid('x1', x1_dtype, [mstype.float32], self.name)
return x1_dtype
class CholeskyTrsm(PrimitiveWithInfer):
class CholeskyTrsm(_Cholesky):
"""
Inner API for resnet50 THOR GPU backend.
"""
@prim_attr_register
def __init__(self, split_dim=0):
self.init_prim_io_names(inputs=['x1'], outputs=['y'])
self.split_dim = split_dim
self.add_prim_attr('split_dim', self.split_dim)
def infer_shape(self, x1_shape):
if self.split_dim != 0:
assert len(x1_shape) == 2
height = x1_shape[0]
width = x1_shape[1]
assert height == width
if height <= self.split_dim:
out_shape = [1, height, width]
else:
batch = height // self.split_dim
if height != batch * self.split_dim:
batch += 1
out_shape = [batch, self.split_dim, self.split_dim]
else:
out_shape = x1_shape
return out_shape
def infer_dtype(self, x1_dtype):
validator.check_tensor_dtype_valid('x1', x1_dtype, [mstype.float32], self.name)
return x1_dtype
class DetTriangle(PrimitiveWithInfer):
"""

37
mindspore/ops/operations/array_ops.py
View File

@ -2362,16 +2362,16 @@ def _get_stack_shape(x_shape, x_type, axis, prim_name):
validator.check_int(len(x_shape), 1, Rel.GE, "len of input_x", prim_name)
validator.check_subclass("input_x[0]", x_type[0], mstype.tensor, prim_name)
rank_base = len(x_shape[0])
N = len(x_shape)
n = len(x_shape)
out_shape = x_shape[0]
validator.check_int_range(axis, -rank_base - 1, rank_base, Rel.INC_BOTH, 'axis', prim_name)
if axis < 0:
axis = axis + rank_base + 1
for i in range(1, N):
for i in range(1, n):
validator.check('x_type[%d]' % i, x_type[i], 'base', x_type[0], Rel.EQ, prim_name, TypeError)
if x_shape[i] != x_shape[0]:
raise ValueError(f"For \'{prim_name}\' element {i} shape in input can not pack with first element")
out_shape.insert(axis, N)
out_shape.insert(axis, n)
return out_shape
@ -2406,7 +2406,8 @@ class Stack(PrimitiveWithInfer):
Stacks the list of input tensors with the same rank `R`, output is a tensor of rank `(R+1)`.
Given input tensors of shape :math:`(x_1, x_2, ..., x_R)`. Set the number of input tensors as `N`.
If :math:`0 \le axis`, the shape of the output tensor is :math:`(x_1, x_2, ..., x_{axis}, N, x_{axis+1}, ..., x_R)`.
If :math:`0 \le axis`, the shape of the output tensor is
:math:`(x_1, x_2, ..., x_{axis}, N, x_{axis+1}, ..., x_R)`.
Args:
axis (int): Dimension to stack. Default: 0.
@ -2848,11 +2849,9 @@ def _compute_slicing_length(begin, end, stride, x_shape, i):
if 0 <= end < x_dim:
end += -x_dim
if end < -x_dim - 1:
# When slicing backward, if end < -x_dim - 1, set end = -x_dim - 1, which means
# slicing to the 0th element.
# Slicing to the 0th element.
end = -x_dim - 1
if begin <= end:
# When slicing backward, if begin <= end, the length of the slicing is 0.
slicing_length = 0
else:
slicing_length = 1 + (end + 1 - begin) // stride
@ -3250,7 +3249,8 @@ class ScatterNd(PrimitiveWithInfer):
r"""
Scatters a tensor into a new tensor depending on the specified indices.
Creates an empty tensor with the given `shape`, and set values by scattering the update tensor depending on indices.
Creates an empty tensor with the given `shape`, and set values by scattering the update tensor
depending on indices.
The empty tensor has rank P and `indices` has rank Q where `Q >= 2`.
@ -3642,7 +3642,8 @@ class ScatterMax(_ScatterOp):
``Ascend`` ``CPU``
Examples:
>>> input_x = Parameter(Tensor(np.array([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]]), mindspore.float32), name="input_x")
>>> input_x = Parameter(Tensor(np.array([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]]), mindspore.float32),
... name="input_x")
>>> indices = Tensor(np.array([[0, 0], [1, 1]]), mindspore.int32)
>>> updates = Tensor(np.ones([2, 2, 3]) * 88, mindspore.float32)
>>> scatter_max = ops.ScatterMax()
@ -3691,7 +3692,8 @@ class ScatterMin(_ScatterOp):
``Ascend`` ``CPU``
Examples:
>>> input_x = Parameter(Tensor(np.array([[0.0, 1.0, 2.0], [0.0, 0.0, 0.0]]), mindspore.float32), name="input_x")
>>> input_x = Parameter(Tensor(np.array([[0.0, 1.0, 2.0], [0.0, 0.0, 0.0]]), mindspore.float32),
... name="input_x")
>>> indices = Tensor(np.array([[0, 0], [1, 1]]), mindspore.int32)
>>> update = Tensor(np.ones([2, 2, 3]), mindspore.float32)
>>> scatter_min = ops.ScatterMin()
