# Copyright 2020 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.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ============================================================================
"""Operators for nn."""
import math
import operator
from functools import reduce
import numpy as np
from ... import context
from ..._checkparam import ParamValidator as validator
from ..._checkparam import Rel, check_bool, check_int_positive
from ...common import dtype as mstype
from ..primitive import Primitive, PrimitiveWithInfer, prim_attr_register
[docs]class Flatten(PrimitiveWithInfer):
r"""
Flattens a tensor without changing its batch size on the 0-th axis.
Inputs:
- **input_x** (Tensor) - Tensor of shape :math:`(N, \ldots)` to be flattened.
Outputs:
Tensor, the shape of the output tensor is :math:`(N, X)`, where :math:`X` is
the product of the remaining dimension.
Examples:
>>> input_tensor = Tensor(np.ones(shape=[1, 2, 3, 4]), mindspore.float32)
>>> flatten = Flatten()
>>> output = flatten(input_tensor)
>>> assert output.shape() == (1, 24)
"""
@prim_attr_register
def __init__(self):
pass
def infer_shape(self, input_x):
validator.check('input_x rank', len(input_x), '', 1, Rel.GE)
prod = 1 if len(input_x) == 1 else reduce(operator.mul, input_x[1:])
return input_x[0], prod
def infer_dtype(self, input_x):
validator.check_subclass("input_x", input_x, mstype.tensor)
return input_x
[docs]class Softmax(PrimitiveWithInfer):
r"""
Softmax operation.
Applies the Softmax operation to the input tensor on the specified axis.
Suppose a slice along the given aixs :math:`x` then for each element :math:`x_i`
the Softmax function is shown as follows:
.. math::
\text{output}(x_i) = \frac{exp(x_i)}{\sum_{j = 0}^{N-1}\exp(x_j)},
where :math:`N` is the length of the tensor.
Args:
axis (Union[int, tuple]): The axis to do the Softmax operation. Default: -1.
Inputs:
- **logits** (Tensor) - The input of Softmax.
Outputs:
Tensor, with the same type and shape as the logits.
"""
@prim_attr_register
def __init__(self, axis=-1):
self.init_prim_io_names(inputs=['x'], outputs=['output'])
validator.check_type("axis", axis, [int, tuple])
if isinstance(axis, int):
self.add_prim_attr('axis', (axis,))
for item in self.axis:
validator.check_type("item of axis", item, [int])
def infer_shape(self, logits):
validator.check_shape_length("axis shape", len(self.axis), 1, Rel.GE)
rank = len(logits)
for axis_v in self.axis:
validator.check_int_range("axis", axis_v, -rank, rank, Rel.INC_LEFT)
return logits
def infer_dtype(self, logits):
validator.check_subclass("logits", logits, mstype.tensor)
return logits
[docs]class LogSoftmax(PrimitiveWithInfer):
r"""
Log Softmax activation function.
Applies the Log Softmax function to the input tensor on the specified axis.
Suppose a slice along the given aixs :math:`x` then for each element :math:`x_i`
the Log Softmax function is shown as follows:
.. math::
\text{output}(x_i) = \log \left(\frac{exp(x_i)} {\sum_{j = 0}^{N-1}\exp(x_j)}\right),
where :math:`N` is the length of the Tensor.
Args:
axis (int): The axis to do the Log softmax operation. Default: -1.
Inputs:
- **logits** (Tensor) - The input of Log Softmax.
Outputs:
Tensor, with the same type and shape as the logits.
"""
@prim_attr_register
def __init__(self, axis=-1):
validator.check_type("axis", axis, [int])
def infer_shape(self, logits):
rank = len(logits)
validator.check_int_range('axis', self.axis, -rank - 1, rank, Rel.INC_BOTH)
return logits
def infer_dtype(self, logits):
validator.check_subclass("logits", logits, mstype.tensor)
return logits
[docs]class ReLU(PrimitiveWithInfer):
r"""
Computes ReLU(Rectified Linear Unit) of input tensor element-wise.
It returns :math:`\max(x,\ 0)` element-wise.
Inputs:
- **input_x** (Tensor) - The input tensor.
Outputs:
Tensor, with the same type and shape as the `input_x`.
Examples:
>>> input_x = Tensor(np.array([[-1.0, 4.0, -8.0], [2.0, -5.0, 9.0]], np.float32))
>>> relu = ReLU()
>>> result = relu(input_x)
[[0, 4.0, 0.0], [2.0, 0.0, 9.0]]
"""
@prim_attr_register
def __init__(self):
"""init ReLU"""
self.init_prim_io_names(inputs=['x'], outputs=['output'])
def infer_shape(self, input_x):
return input_x
def infer_dtype(self, input_x):
validator.check_subclass("input_x", input_x, mstype.tensor)
validator.check_typename("input_x", input_x, mstype.number_type)
return input_x
[docs]class ReLU6(PrimitiveWithInfer):
r"""
Computes ReLU(Rectified Linear Unit) upper bounded by 6 of input tensor element-wise.
It returns :math:`\min(\max(0,x), 6)` element-wise.
Inputs:
- **input_x** (Tensor) - The input tensor.
Outputs:
Tensor, with the same type and shape as the `input_x`.
Examples:
>>> input_x = Tensor(np.array([[-1.0, 4.0, -8.0], [2.0, -5.0, 9.0]], np.float32))
>>> relu6 = ReLU6()
>>> result = relu6(input_x)
>>> assert result.asnumpy() == Tensor(np.array([[0, 4.0, 0.0], [2.0, 0.0, 6.0]], np.float32)).asnumpy()
"""
@prim_attr_register
def __init__(self):
"""init ReLU6"""
self.init_prim_io_names(inputs=['x'], outputs=['output'])
def infer_shape(self, input_x):
return input_x
def infer_dtype(self, input_x):
validator.check_subclass("input_x", input_x, mstype.tensor)
validator.check_typename("input_x", input_x, (mstype.float16, mstype.float32))
return input_x
[docs]class Sigmoid(PrimitiveWithInfer):
r"""
Sigmoid activation function.
Computes Sigmoid of input element-wise. The Sigmoid function is defined as:
.. math::
\text{sigmoid}(x_i) = \frac{1}{1 + exp(-x_i)},
where :math:`x_i` is the element of the input.
Inputs:
- **input_x** (Tensor) - The input of Sigmoid.
Outputs:
Tensor, with the same type and shape as the input_x.
"""
@prim_attr_register
def __init__(self):
self.init_prim_io_names(inputs=['x'], outputs=['output'])
def infer_shape(self, input_x):
return input_x
def infer_dtype(self, input_x):
validator.check_subclass("input_x", input_x, mstype.tensor)
validator.check_typename("input_x", input_x, (mstype.float16, mstype.float32))
return input_x
[docs]class Tanh(PrimitiveWithInfer):
r"""
Tanh activation function.
Computes hyperbolic tangent of input element-wise. The Tanh function is defined as:
.. math::
tanh(x_i) = \frac{\exp(x_i) - \exp(-x_i)}{\exp(x_i) + \exp(-x_i)} = \frac{\exp(2x_i) - 1}{\exp(2x_i) + 1},
where :math:`x_i` is an element of the input Tensor.
Inputs:
- **input_x** (Tensor) - The input of Tanh.
Outputs:
Tensor, with the same type and shape as the input_x.
"""
@prim_attr_register
def __init__(self):
pass
def infer_shape(self, input_x):
return input_x
def infer_dtype(self, input_x):
validator.check_subclass("input_x", input_x, mstype.tensor)
return input_x
[docs]class FusedBatchNorm(Primitive):
r"""
FusedBatchNorm is a BatchNorm that moving mean and moving variance will be computed instead of being loaded.
Batch Normalization is widely used in convolutional networks. This operation applies
Batch Normalization over input to avoid internal covariate shift as described in the
paper `Batch Normalization: Accelerating Deep Network Training by Reducing Internal
Covariate Shift <https://arxiv.org/abs/1502.03167>`_. It rescales and recenters the
feature using a mini-batch of data and the learned parameters which can be described
in the following formula.