@ -4142,7 +4144,8 @@ class DepthToSpace(PrimitiveWithInfer):
- **x** (Tensor) - The target tensor. It must be a 4-D tensor with shape :math:`(N, C_{in}, H_{in}, W_{in})`.
Outputs:
Tensor of shape :math:`(N, C_{in} / \text{block_size}, H_{in} * \text{block_size}, W_{in} * \text{block_size})`.
Tensor of shape :math:`(N, C_{in} / \text{block_size}, H_{in} * \text{block_size},
W_{in} * \text{block_size})`.
Raises:
TypeError: If `block_size` is not an int.
@ -4275,8 +4278,8 @@ class BatchToSpace(PrimitiveWithInfer):
Divides batch dimension with blocks and interleaves these blocks back into spatial dimensions.
This operation will divide batch dimension N into blocks with block_size, the output tensor's N dimension
is the corresponding number of blocks after division. The output tensor's H, W dimension is product of original H, W
dimension and block_size with given amount to crop from dimension, respectively.
is the corresponding number of blocks after division. The output tensor's H, W dimension is product of
original H, W dimension and block_size with given amount to crop from dimension, respectively.
Args:
block_size (int): The block size of division, has the value not less than 2.
@ -4466,8 +4469,8 @@ class BatchToSpaceND(PrimitiveWithInfer):
Divides batch dimension with blocks and interleaves these blocks back into spatial dimensions.
This operation will divide batch dimension N into blocks with block_shape, the output tensor's N dimension
is the corresponding number of blocks after division. The output tensor's H, W dimension is product of original H, W
dimension and block_shape with given amount to crop from dimension, respectively.
is the corresponding number of blocks after division. The output tensor's H, W dimension is product of
original H, W dimension and block_shape with given amount to crop from dimension, respectively.
Args:
block_shape (Union[list(int), tuple(int), int]): The block shape of dividing block with all value greater
@ -4580,8 +4583,8 @@ class BroadcastTo(Primitive):
where it will be substituted by the input tensor's shape in that position, see example.
Inputs:
- **input_x** (Tensor) - The input tensor. The data type should be one of the following types: float16, float32,
int32, int8, uint8.
- **input_x** (Tensor) - The input tensor. The data type should be one of the following types:
float16, float32, int32, int8, uint8.
Outputs:
Tensor, with the given `shape` and the same data type as `input_x`.

7
mindspore/ops/operations/comm_ops.py
View File

@ -23,6 +23,7 @@ from ...common import dtype as mstype
from ..primitive import PrimitiveWithInfer, PrimitiveWithCheck, prim_attr_register
from ...common.api import context
class ReduceOp:
"""
Operation options for reducing tensors.
@ -45,12 +46,14 @@ class ReduceOp:
target_dtypes = (mstype.int8, mstype.int32, mstype.float16, mstype.float32)
def check_hcom_group_valid(group):
if context.get_context("mode") == context.PYNATIVE_MODE and \
context.get_context("device_target") == "Ascend" and \
group != GlobalComm.WORLD_COMM_GROUP:
raise RuntimeError("Only hccl_world_group is supported in Pynative mode, but got {}".format(group))
class AllReduce(PrimitiveWithInfer):
"""
Reduces the tensor data across all devices in such a way that all devices will get the same final result.
@ -149,7 +152,7 @@ class AllGather(PrimitiveWithInfer):
``Ascend`` ``GPU``
Examples:
>>> # This example should be run with two devices. Refer to the tutorial > Distributed Training on mindspore.cn.
>>> # This example should be run with two devices. Refer to the tutorial > Distributed Training on mindspore.cn
>>> import numpy as np
>>> import mindspore.ops.operations as ops
>>> import mindspore.nn as nn
@ -304,7 +307,7 @@ class ReduceScatter(PrimitiveWithInfer):
``Ascend`` ``GPU``
Examples:
>>> # This example should be run with two devices. Refer to the tutorial > Distributed Training on mindspore.cn.
>>> # This example should be run with two devices. Refer to the tutorial > Distributed Training on mindspore.cn
>>> from mindspore import Tensor, context
>>> from mindspore.communication import init
>>> from mindspore.ops.operations.comm_ops import ReduceOp