.. math::
y = \frac{x - mean}{\sqrt{variance + \epsilon}} * \gamma + \beta
where :math:`\gamma` is scale, :math:`\beta` is bias, :math:`\epsilon` is epsilon.
Args:
mode (int): Mode of batch normalization, value is 0 or 1. Default: 0.
epsilon (float): A small value added for numerical stability. Default: 1e-5.
momentum (float): The hyper parameter to compute moving average for running_mean and running_var
(e.g. :math:`new\_running\_mean = momentum * running\_mean + (1 - momentum) * current\_mean`).
Momentum value should be [0, 1]. Default: 0.9.
Inputs:
- **input_x** (Tensor) - Tensor of shape :math:`(N, C)`.
- **scale** (Tensor) - Tensor of shape :math:`(C,)`.
- **bias** (Tensor) - Tensor of shape :math:`(C,)`.
- **mean** (Tensor) - Tensor of shape :math:`(C,)`.
- **variance** (Tensor) - Tensor of shape :math:`(C,)`.
Outputs:
Tuple of 5 Tensor, the normalized input and the updated parameters.
- **output_x** (Tensor) - The same type and shape as the `input_x`.
- **updated_scale** (Tensor) - Tensor of shape :math:`(C,)`.
- **updated_bias** (Tensor) - Tensor of shape :math:`(C,)`.
- **updated_moving_mean** (Tensor) - Tensor of shape :math:`(C,)`.
- **updated_moving_variance** (Tensor) - Tensor of shape :math:`(C,)`.
"""
@prim_attr_register
def __init__(self, mode=0, epsilon=1e-5, momentum=0.1):
self.init_prim_io_names(inputs=['x', 'scale', 'b', 'mean', 'variance'],
outputs=['y', 'running_mean', 'running_variance', 'save_mean', 'save_inv_variance'])
self.mode = validator.check_integer('mode', mode, [0, 1], Rel.IN)
self.epsilon = validator.check_number_range('epsilon', epsilon, 0, 1, Rel.INC_RIGHT)
self.momentum = validator.check_number_range('momentum', momentum, 0, 1, Rel.INC_BOTH)
[docs]class BatchNorm(PrimitiveWithInfer):
r"""
Batch Normalization for input data and updated parameters.
Batch Normalization is widely used in convolutional neural networks. This operation
applies Batch Normalization over input to avoid internal covariate shift as described
in the paper `Batch Normalization: Accelerating Deep Network Training by Reducing Internal
Covariate Shift <https://arxiv.org/abs/1502.03167>`_. It rescales and recenters the
features using a mini-batch of data and the learned parameters which can be described
in the following formula,
.. math::
y = \frac{x - mean}{\sqrt{variance + \epsilon}} * \gamma + \beta
where :math:`\gamma` is scale, :math:`\beta` is bias, :math:`\epsilon` is epsilon.
Args:
is_training (bool): If `is_training` is True, `mean` and `variance` are computed during training.
If `is_training` is False, they're loaded from checkpoint during inference. Default: False.
epsilon (float): A small value added for numerical stability. Default: 1e-5.
Inputs:
- **input_x** (Tensor) - Tensor of shape :math:`(N, C)`.
- **scale** (Tensor) - Tensor of shape :math:`(C,)`.
- **bias** (Tensor) - Tensor of shape :math:`(C,)`.
- **mean** (Tensor) - Tensor of shape :math:`(C,)`.
- **variance** (Tensor) - Tensor of shape :math:`(C,)`.
Outputs:
Tuple of 5 Tensor, the normalized inputs and the updated parameters.
- **output_x** (Tensor) - The same type and shape as the input_x. The shape is :math:`(N, C)`.
- **updated_scale** (Tensor) - Tensor of shape :math:`(C,)`.
- **updated_bias** (Tensor) - Tensor of shape :math:`(C,)`.
- **reserve_space_1** (Tensor) - Tensor of shape :math:`(C,)`.
- **reserve_space_2** (Tensor) - Tensor of shape :math:`(C,)`.
- **reserve_space_3** (Tensor) - Tensor of shape :math:`(C,)`.
"""
@prim_attr_register
def __init__(self, is_training=False, epsilon=1e-5):
self.is_training = validator.check_type('is_training', is_training, (bool,))
self.epsilon = validator.check_number_range('epsilon', epsilon, 0, 1, Rel.INC_RIGHT)
self.add_prim_attr('data_format', "NCHW")
self.init_prim_io_names(inputs=['x', 'scale', 'offset', 'mean', 'variance'],
outputs=['y', 'batch_mean', 'batch_variance', 'reserve_space_1', 'reserve_space_2',
'reserve_space_3'])
def infer_shape(self, input_x, scale, bias, mean, variance):
validator.check("BatchNorm scale shape length", len(scale), "1", 1, Rel.EQ)
validator.check("BatchNorm scale shape", scale, "BatchNorm bias shape", bias)
validator.check("BatchNorm scale shape", scale[0], "BatchNorm input_x shape[1]", input_x[1])
if not self.is_training:
validator.check("BatchNorm mean shape length", len(mean), "1", 1, Rel.EQ)
validator.check("BatchNorm mean shape", mean, "BatchNorm variance shape", variance)
validator.check("BatchNorm mean shape", mean, "BatchNorm scale shape", scale)
return (input_x, scale, scale, scale, scale, scale)
def infer_dtype(self, input_x, scale, bias, mean, variance):
args = {"BatchNorm scale type": scale, "BatchNorm bias type": bias}
args_moving = {"BatchNorm mean type": mean, "BatchNorm variance type": variance}
validator.check_typename("input_x", input_x, [mstype.float32, mstype.float16])
validator.check_type_same(args, [mstype.float32, mstype.float16])
if self.is_training:
validator.check_type_same(args_moving, [mstype.float32, mstype.float16, None])
else:
validator.check_type_same(args_moving, [mstype.float32, mstype.float16])
return (input_x, scale, bias, input_x, input_x, input_x)
[docs]class Conv2D(PrimitiveWithInfer):
r"""
2D convolution layer.
Applies a 2D convolution over an input tensor which is typically of shape :math:`(N, C_{in}, H_{in}, W_{in})`,
where :math:`N` is batch size and :math:`C_{in}` is channel number. For each batch of shape
:math:`(C_{in}, H_{in}, W_{in})`, the formula is defined as:
.. math::
out_j = \sum_{i=0}^{C_{in} - 1} ccor(W_{ij}, X_i) + b_j,
where :math:`ccor` is cross correlation operator, :math:`C_{in}` is the input channel number, :math:`j` ranges
from :math:`0` to :math:`C_{out} - 1`, :math:`W_{ij}` corresponds to :math:`i`-th channel of the :math:`j`-th
filter and :math:`out_{j}` corresponds to the :math:`j`-th channel of the output. :math:`W_{ij}` is a slice
of kernel and it has shape :math:`(\text{ks_h}, \text{ks_w})`, where :math:`\text{ks_h}` and
:math:`\text{ks_w}` are height and width of the convolution kernel. The full kernel has shape
:math:`(C_{out}, C_{in} // \text{group}, \text{ks_h}, \text{ks_w})`, where group is the group number
to split the input in the channel dimension.
If the 'pad_mode' is set to be "valid", the output height and width will be
:math:`\left \lfloor{1 + \frac{H_{in} + 2 \times \text{padding} - \text{ks_h} -
(\text{ks_h} - 1) \times (\text{dilation} - 1) }{\text{stride}}} \right \rfloor` and
:math:`\left \lfloor{1 + \frac{W_{in} + 2 \times \text{padding} - \text{ks_w} -
(\text{ks_w} - 1) \times (\text{dilation} - 1) }{\text{stride}}} \right \rfloor` respectively.
The first introduction can be found in paper `Gradient Based Learning Applied to Document Recognition
<http://vision.stanford.edu/cs598_spring07/papers/Lecun98.pdf>`_. More detailed introduction can be found here:
http://cs231n.github.io/convolutional-networks/.