3
mindspore/ops/operations/math_ops.py
View File

@ -1844,7 +1844,8 @@ class Log1p(Primitive):
Returns the natural logarithm of one plus the input tensor element-wise.
Inputs:
- **input_x** (Tensor) - The input tensor. With float16 or float32 data type. The value must be greater than -1.
- **input_x** (Tensor) - The input tensor. With float16 or float32 data type.
The value must be greater than -1.
Outputs:
Tensor, has the same shape as the `input_x`.

70
mindspore/ops/operations/nn_ops.py
View File

@ -1019,8 +1019,8 @@ class BNTrainingUpdate(PrimitiveWithInfer):
Inputs:
- **x** (Tensor) - A 4-D Tensor with float16 or float32 data type. Tensor of shape :math:`(N, C, A, B)`.
- **sum** (Tensor) - A 1-D Tensor with float16 or float32 data type for the output of operator BNTrainingReduce.
Tensor of shape :math:`(C,)`.
- **sum** (Tensor) - A 1-D Tensor with float16 or float32 data type for the output of operator
BNTrainingReduce. Tensor of shape :math:`(C,)`.
- **square_sum** (Tensor) - A 1-D Tensor with float16 or float32 data type for the output of operator
BNTrainingReduce. Tensor of shape :math:`(C,)`.
- **scale** (Tensor) - A 1-D Tensor with float16 or float32, for the scaling factor.
@ -2690,7 +2690,8 @@ class SGD(PrimitiveWithCheck):
>>> stat = Tensor(np.array([1.5, -0.3, 0.2, -0.7]), mindspore.float32)
>>> output = sgd(parameters, gradient, learning_rate, accum, momentum, stat)
>>> print(output)
(Tensor(shape=[4], dtype=Float32, value= [ 1.98989999e+00, -4.90300000e-01, 1.69520009e+00, 3.98009992e+00]),)
(Tensor(shape=[4], dtype=Float32,
value= [ 1.98989999e+00, -4.90300000e-01, 1.69520009e+00, 3.98009992e+00]),)
"""
@prim_attr_register
@ -3234,14 +3235,15 @@ class OneHot(Primitive):
If the input indices is rank `N`, the output will have rank `N+1`. The new axis is created at dimension `axis`.
Args:
axis (int): Position to insert the value. e.g. If `indices` shape is [n, c], and `axis` is `-1` the output shape
will be [n, c, depth], If `axis` is `0` the output shape will be [depth, n, c]. Default: -1.
axis (int): Position to insert the value. e.g. If `indices` shape is [n, c], and `axis` is `-1` the output
shape will be [n, c, depth], If `axis` is `0` the output shape will be [depth, n, c]. Default: -1.
Inputs:
- **indices** (Tensor) - A tensor of indices. Tensor of shape :math:`(X_0, \ldots, X_n)`.
Data type must be int32 or int64.
- **depth** (int) - A scalar defining the depth of the one hot dimension.
- **on_value** (Tensor) - A value to fill in output when `indices[j] = i`. With data type of float16 or float32.
- **on_value** (Tensor) - A value to fill in output when `indices[j] = i`.
With data type of float16 or float32.
- **off_value** (Tensor) - A value to fill in output when `indices[j] != i`.
Has the same data type with as `on_value`.
@ -3622,7 +3624,6 @@ class LSTM(PrimitiveWithInfer):
self.num_directions = 1
def infer_shape(self, x_shape, h_shape, c_shape, w_shape):
# (seq, batch_size, feature)
validator.check_equal_int(len(x_shape), 3, "x rank", self.name)
validator.check_equal_int(x_shape[2], self.input_size, "x[2]", self.name)
@ -3630,7 +3631,6 @@ class LSTM(PrimitiveWithInfer):
validator.check_equal_int(len(h_shape), 3, "h rank", self.name)
validator.check("h_shape", h_shape, "c_shape", c_shape, Rel.EQ, self.name)
# (num_layers * num_directions, batch, hidden_size)
validator.check_int(h_shape[0], self.num_layers * self.num_directions, Rel.EQ, "h[0]", self.name)
validator.check_equal_int(h_shape[1], x_shape[1], "h[1]", self.name)
validator.check_int(h_shape[2], self.hidden_size, Rel.EQ, "h[2]", self.name)
@ -4238,11 +4238,10 @@ class AdamNoUpdateParam(PrimitiveWithInfer):
- **beta2_power** (Tensor) - :math:`beta_2^t(\beta_1^{t})` in the updating formula.
The data type must be float32.
- **lr** (Tensor) - :math:`l` in the updating formula. The data type must be float32.
The paper suggested value is :math:`10^{-8}`
- **beta1** (Tensor) - The exponential decay rate for the 1st moment estimations. The data type must be float32.
The paper suggested value is :math:`0.9`
- **beta2** (Tensor) - The exponential decay rate for the 2nd moment estimations. The data type must be float32.
The paper suggested value is :math:`0.999`