Args:
out_channel (int): The dimension of the output.
kernel_size (Union[int, tuple[int]]): The kernel size of the 2D convolution.
mode (int): 0 Math convolutiuon, 1 cross-correlation convolution ,
2 deconvolution, 3 depthwise convolution. Default: 1.
pad_mode (str): "valid", "same", "pad" the mode to fill padding. Default: "valid".
pad (int): The pad value to fill. Default: 0.
stride (int): The stride to apply conv filter. Default: 1.
dilation (int): Specify the space to use between kernel elements. Default: 1.
group (int): Split input into groups. Default: 1.
Returns:
Tensor, the value that applied 2D convolution.
Inputs:
- **input** (Tensor) - Tensor of shape :math:`(N, C_{in}, H_{in}, W_{in})`.
- **weight** (Tensor) - Set size of kernel is :math:`(K_1, K_2)`, then the shape is
:math:`(C_{out}, C_{in}, K_1, K_2)`.
Outputs:
Tensor of shape :math:`(N, C_{out}, H_{out}, W_{out})`.
"""
@prim_attr_register
def __init__(self,
out_channel,
kernel_size,
mode=1,
pad_mode="valid",
pad=0,
stride=1,
dilation=1,
group=1):
"""init Conv2D"""
self.init_prim_io_names(inputs=['x', 'w'], outputs=['output'])
self.kernel_size = kernel_size
self.kernel_size = validator.check_type('kernel_size', kernel_size, (int, tuple))
if isinstance(self.kernel_size, int):
self.kernel_size = (self.kernel_size, self.kernel_size)
validator.check_integer('length of kernel_size', len(self.kernel_size), 2, Rel.GE)
validator.equal('type of pad', type(pad), 'not bool', not isinstance(pad, bool))
validator.equal('type of pad', type(pad), 'int', isinstance(pad, int))
self.pad_mode = validator.check_string('pad_mode', pad_mode, ['valid', 'same', 'pad'])
self.pad = validator.check_pad_value_by_mode(self.__class__.__name__, pad_mode, pad)
if self.pad_mode == 'pad':
validator.check_integer('pad', self.pad, 0, Rel.GE)
self.mode = validator.check_integer('mode', mode, 1, Rel.EQ)
self.add_prim_attr('data_format', "NCHW")
self.out_channel = validator.check_integer('out_channel', out_channel, 0, Rel.GT)
self.group = validator.check_integer('group', group, 0, Rel.GT)
self.dilation = validator.check_integer('dilation', dilation, 1, Rel.GE)
validator.check_type('kernel_size', kernel_size, [int, tuple])
if isinstance(kernel_size, int) and kernel_size < 1:
raise ValueError('Attr \'kernel_size\' of \'Conv2D\' Op passed '
+ str(self.kernel_size) + ', should be a int or tuple and equal to or greater than 1.')
if isinstance(kernel_size, tuple) and (len(kernel_size) != 2 or
(not isinstance(kernel_size[0], int)) or
(not isinstance(kernel_size[1], int)) or
kernel_size[0] < 1 or kernel_size[1] < 1):
raise ValueError('Attr \'kernel_size\' of \'Conv2D\' Op passed '
+ str(self.kernel_size) + ', should be a int or tuple and equal to or greater than 1.')
self.stride = validator.check_integer('stride', stride, 1, Rel.GE)
def infer_shape(self, x_shape, w_shape):
validator.check_integer("weight_shape", len(w_shape), 4, Rel.EQ)
validator.check_integer("x_shape", len(x_shape), 4, Rel.EQ)
validator.check_param_equal("x_shape[1]", x_shape[1] // self.group, "w_shape[1]", w_shape[1])
validator.check_param_equal('out_channel', self.out_channel, 'w_shape[0]', w_shape[0])
validator.check_param_equal('kernel_size', self.kernel_size, 'w_shape[2:4]', tuple(w_shape[2:4]))
kernel_size_h = w_shape[2]
kernel_size_w = w_shape[3]
if self.pad_mode == "valid":
h_out = math.ceil((x_shape[2] - self.dilation * (kernel_size_h - 1)) / self.stride)
w_out = math.ceil((x_shape[3] - self.dilation * (kernel_size_w - 1)) / self.stride)
pad_top, pad_bottom, pad_left, pad_right = 0, 0, 0, 0
elif self.pad_mode == "same":
h_out = math.ceil(x_shape[2] / self.stride)
w_out = math.ceil(x_shape[3] / self.stride)
pad_needed_h = max(0, (h_out - 1) * self.stride + self.dilation * (kernel_size_h - 1) + 1 - x_shape[2])
pad_top = math.floor(pad_needed_h / 2)
pad_bottom = pad_needed_h - pad_top
pad_needed_w = max(0, (w_out - 1) * self.stride + self.dilation * (kernel_size_w - 1) + 1 - x_shape[3])
pad_left = math.floor(pad_needed_w / 2)
pad_right = pad_needed_w - pad_left
elif self.pad_mode == 'pad':
pad_top, pad_bottom, pad_left, pad_right = self.pad, self.pad, self.pad, self.pad
h_out = 1 + (x_shape[2] + 2 * self.pad - kernel_size_h - (kernel_size_h - 1) * (self.dilation - 1)) \
/ self.stride
w_out = 1 + (x_shape[3] + 2 * self.pad - kernel_size_w - (kernel_size_w - 1) * (self.dilation - 1)) \
/ self.stride
h_out = math.floor(h_out)
w_out = math.floor(w_out)
self.pad_list = [pad_top, pad_bottom, pad_left, pad_right]
self.add_prim_attr('pad_list', (pad_top, pad_bottom, pad_left, pad_right))
out_channel = self.out_channel
out_shape = [x_shape[0], out_channel, h_out, w_out]
return out_shape
def infer_dtype(self, x_dtype, w_dtype):
args = {'x_dtype': x_dtype, 'w_dtype': w_dtype}
validator.check_subclass('input', x_dtype, mstype.tensor)
validator.check_subclass('weight', w_dtype, mstype.tensor)
validator.check_type_same(args, [mstype.int8, mstype.int32, mstype.float16, mstype.float32])
return x_dtype
[docs]class DepthwiseConv2dNative(PrimitiveWithInfer):
r"""
Returns the depth-wise convolution value for the input.
Applies depthwise conv2d for the input, which will generate more channels with channel_multiplier.
Given an input tensor of shape :math:`(N, C_{in}, H_{in}, W_{in})` where :math:`N` is the batch size and a
filter tensor with kernel size :math:`(ks_{h}, ks_{w})`, containing :math:`C_{in} * \text{channel_multiplier}`
convolutional filters of depth 1; it applies different filters to each input channel (channel_multiplier channels
for each with default value 1), then concatenates the results together. The output has
:math:`\text{in_channels} * \text{channel_multiplier}` channels.
Args:
channel_multiplier (int): The multipiler for the original output conv.
kernel_size (int or tuple): The size of the conv kernel.
mode (int): 0 Math convolution, 1 cross-correlation convolution ,
2 deconvolution, 3 depthwise convolution. Default: 3.
pad_mode (str): "valid", "same", "pad" the mode to fill padding. Default: "valid".
pad (int): The pad value to fill. Default: 0.
stride (int): The stride to apply conv filter. Default: 1.
dilation (int): Specifies the dilation rate to use for dilated convolution. Default: 1.
group (int): Splits input into groups. Default: 1.
Inputs:
- **input** (Tensor) - Tensor of shape :math:`(N, C_{in}, H_{in}, W_{in})`.
- **weight** (Tensor) - Set size of kernel is :math:`(K_1, K_2)`, then the shape is
:math:`(C_{out}, C_{in}, K_1, K_2)`.
Outputs:
Tensor of shape :math:`(N, C_{in} * \text{channel_multiplier}, H_{out}, W_{out})`.
"""
@prim_attr_register
def __init__(self,
channel_multiplier,
kernel_size,
mode=3,
pad_mode="valid",
pad=0,
stride=1,
dilation=1,
group=1):
"""init DepthwiseConv2dNative"""
validator.check_pad_value_by_mode(self.__class__.__name__, pad_mode, pad)
validator.check_type("kernel_size", kernel_size, (int, tuple))
if isinstance(kernel_size, int):
kernel_size = (kernel_size, kernel_size)
if isinstance(kernel_size, tuple) and (len(kernel_size) != 2 or
(not isinstance(kernel_size[0], int)) or
(not isinstance(kernel_size[1], int)) or
kernel_size[0] < 1 or kernel_size[1] < 1):
raise ValueError(f"Attr kernel_size of DepthwiseConv2dNative Op not passed "
f"{kernel_size}, should be a int or tuple and equal to or greater than 1.")
if pad_mode not in ("same", "valid", "pad"):
raise ValueError(f"Attr pad_mode of DepthwiseConv2dNative Op not passed"
f"{pad_mode} not in valid, same, pad.")
self.pad_mode = pad_mode
self.mode = validator.check_integer("mode", mode, 3, Rel.EQ)
self.add_prim_attr('data_format', "NCHW")
self.channel_multiplier = validator.check_integer("channel_multiplier", channel_multiplier, 0, Rel.GT)
self.group = validator.check_integer("group", group, 0, Rel.GT)
self.dilation = validator.check_integer("dilation", dilation, 1, Rel.GE)
self.kernel_size = validator.check_value_on_integer("kernel_size", kernel_size, 1, Rel.GE)
self.stride = validator.check_integer("stride", stride, 1, Rel.GE)
self.pad = pad
def infer_shape(self, x_shape, w_shape):
validator.check_integer("weight_shape", len(w_shape), 4, Rel.EQ)
validator.check_integer("x_shape", len(x_shape), 4, Rel.EQ)
validator.check_param_equal("x_shape[1]", x_shape[1], "w_shape[1]", w_shape[1])
validator.check_param_equal('kernel_size', self.kernel_size, 'w_shape[2:4]', tuple(w_shape[2:4]))
kernel_size_h = w_shape[2]
kernel_size_w = w_shape[3]
if self.pad_mode == "valid":
h_out = math.ceil((x_shape[2] - self.dilation * (kernel_size_h - 1)) / self.stride)
w_out = math.ceil((x_shape[3] - self.dilation * (kernel_size_w - 1)) / self.stride)
pad_top, pad_bottom, pad_left, pad_right = 0, 0, 0, 0
elif self.pad_mode == "same":
h_out = math.ceil(x_shape[2] / self.stride)
w_out = math.ceil(x_shape[3] / self.stride)
pad_needed_h = max(0, (h_out - 1) * self.stride + self.dilation * (kernel_size_h - 1) + 1 - x_shape[2])
pad_top = math.floor(pad_needed_h / 2)
pad_bottom = pad_needed_h - pad_top
pad_needed_w = max(0, (w_out - 1) * self.stride + self.dilation * (kernel_size_w - 1) + 1 - x_shape[3])
pad_left = math.floor(pad_needed_w / 2)
pad_right = pad_needed_w - pad_left
elif self.pad_mode == 'pad':
pad_top, pad_bottom, pad_left, pad_right = self.pad, self.pad, self.pad, self.pad
h_out = 1 + (x_shape[2] + 2 * self.pad - kernel_size_h - (kernel_size_h - 1) * (self.dilation - 1)) \
/ self.stride
w_out = 1 + (x_shape[3] + 2 * self.pad - kernel_size_w - (kernel_size_w - 1) * (self.dilation - 1)) \
/ self.stride
h_out = math.floor(h_out)
w_out = math.floor(w_out)
else:
raise ValueError(f"Attr pad_mode of DepthwiseConv2dNative Op not passed"
"{pad_mode} not in valid, same, pad.")
self.pad_list = (pad_top, pad_bottom, pad_left, pad_right)
self.add_prim_attr('pads', self.pad_list)
out_channel = self.channel_multiplier * x_shape[1]
out_shape = [x_shape[0], out_channel, h_out, w_out]
return out_shape
def infer_dtype(self, x_dtype, w_dtype):
args = {'x_dtype': x_dtype, 'w_dtype': w_dtype}
validator.check_type_same(args, mstype.number_type)
return x_dtype
[docs]class MaxPoolWithArgmax(PrimitiveWithInfer):
r"""
Performs max pooling on the input Tensor and return both max values and indices.
Typically the input is of shape :math:`(N_{in}, C_{in}, H_{in}, W_{in})`, MaxPool outputs
regional maximum in the :math:`(H_{in}, W_{in})`-dimension. Given kernel size
:math:`ks = (h_{ker}, w_{ker})` and stride :math:`s = (s_0, s_1)`, the operation is as follows.
.. math::
\text{output}(N_i, C_j, h, w) = \max_{m=0, \ldots, h_{ker}-1} \max_{n=0, \ldots, w_{ker}-1}
\text{input}(N_i, C_j, s_0 \times h + m, s_1 \times w + n)
Args:
pad_mode (str): "valid", "same", "pad" the mode to fill padding. Default: "valid".
window (Union[int, tuple[int]]): The size of window, which is the kernel size, two `int` for width
and height. Default: 1.
pad (Union[int, tuple[int]]): If `pad_mode` is `pad`, the pad value to fill, two `int` for width
and height. Default: 0.
stride (Union[int, tuple[int]]): The stride of the window, that should be a tuple of two `int` for
width and height. Default: 1.
Inputs:
- **input** (Tensor) - Tensor of shape :math:`(N, C_{in}, H_{in}, W_{in})`.
Outputs:
Tuple of 2 Tensor, the maxpool result and where max values from.
- **output** (Tensor) - Maxpooling result, with shape :math:`(N, C_{out}, H_{out}, W_{out})`.
- **mask** (Tensor) - Max values' index represented by the mask.
"""
@prim_attr_register
def __init__(self,
pad_mode="valid",
window=1,
pad=0,
stride=1,
data_mode=1,
ceil_mode=0,
alpha=1.0,
beta=0.0):
self.init_prim_io_names(inputs=['x'], outputs=['output', 'argmax'])
self.window = validator.check_type('window', window, [int, tuple])
if isinstance(window, int) and window <= 0:
raise ValueError('Attr \'window\' of \'MaxPoolWithArgmax\' Op passed '
+ str(self.window)+', should be a int or tuple and greater than 0.')
if isinstance(window, tuple) and (len(window) != 2 or
(not isinstance(window[0], int)) or
(not isinstance(window[1], int)) or
window[0] <= 0 or window[1] <= 0):
raise ValueError('Attr \'window\' of \'MaxPoolWithArgmax\' Op passed '
+ str(self.window)+', should be a int or tuple and greater than 0.')
self.pool_h = self.pool_w = window
self.pad_mode = validator.check_string('pad_mode', pad_mode, ['valid', 'same', 'pad'])
if self.pad_mode == "valid":
self.pad = 0
elif self.pad_mode == "same":
self.pad = math.floor((self.window - 1) / 2)
elif self.pad_mode == "pad":
self.pad = validator.check_integer('pad', pad, 0, Rel.GE)
self.data_mode = validator.check_integer('data_mode', data_mode, 1, Rel.EQ)
self.ceil_mode = validator.check_integer('ceil_mode', ceil_mode, 0, Rel.EQ)
self.stride = validator.check_integer('stride', stride, 1, Rel.GE)
self.alpha = validator.check_type('alpha', alpha, [int, float])
self.beta = validator.check_type('beta', beta, [int, float])
self.is_tbe = not context.get_context("enable_ge") and context.get_context("device_target") == "Ascend"
def infer_shape(self, x_shape):
validator.check_integer("x_shape", len(x_shape), 4, Rel.EQ)
pad = self.pad
h_input = x_shape[2]
w_input = x_shape[3]
h_out = (h_input + 2 * pad - (self.window - 1) - 1) / self.stride + 1
h_out = math.floor(h_out)
w_out = (w_input + 2 * pad - (self.window - 1) - 1) / self.stride + 1
w_out = math.floor(w_out)
out_shape = [x_shape[0], x_shape[1], h_out, w_out]
for shape_value in out_shape:
if shape_value <= 0:
raise ValueError("The kernel size is not valid please check it if is larger than data's shape size.")
k_size_vec = [1, self.window, self.window, 1]
argmax_shape = []
if self.is_tbe:
for i in range(4):
if i == 2:
dim = k_size_vec[i - 1] * k_size_vec[i]
argmax_shape.append(dim)
elif i == 3:
dim = math.ceil(out_shape[i - 1] * out_shape[i] / 16) + 1
argmax_shape.append(dim)
else:
argmax_shape.append(x_shape[i])
else:
argmax_shape = out_shape
return out_shape, argmax_shape
def infer_dtype(self, x_dtype):
out_dtype = x_dtype
validator.check_typename("x_type", x_dtype, (mstype.float16, mstype.float32))
argmax_dtype = mstype.int32
return out_dtype, argmax_dtype
class _Pool(PrimitiveWithInfer):
r"""
Performs max/avg pooling operation.