- **beta1** (Tensor) - The exponential decay rate for the 1st moment estimations.
The data type must be float32.
- **beta2** (Tensor) - The exponential decay rate for the 2nd moment estimations.
The data type must be float32.
- **epsilon** (Tensor) - Term added to the denominator to improve numerical stability. The data type must be
float32.
- **gradient** (Tensor) - Gradient, the shape must be the same as `m`, the data type must be float32.
@ -4253,8 +4252,8 @@ class AdamNoUpdateParam(PrimitiveWithInfer):
Raises:
TypeError: If neither `use_locking` nor `use_nesterov` is a bool.
TypeError: If `m`, `v`, `beta1_power`, `beta2_power1`, `lr`, `beta1`, `beta2`, `epsilon` or `gradient` is not a
Tensor.
TypeError: If `m`, `v`, `beta1_power`, `beta2_power1`, `lr`, `beta1`, `beta2`, `epsilon` or `gradient`
is not a Tensor.
Supported Platforms:
``CPU``
@ -4351,8 +4350,8 @@ class FusedSparseAdam(PrimitiveWithInfer):
- **var** (Parameter) - Parameters to be updated with float32 data type.
- **m** (Parameter) - The 1st moment vector in the updating formula, has the same type as `var` with
float32 data type.
- **v** (Parameter) - The 2nd moment vector in the updating formula. Mean square gradients, has the same type as
`var` with float32 data type.
- **v** (Parameter) - The 2nd moment vector in the updating formula.
Mean square gradients, has the same type as `var` with float32 data type.
- **beta1_power** (Tensor) - :math:`beta_1^t` in the updating formula with float32 data type.
- **beta2_power** (Tensor) - :math:`beta_2^t` in the updating formula with float32 data type.
- **lr** (Tensor) - :math:`l` in the updating formula. With float32 data type.
@ -4498,8 +4497,8 @@ class FusedSparseLazyAdam(PrimitiveWithInfer):
- **var** (Parameter) - Parameters to be updated with float32 data type.
- **m** (Parameter) - The 1st moment vector in the updating formula, has the same type as `var` with
float32 data type.
- **v** (Parameter) - The 2nd moment vector in the updating formula. Mean square gradients, has the same type as
`var` with float32 data type.
- **v** (Parameter) - The 2nd moment vector in the updating formula.
Mean square gradients, has the same type as `var` with float32 data type.
- **beta1_power** (Tensor) - :math:`beta_1^t` in the updating formula with float32 data type.
- **beta2_power** (Tensor) - :math:`beta_2^t` in the updating formula with float32 data type.
- **lr** (Tensor) - :math:`l` in the updating formula with float32 data type.
@ -4995,7 +4994,8 @@ class BinaryCrossEntropy(PrimitiveWithInfer):
valid_dtypes = (mstype.float16, mstype.float32)
validator.check_tensors_dtypes_same_and_valid(args, valid_dtypes, self.name)
if weight_type:
validator.check_tensors_dtypes_same_and_valid({'x': x_type, 'weight': weight_type}, valid_dtypes, self.name)
validator.check_tensors_dtypes_same_and_valid({'x': x_type, 'weight': weight_type}, valid_dtypes,
self.name)
return x_type
@ -6953,8 +6953,8 @@ class CTCLoss(Primitive):
(`max_time`, `batch_size`, `num_classes`). `num_classes` must be `num_labels + 1` classes, `num_labels`
indicates the number of actual labels. Blank labels are reserved. Default blank label is `num_classes - 1`.
Data type must be float16, float32 or float64.
- **labels_indices** (Tensor) - The indices of labels. `labels_indices[i, :] == [b, t]` means `labels_values[i]`
stores the id for `(batch b, time t)`. The type must be int64 and rank must be 2.
- **labels_indices** (Tensor) - The indices of labels. `labels_indices[i, :] == [b, t]` means
`labels_values[i]` stores the id for `(batch b, time t)`. The type must be int64 and rank must be 2.
- **labels_values** (Tensor) - A `1-D` input tensor. The values are associated with the given batch and time.
The type must be int32. `labels_values[i]` must in the range of `[0, num_classes)`.
- **sequence_length** (Tensor) - A tensor containing sequence lengths with the shape of (`batch_size`).
@ -6966,8 +6966,8 @@ class CTCLoss(Primitive):
- **gradient** (Tensor) - The gradient of `loss`, has the same type and shape with `inputs`.
Raises:
TypeError: If `preprocess_collapse_repeated`, `ctc_merge_repeated` or `ignore_longer_outputs_than_inputs` is not
a bool.