Args:
ksize (Union[int, tuple[int]]): The size of the window to take a max over, that should be a tuple
of two `int` for width and height. Default: 1.
stride (Union[int, tuple[int]]): The stride of the window, that should be a tuple of two `int` for
width and height. Default: 1.
padding (str): The optional values for pad mode "SAME", "VALID". Default: "VALID".
"""
@prim_attr_register
def __init__(self, ksize=1, strides=1, padding="VALID"):
self.init_prim_io_names(inputs=['x'], outputs=['output'])
validator.check_type('padding', padding, [str])
self.ksize = ksize
self.strides = strides
self.padding = padding.upper()
self.ksize = validator.check_type('ksize', self.ksize, [int, tuple])
self.strides = validator.check_type('strides', self.strides, [int, tuple])
self.padding = validator.check_string('padding', self.padding, ['VALID', 'SAME'])
self.init_prim_io_names(inputs=['x'], outputs=['output'])
self.add_prim_attr("padding", self.padding)
self.add_prim_attr('data_format', "NCHW")
if isinstance(self.ksize, int):
self.pool_h = validator.check_integer("ksize", self.ksize, 1, Rel.GE)
self.pool_w = self.pool_h
self.add_prim_attr("ksize", (1, 1, self.ksize, self.ksize))
elif isinstance(self.ksize, tuple):
if (len(self.ksize) != 2 or (not isinstance(self.ksize[0], int)) or (not isinstance(self.ksize[1], int))
or self.ksize[0] <= 0 or self.ksize[1] <= 0):
raise ValueError('Each value of attr \'ksize\' of \'MaxPool\' Op passed ' +
str(self.ksize) + ', should be a int or a tuple of length 2 and greater than 0.')
self.pool_h = self.ksize[0]
self.pool_w = self.ksize[1]
self.add_prim_attr("ksize", (1, 1, self.ksize[0], self.ksize[1]))
if isinstance(self.strides, int):
self.stride_h = validator.check_integer("strides", self.strides, 1, Rel.GE)
self.stride_w = self.stride_h
self.add_prim_attr("strides", (1, 1, self.strides, self.strides))
elif isinstance(self.strides, tuple):
if (len(self.strides) != 2 or (not isinstance(self.strides[0], int)) or
(not isinstance(self.strides[1], int)) or self.strides[0] <= 0 or self.strides[1] <= 0):
raise ValueError('Each value of attr \'strides\' of \'MaxPool\' Op passed ' +
str(self.strides) + ', should be a int or a tuple of length 2 and greater than 0.')
self.stride_h = self.strides[0]
self.stride_w = self.strides[1]
self.add_prim_attr("strides", (1, 1, self.strides[0], self.strides[1]))
if self.padding == "VALID":
self.pad = 0
elif self.padding == "SAME":
self.pad = math.floor((self.pool_h - 1) / 2)
else:
raise ValueError('The padding should be str and must be SAME or VALID,'
' but got {}.'.format(self.padding))
self.add_prim_attr('pad', self.pad)
def infer_shape(self, x_shape):
validator.check_integer("x_shape", len(x_shape), 4, Rel.EQ)
h_input = x_shape[2]
w_input = x_shape[3]
if self.padding == "VALID":
h_out = math.ceil((h_input - (self.pool_h - 1)) / self.stride_h)
w_out = math.ceil((w_input - (self.pool_w - 1)) / self.stride_w)
elif self.padding == "SAME":
h_out = math.ceil(h_input / self.stride_h)
w_out = math.ceil(w_input / self.stride_w)
else:
raise ValueError('The padding should be str and must be SAME or VALID,'
' but got {}.'.format(self.padding))
out_shape = [x_shape[0], x_shape[1], h_out, w_out]
for shape_value in out_shape:
if shape_value <= 0:
raise ValueError("The kernel size is not valid please check it if is larger than data's shape size.")
return out_shape
def infer_dtype(self, x_dtype):
validator.check_subclass("input", x_dtype, mstype.tensor)
return x_dtype
[docs]class MaxPool(_Pool):
r"""
Max pooling operation.
Applies a 2D max pooling over an input Tensor which can be regarded as a composition of 2D planes.
Typically the input is of shape :math:`(N_{in}, C_{in}, H_{in}, W_{in})`, MaxPool outputs
regional maximum in the :math:`(H_{in}, W_{in})`-dimension. Given kernel size
:math:`ks = (h_{ker}, w_{ker})` and stride :math:`s = (s_0, s_1)`, the operation is as follows.
.. math::
\text{output}(N_i, C_j, h, w) = \max_{m=0, \ldots, h_{ker}-1} \max_{n=0, \ldots, w_{ker}-1}
\text{input}(N_i, C_j, s_0 \times h + m, s_1 \times w + n)
Args:
ksize (Union[int, tuple[int]]): The size of the window to take a max over, that should be a tuple
of two `int` for width and height. Default: 1.
stride (Union[int, tuple[int]]): The stride of the window, that should be a tuple of two `int` for
width and height. Default: 1.
padding (str): The optional values for pad mode "SAME", "VALID". Default: "VALID".
Inputs:
- **input** (Tensor) - Tensor of shape :math:`(N, C_{in}, H_{in}, W_{in})`.
Outputs:
Tensor, with shape :math:`(N, C_{out}, H_{out}, W_{out})`.
"""
@prim_attr_register
def __init__(self, ksize=1, strides=1, padding="VALID"):
super(MaxPool, self).__init__(ksize, strides, padding)
[docs]class AvgPool(_Pool):
r"""
Average pooling operation.
Applies a 2D average pooling over an input Tensor which can be regarded as a composition of 2D input planes.
Typically the input is of shape :math:`(N_{in}, C_{in}, H_{in}, W_{in})`, AvgPool2d outputs
regional average in the :math:`(H_{in}, W_{in})`-dimension. Given kernel size
:math:`ks = (h_{ker}, w_{ker})` and stride :math:`s = (s_0, s_1)`, the operation is as follows.
.. math::
\text{output}(N_i, C_j, h, w) = \frac{1}{h_{ker} * w_{ker}} \sum_{m=0}^{h_{ker}-1} \sum_{n=0}^{w_{ker}-1}
\text{input}(N_i, C_j, s_0 \times h + m, s_1 \times w + n)
Args:
ksize (Union[int, tuple[int]]): The size of the window to take a average over, that should be a tuple
of two `int` for width and height. Default: 1.
stride (Union[int, tuple[int]]): The stride of the window, that should be a tuple of two `int` for
width and height. Default: 1.
padding (str): The optional values for pad mode "SAME", "VALID". Default: "VALID".
Inputs:
- **input** (Tensor) - Tensor of shape :math:`(N, C_{in}, H_{in}, W_{in})`.
Outputs:
Tensor, with shape :math:`(N, C_{out}, H_{out}, W_{out})`.
"""
@prim_attr_register
def __init__(self, ksize=1, strides=1, padding="VALID"):
if context.get_context("device_target") == "GPU":
self.target = "GPU"
else:
self.target = "OTHER"
super(AvgPool, self).__init__(ksize, strides, padding)
[docs]class BiasAdd(PrimitiveWithInfer):
r"""
Returns sum of input and bias tensor.
Adds the 1-D bias tensor to the input tensor, and boardcasts the shape on all axis
except for the channel axis.
Inputs:
- **input_x** (Tensor) - Input value, with shape :math:`(N, C)` or :math:`(N, C, H, W)`.
- **bias** (Tensor) - Bias value, with shape :math:`(C)`.
Outputs:
Tensor, with the same shape and type as `input_x`.
"""
@prim_attr_register
def __init__(self):
self.init_prim_io_names(inputs=['x', 'b'], outputs=['output'])
self.add_prim_attr('data_format', 'NCHW')
def infer_shape(self, x_shape, b_shape):
if len(b_shape) != 1 or len(x_shape) < 2 or b_shape[0] != x_shape[1]:
raise ValueError("Input_x and bias shapes do not match",
"(require: rank of input_x must be at least 2, rank of bias must be 1, "
"input_x.dim[1] must equal bias.dim[0]),"
" but got input_x shape {}, bias shape {}.".format(x_shape, b_shape))
return x_shape
def infer_dtype(self, x_type, b_type):
args = {"input_x type": x_type, "bias type": b_type}
validator.check_type_same(args, (mstype.float16, mstype.float32, mstype.int8, mstype.int32))
return x_type
[docs]class TopK(PrimitiveWithInfer):
"""
Finds values and indices of the `k` largest entries along the last dimension.
Args:
sorted (bool): If true, the resulting elements will
be sorted by the values in descending order. Default: False.
Inputs:
- **input_x** (Tensor) - Input to be computed.
- **k** (int) - Number of top elements to be computed along the last dimension, constant input is needed.
Outputs:
Tuple of 2 Tensor, the values and the indices.
- **values** (Tensor) - The `k` largest elements along each last dimensional slice.
- **indices** (Tensor) - The indices of values within the last dimension of input.
Examples:
>>> topk = TopK(sorted=True)
>>> x = Tensor(np.array([1, 2, 3, 4, 5]).astype(np.float16))
>>> values, indices = topk(x)
>>> assert values == Tensor(np.array([5, 4, 3]))
>>> assert indices == Tensor(np.array([4, 3, 2]))
"""
@prim_attr_register
def __init__(self, sorted=False):
validator.check_type("sorted", sorted, [bool])
self.init_prim_io_names(inputs=['input', 'k'],
outputs=['values', 'indices'])
def __infer__(self, input_x, k):
x_shape = list(input_x['shape'])
ndim = len(x_shape) - 1
k_v = k['value']
x_shape[ndim] = k_v
input_dtype = input_x['dtype']
validator.check_typename("TopK input_dtype",
input_dtype, (mstype.float16, mstype.float32, mstype.int32))
if not isinstance(k_v, int):
raise ValueError('The k must int.', k)
return {'shape': (x_shape, x_shape),
'dtype': (input_dtype, mstype.int32),
'value': None}
[docs]class SoftmaxCrossEntropyWithLogits(PrimitiveWithInfer):
r"""
Gets the softmax cross-entropy value between logits and labels which shoule be one-hot encoding.
Note:
Sets input logits as `X`, input label as `Y`, output as `loss`. Then,
.. math::
p_{ij} = softmax(X_{ij}) = \frac{exp(x_i)}{\sum_{j = 0}^{N-1}\exp(x_j)}
.. math::
loss_{ij} = -\sum_j{Y_{ij} * ln(p_{ij})}
Inputs:
- **logits** (Tensor) - Input logits, with shape :math:`(N, C)`.
- **labels** (Tensor) - Ground truth labels, with shape :math:`(N, C)`.
Outputs:
Tuple of 2 Tensor, the loss shape is `(N,)`, and the dlogits with the same shape as `logits`.
"""
@prim_attr_register
def __init__(self):
pass
def infer_shape(self, logits_shape, labels_shape):
validator.check_param_equal("SoftmaxCrossEntropyWithLogits logits_shape", logits_shape,
"SoftmaxCrossEntropyWithLogits labels_shape", labels_shape)
loss_shape = [logits_shape[0]]
dlogits_shape = logits_shape
return (loss_shape, dlogits_shape)
def infer_dtype(self, logits_type, labels_type):
args = {"SoftmaxCrossEntropyWithLogits logits_type": logits_type,
"SoftmaxCrossEntropyWithLogits labels_type": labels_type}
validator.check_type_same(args, (mstype.float16, mstype.float32))
return (logits_type, logits_type)
[docs]class SparseSoftmaxCrossEntropyWithLogits(PrimitiveWithInfer):
r"""
Computes the softmax cross-entropy value between logits and sparse encoding labels.
Note:
Sets input logits as `X`, input label as `Y`, output as `loss`. Then,
.. math::
p_{ij} = softmax(X_{ij}) = \frac{exp(x_i)}{\sum_{j = 0}^{N-1}\exp(x_j)}
.. math::
loss_{ij} = \begin{cases} -ln(p_{ij}), &j = y_i \cr -ln(1 - p_{ij}), & j \neq y_i \end{cases}
.. math::
loss = \sum_{ij} loss_{ij}
Args:
is_grad (bool): If it's true, this operation returns the computed gradient. Default: False.
Inputs:
- **logits** (Tensor) - Input logits, with shape :math:`(N, C)`.
- **labels** (Tensor) - Ground truth labels, with shape :math:`(N)`.
Outputs:
Tensor, if `is_grad` is False, the output tensor is the value of loss which is a scalar tensor;
if `is_grad` is True, the output tensor is the gradient of input with the same shape as `logits`.
"""
@prim_attr_register
def __init__(self, is_grad=False):
self.init_prim_io_names(inputs=['features', 'labels'], outputs=['output'])
self.is_grad = is_grad
self.add_prim_attr('sens', 1.0)
def infer_shape(self, logits_shape, labels_shape):
validator.check_param_equal("SparseSoftmaxCrossEntropyWithLogits logits_shape", logits_shape[0],
"SparseSoftmaxCrossEntropyWithLogits labels_shape", labels_shape[0])
loss_shape = []
if self.is_grad:
return logits_shape
return loss_shape
def infer_dtype(self, logits_type, labels_type):
validator.check_typename("SparseSoftmaxCrossEntropyWithLogits logits_type",
logits_type, (mstype.float16, mstype.float32))
validator.check_typename("SparseSoftmaxCrossEntropyWithLogits labels_type",
labels_type, (mstype.int32, mstype.int64))
return logits_type
[docs]class ApplyMomentum(PrimitiveWithInfer):
"""
Optimizer that implements the Momentum algorithm.
Refer to the paper `On the importance of initialization and momentum in deep
learning <https://dl.acm.org/doi/10.5555/3042817.3043064>`_ for more details.
Args:
use_locking (bool): Enable a lock to protect the update of variable and accumlation tensors. Default: False.
use_nesterov (bool): Enable Nesterov momentum. Default: False.
gradient_scale (float): The scale of the gradient. Default: 1.0.
Inputs:
- **variable** (Tensor) - Weights to be updated.
- **accumulation** (Tensor) - Accumulated gradient value by moment weight.
- **learning_rate** (float) - Learning rate.
- **gradient** (Tensor) - Gradients.
- **momentum** (float) - Momentum.
Outputs:
Tensor, parameters to be updated.