TypeError: If `preprocess_collapse_repeated`, `ctc_merge_repeated` or `ignore_longer_outputs_than_inputs`
is not a bool.
TypeError: If `inputs`, `labels_indices`, `labels_values` or `sequence_length` is not a Tensor.
TypeError: If dtype of `inputs` is not one of the following: float16, float32 or float64.
TypeError: If dtype of `labels_indices` is not int64.
@ -7088,8 +7088,6 @@ class BasicLSTMCell(PrimitiveWithInfer):
Please use DynamicRNN instead.
"""
# deprecate_new_name = "BasicLSTMCell"
@prim_attr_register
def __init__(self, keep_prob=1.0, forget_bias=1.0, state_is_tuple=True, activation='tanh'):
self.keep_prob = validator.check_value_type("keep_prob", keep_prob, [float], self.name)
@ -7331,8 +7329,8 @@ class DynamicGRUV2(PrimitiveWithInfer):
The data type must be float16.
- **bias_input** (Tensor) - Input-hidden bias. Tensor of shape :math:`(3 \times \text{hidden_size})`, or None.
Has the same data type with input `init_h`.
- **bias_hidden** (Tensor) - Hidden-hidden bias. Tensor of shape :math:`(3 \times \text{hidden_size})`, or None.
Has the same data type with input `init_h`.
- **bias_hidden** (Tensor) - Hidden-hidden bias. Tensor of shape :math:`(3 \times \text{hidden_size})`,
or None. Has the same data type with input `init_h`.
- **seq_length** (Tensor) - The length of each batch. Tensor of shape :math:`(\text{batch_size})`.
Only `None` is currently supported.
- **init_h** (Tensor) - Hidden state of initial time.
@ -7475,7 +7473,8 @@ class InTopK(PrimitiveWithInfer):
k (int): Specifies the number of top elements to be used for computing precision.
Inputs:
- **x1** (Tensor) - A 2D Tensor defines the predictions of a batch of samples with float16 or float32 data type.
- **x1** (Tensor) - A 2D Tensor defines the predictions of a batch of samples with float16 or float32
data type.
- **x2** (Tensor) - A 1D Tensor defines the labels of a batch of samples with int32 data type. The size of x2
must be equal to x1's first dimension. The values of `x2` can not be negative and
must be equal to or less than index of x1's second dimension.
@ -7712,10 +7711,10 @@ class Conv3D(PrimitiveWithInfer):
Args:
out_channels (int): The number of output channel :math:`C_{out}`.
kernel_size (Union[int, tuple[int]]): The data type is int or a tuple of 3 integers. Specifies the depth, height
and width of the 3D convolution window. Single int means the value is for the depth, height and the width
of the kernel. A tuple of 3 ints means the first value is for the depth, height and the other is for the
width of the kernel.
kernel_size (Union[int, tuple[int]]): The data type is int or a tuple of 3 integers. Specifies the depth,
height and width of the 3D convolution window. Single int means the value is for the depth, height
and the width of the kernel. A tuple of 3 ints means the first value is for the depth, height and
the other is for the width of the kernel.
mode (int): Modes for different convolutions. Not currently used.
stride (Union[int, tuple[int]]): The distance of kernel moving, an int number that represents
the depth, height and width of movement are both strides, or a tuple of three int numbers that
@ -7755,7 +7754,8 @@ class Conv3D(PrimitiveWithInfer):
- **input** (Tensor) - Tensor of shape :math:`(N, C_{in}, D_{in}, H_{in}, W_{in})`.
Currently input data type only support float16 and float32.
- **weight** (Tensor) - Set size of kernel is :math:`(k_d, K_h, K_w)`, then the shape is
:math:`(C_{out}, C_{in}//groups, k_d, K_h, K_w)`. Currently weight data type only support float16 and float32.
:math:`(C_{out}, C_{in}//groups, k_d, K_h, K_w)`.
Currently weight data type only support float16 and float32.
- **bias** (Tensor) - Tensor of shape :math:`C_{in}`. Currently, only support none.
Outputs:
@ -8185,7 +8185,7 @@ class Conv3DTranspose(PrimitiveWithInfer):
if isinstance(pad, int):
pad = (pad,) * 6
if len(pad) != 6:
raise ValueError(f"For `conv3d` attr 'pad' should be an positive int number or a tuple of "
raise ValueError(f"For `Conv3DTranspose` attr 'pad' should be an positive int number or a tuple of "
f"six positive int numbers, but got `{len(pad)}`.")
self.pad_list = pad
validator.check_value_type('pad_mode', pad_mode, [str], self.name)