Examples:
>>> net = ResNet50()
>>> loss = SoftmaxCrossEntropyWithLogits()
>>> opt = ApplyMomentum(Tensor(np.array([0.001])), Tensor(np.array([0.9])),
filter(lambda x: x.requires_grad, net.get_parameters()))
>>> model = Model(net, loss, opt)
"""
@prim_attr_register
def __init__(self, use_nesterov=False, use_locking=False, gradient_scale=1.0):
self.init_prim_io_names(inputs=['variable', 'accumulation', 'learning_rate', 'gradient', 'momentum'],
outputs=['output'])
def infer_shape(self, v_shape, a_shape, l_shape, g_shape, m_shape):
validator.check(f'variable shape {v_shape}', len(v_shape), '', 0, Rel.GT)
validator.check(f'accumulation shape {a_shape}', len(a_shape), '', 0, Rel.GT)
validator.check(f'learning rate shape {l_shape}', len(l_shape), '', 0, Rel.GE)
validator.check(f'gradient shape {g_shape}', len(g_shape), '', 0, Rel.GE)
validator.check(f'momentum shape {m_shape}', len(m_shape), '', 0, Rel.GE)
return v_shape
def infer_dtype(self, v_dtype, a_dtype, l_dtype, g_dtype, m_dtype):
validator.check_subclass("v_dtype", v_dtype, mstype.tensor)
validator.check_subclass("a_dtype", a_dtype, mstype.tensor)
v_type = validator.check_typename("v_dtype", v_dtype, [mstype.float16, mstype.float32, mstype.float64])
validator.check_typename("a_dtype", a_dtype, [mstype.float16, mstype.float32, mstype.float64])
validator.check_typename("l_dtype", l_dtype, [mstype.float16, mstype.float32, mstype.float64])
validator.check_typename("g_dtype", g_dtype, [mstype.float16, mstype.float32, mstype.float64])
validator.check_typename("m_dtype", m_dtype, [mstype.float16, mstype.float32, mstype.float64])
return v_type
[docs]class SmoothL1Loss(PrimitiveWithInfer):
r"""
Computes smooth L1 loss, a robust L1 loss.
SmoothL1Loss is a Loss similar to MSELoss but less sensitive to outliers as described in the
`Fast R-CNN <https://arxiv.org/abs/1504.08083>`_ by Ross Girshick.
Note:
Sets input prediction as `X`, input target as `Y`, output as `loss`. Then,
.. math::
\text{SmoothL1Loss} = \begin{cases}0.5x^{2}, &if \left |x \right |\leq \text{sigma} \cr
\left |x \right|-0.5, &\text{otherwise}\end{cases}
Args:
sigma (float): A parameter used to control the point where the function will change from
quadratic to linear. Default: 1.0.
Inputs:
- **prediction** (Tensor) - Predict data.
- **target** (Tensor) - Ground truth data, with the same type and shape as `prediction`.
Outputs:
Tensor, with the same type and shape as `prediction`.
"""
@prim_attr_register
def __init__(self, sigma=1.0):
validator.check_type('sigma', sigma, [float])
validator.check('sigma', sigma, '', 0, Rel.GT)
self.init_prim_io_names(inputs=['prediction', 'target'], outputs=['output'])
def infer_shape(self, prediction, target):
validator.check_param_equal('prediction shape', prediction, 'target shape', target)
return prediction
def infer_dtype(self, prediction, target):
args = {"prediction": prediction, "target": target}
validator.check_type_same(args, (mstype.float16, mstype.float32))
return prediction
[docs]class SGD(PrimitiveWithInfer):
"""
Computes stochastic gradient descent (optionally with momentum).
Nesterov momentum is based on the formula from On the importance of
initialization and momentum in deep learning.
Note:
For details, please refer to `nn.SGD` source code.
Args:
dampening (float): The dampening for momentum. Default: 0.0.
weight_decay (float): Weight decay (L2 penalty). Default: 0.0.
nesterov (bool): Enable Nesterov momentum. Default: False.
Inputs:
- **parameters** (Tensor) - Parameters to be updated.
- **gradient** (Tensor) - Gradients.
- **learning_rate** (Tensor) - Learning rate. e.g. Tensor(0.1, mindspore.float32).
- **accum** (Tensor) - Accum(velocity) to be updated.
- **momentum** (Tensor) - Momentum. e.g. Tensor(0.1, mindspore.float32).
- **stat** (Tensor) - States to be updated with the same shape as gradient.
Outputs:
Tensor, parameters to be updated.
"""
@prim_attr_register
def __init__(self, dampening=0.0, weight_decay=0.0, nesterov=False):
self.init_prim_io_names(inputs=['parameters', 'gradient', 'learning_rate', 'accum', 'momentum', 'stat'],
outputs=['output'])
def infer_shape(self, parameters_shape, gradient_shape, learning_rate_shape,
accum_shape, momentum_shape, stat_shape):
validator.check(f'parameters shape {parameters_shape}', len(parameters_shape), '', 0, Rel.GT)
validator.check(f'gradient shape {gradient_shape}', len(gradient_shape), '', 0, Rel.GE)
validator.check(f'learning rate shape {learning_rate_shape}', len(learning_rate_shape), '', 0, Rel.GE)
validator.check(f'accumulation shape {accum_shape}', len(accum_shape), '', 0, Rel.GT)
validator.check(f'momentum shape {momentum_shape}', len(momentum_shape), '', 0, Rel.GE)
validator.check(f'stat shape {stat_shape}', len(stat_shape), '', 0, Rel.GE)
validator.check("gradient shape", gradient_shape, "stat shape", stat_shape)
return parameters_shape
def infer_dtype(self, parameters_dtype, gradient_dtype, learning_rate_dtype,
accum_dtype, momentum_dtype, stat_dtype):
validator.check_typename("parameters_dtype", parameters_dtype, [mstype.float16, mstype.float32])
validator.check_typename("gradient_dtype", gradient_dtype, [mstype.float16, mstype.float32])
validator.check_typename("learning_rate_dtype", learning_rate_dtype, [mstype.float16, mstype.float32])
validator.check_typename("accum_dtype", accum_dtype, [mstype.float16, mstype.float32])
validator.check_typename("momentum_dtype", momentum_dtype, [mstype.float16, mstype.float32])
validator.check_typename("stat_dtype", stat_dtype, [mstype.float16, mstype.float32])
return parameters_dtype
[docs]class LayerNorm(Primitive):
r"""
Applies the Layer Normalization to the input tensor.
This operator will normalize the input tensor on given axis. LayerNorm is described in the paper
`Layer Normalization <https://arxiv.org/abs/1607.06450>`_.
.. math::
y = \frac{x - mean]}{\sqrt{variance + \epsilon}} * \gamma + \beta
where :math:`\gamma` is scale, :math:`\beta` is bias, :math:`\epsilon` is epsilon.
Args:
begin_norm_axis (int): The begin axis of the `input_x` to apply LayerNorm,
the value should be in [-1, rank(input)). Default: 1.
begin_params_axis (int): The begin axis of the parameter input (`gamma`, `beta`) to
apply LayerNorm, the value should be in [-1, rank(input)). Default: 1.
Inputs:
- **input_x** (Tensor) - Tensor of shape :math:`(N, \ldots)`.
The input of LayerNorm.
- **gamma** (Tensor) - Tensor of shape :math:`(P_0, \ldots, P_\text{begin_params_axis})`.
The learnable parameter `gamma` as the scale on norm.
- **beta** (Tensor) - Tensor of shape :math:`(P_0, \ldots, P_\text{begin_params_axis})`.
The learnable parameter `beta` as the scale on norm.
Outputs:
tuple[Tensor], tuple of 3 tensors, the normalized input and the updated parameters.
- **output_x** (Tensor) - The normalized input, has the same type and shape as the `input_x`.
The shape is :math:`(N, C)`.
- **updated_gamma** (Tensor) - Tensor of shape :math:`(C,)`.
- **updated_beta** (Tensor) - Tensor of shape :math:`(C,)`.
"""
@prim_attr_register
def __init__(self, begin_norm_axis=1, begin_params_axis=1):
validator.check_type('begin_norm_axis', begin_norm_axis, [int])
validator.check_type('begin_params_axis', begin_params_axis, [int])
[docs]class L2Normalize(PrimitiveWithInfer):
r"""
L2 normalization Operator.
This operator will normalizes the input using the given axis. The function is shown as follows:
.. math::
\text{output} = \frac{x}{\sqrt{\text{max}(\text{sum} (\text{input_x}^2), \epsilon)}},
where :math:`\epsilon` is epsilon.