4
mindspore/ops/operations/other_ops.py
View File

@ -119,6 +119,7 @@ class InplaceAssign(PrimitiveWithInfer):
def infer_dtype(self, x, y, z):
return z
class Load(PrimitiveWithCheck):
"""
Load `Parameter` to a value.
@ -142,6 +143,7 @@ class Load(PrimitiveWithCheck):
if variable != mstype.type_refkey:
validator.check_tensor_type_same({"variable": variable}, mstype.number_type, self.name)
class BoundingBoxEncode(PrimitiveWithInfer):
"""
Encodes bounding boxes locations.
@ -476,6 +478,7 @@ class Depend(Primitive):
def __call__(self, value, expr):
return value
class UpdateState(Primitive):
"""
UpdateState is used for update side-effect state.
@ -495,6 +498,7 @@ class UpdateState(Primitive):
def __call__(self, state, expr):
return state
class CheckBprop(PrimitiveWithInfer):
"""
Checks whether the data type and the shape of corresponding elements from tuples x and y are the same.

7
mindspore/ops/primitive.py
View File

@ -249,8 +249,8 @@ class Primitive(Primitive_):
class PrimitiveWithCheck(Primitive):
"""
PrimitiveWithCheck is the base class of primitives in python defines functions for checking operator input arguments
but used the infer method registered in c++ source codes.
PrimitiveWithCheck is the base class of primitives in python defines functions for checking operator
input arguments but used the infer method registered in c++ source codes.
There are three methods can be override to define the check logic of the primitive: __check__(), check_shape(),
check_dtype(). If __check__() is defined in primitive, the __check__() has highest priority to be called.
@ -330,7 +330,8 @@ class PrimitiveWithCheck(Primitive):
class PrimitiveWithInfer(Primitive):
"""
PrimitiveWithInfer is the base class of primitives in python and defines functions for tracking inference in python.
PrimitiveWithInfer is the base class of primitives in python and defines functions for tracking inference
in python.
There are four method can be override to define the infer logic of the primitive: __infer__(), infer_shape(),
infer_dtype(), and infer_value(). If __infer__() is defined in primitive, the __infer__() has highest priority