Args:
axis (int): The begin axis for the input to apply L2 normalize. Default: 0.
epsilon (float): A small value added for numerical stability. Default: 1e-4.
Inputs:
- **input_x** (Tensor) - Input to compute the normalization.
Outputs:
Tensor, with the same type and shape as the input.
"""
@prim_attr_register
def __init__(self, axis=0, epsilon=1e-4):
validator.check_type('axis', axis, [int])
validator.check_type('epsilon', epsilon, [int, float])
def infer_shape(self, input_x):
dim = len(input_x)
validator.check_int_range('axis value', self.axis, -dim, dim, Rel.INC_LEFT)
return input_x
def infer_dtype(self, input_x):
validator.check_subclass("x", input_x, mstype.tensor)
return input_x
[docs]class DropoutGenMask(Primitive):
"""
Generates the mask value for the input shape.
Args:
Seed0 (int): Seed0 value for random generating. Default: 0.
Seed1 (int): Seed1 value for random generating. Default: 0.
Inputs:
- **shape** (tuple[int]) - The shape of target mask.
- **keep_prob** (Tensor) - The keep rate, between 0 and 1, e.g. keep_prob = 0.9,
means dropping out 10% of input units.
Outputs:
Tensor, the value of generated mask for input shape.
Examples:
>>> dropout_gen_mask = DropoutGenMask()
>>> shape = (20, 16, 50)
>>> keep_prob = Tensor(0.5, mindspore.float32)
>>> mask = dropout_gen_mask(shape, keep_prob)
"""
@prim_attr_register
def __init__(self, Seed0=0, Seed1=0):
self.init_prim_io_names(inputs=['shape', 'keep_prob'], outputs=['output'])
validator.check_type("Seed0", Seed0, [int])
validator.check_type("Seed1", Seed1, [int])
[docs]class DropoutDoMask(PrimitiveWithInfer):
"""
Applies dropout mask on the input tensor.
Take the mask output of DropoutGenMask as input, and apply dropout on the input.
Inputs:
- **input_x** (Tensor) - The input tensor.
- **mask** (Tensor) - The mask to be applied on `input_x`, which is the output of `DropoutGenMask`. And the
shape of `input_x` must be same as the value of `DropoutGenMask`'s input `shape`. If input wrong `mask`,
the output of `DropoutDoMask` are unpredictable.
- **keep_prob** (Tensor) - The keep rate, between 0 and 1, e.g. keep_prob = 0.9,
means dropping out 10% of input units. The value of `keep_prob` is same as the input `keep_prob` of
`DropoutGenMask`.
Outputs:
Tensor, the value that applied dropout on.
Examples:
>>> x = Tensor(np.ones([20, 16, 50]), mindspore.float32)
>>> shape = (20, 16, 50)
>>> keep_prob = Tensor(0.5, mindspore.float32)
>>> dropout_gen_mask = DropoutGenMask()
>>> dropout_do_mask = DropoutDoMask()
>>> mask = dropout_gen_mask(shape, keep_prob)
>>> output = dropout_do_mask(x, mask, keep_prob)
>>> assert output.shape() == (20, 16, 50)
"""
@prim_attr_register
def __init__(self):
pass
def __infer__(self, input_x, mask, keep_prob):
input_x_shape = input_x['shape']
mask_shape = mask['shape']
keep_prob_shape = keep_prob['shape']
validator.check("keep_prob's dim", len(keep_prob_shape), '0(scalar)', 0)
size_x = reduce(lambda x, y: x * y, input_x_shape)
if len(mask_shape) != 1:
raise ValueError("DropoutDoMask mask shape should be 1-dimension.")
size_y = mask_shape[0] * 8
if size_x > size_y:
raise ValueError(f"DropoutDoMask y mask do not math input input_x shape:"
"{input_x_shape}, mask shape: {mask_shape}.")
validator.check_typename("input_x type", input_x['dtype'], [mstype.float32, mstype.float16, mstype.int32])
validator.check_typename("input_mask type", mask['dtype'], [mstype.uint8])
keep_prob_v = keep_prob['value']
if keep_prob_v is not None:
validator.check_const_input('keep_prob', keep_prob_v)
validator.check_number_range('keep_prob', keep_prob_v.asnumpy(), 0, 1, Rel.INC_BOTH)
out = {'shape': input_x_shape,
'dtype': input_x['dtype'],
'value': None}
return out
[docs]class ResizeBilinear(PrimitiveWithInfer):
r"""
Resizes the image to certain size using bilinear interpolation.
The resizing only affects the lower two dimensions which represent the height and width. The input images
can be represented by different data types, but the data types of output images are always float32.
Args:
size (tuple[int]): A tuple of 2 int elements `(new_height, new_width)`, the new size for the images.
align_corners (bool): If it's true, rescale input by `(new_height - 1) / (height - 1)`,
which exactly aligns the 4 corners of images and resized images. If it's false,
rescale by `new_height / height`. Default: False.
Inputs:
- **input** (Tensor) - Image to be resized. Tensor of shape `(N_i, ..., N_n, height, width)`.
Outputs:
Tensor, resized image. Tensor of shape `(N_i, ..., N_n, new_height, new_width)` in `float32`.
Examples:
>>> tensor = Tensor([[[[1, 2, 3, 4, 5], [1, 2, 3, 4, 5]]]], mindspore.int32)
>>> resize_bilinear = P.ResizeBilinear((5, 5))
>>> result = resize_bilinear(tensor)
>>> assert result.shape() == (5, 5)
"""
@prim_attr_register
def __init__(self, size, align_corners=False):
pass
def infer_shape(self, input_shape):
input_shape = list(input_shape)
batch, channel, _, _ = input_shape
out_shape = [batch, channel]
for i in self.size:
out_shape.append(int(i))
return out_shape
def infer_dtype(self, input_dtype):
return mstype.tensor_type(mstype.float32)
[docs]class OneHot(PrimitiveWithInfer):
r"""
Computes a one-hot tensor.
Makes a new tensor, whose locations represented by indices in `indices` take value `on_value`, while all
other locations take value `off_value`.
Note:
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.
Inputs:
- **indices** (Tensor) - A tensor of indices. Tensor of shape :math:`(X_0, \ldots, X_n)`.
- **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`.
- **off_value** (Tensor) - A value to fill in output when `indices[j] != i`.
Outputs:
Tensor, one_hot tensor. Tensor of shape :math:`(X_0, \ldots, X_{axis}, \text{depth} ,X_{axis+1}, \ldots, X_n)`.
Examples:
>>> indices = Tensor(np.array([0, 1, 2]), mindspore.int32)
>>> depth, on_value, off_value = 3, Tensor(1.0, mindspore.float32), Tensor(0.0, mindspore.float32)
>>> onehot = OneHot()
>>> result = onehot(indices, depth, on_value, off_value)
[[1, 0, 0], [0, 1, 0], [0, 0, 1]]
"""
@prim_attr_register
def __init__(self, axis=-1):
self.init_prim_io_names(inputs=['indices', 'depth', 'on_value', 'off_value'], outputs=['output'])
validator.check_type("axis", axis, [int])
def __infer__(self, indices, depth, on_value, off_value):
# check type
validator.check_subclass("indices", indices['dtype'], mstype.tensor)
validator.check_typename("indices", indices['dtype'], (mstype.int32,))
validator.check_typename("depth", depth['dtype'], mstype.int_type)
validator.check_subclass("on_value", on_value['dtype'], mstype.tensor)
validator.check_subclass("off_value", off_value['dtype'], mstype.tensor)
args = {"on_value dtype": on_value['dtype'], "off_value dtype": off_value['dtype']}
validator.check_type_same(args, (mstype.float16, mstype.float32))
# check shape
indices_shp = indices['shape']
validator.check_int_range("axis", self.axis, -1, len(indices_shp), Rel.INC_BOTH)
depth_val = depth['value']
validator.check_integer("depth", depth_val, 0, Rel.GE)
# create new dimension at end if self.axis is -1
indices_shp.insert(self.axis, depth_val) if self.axis >= 0 else indices_shp.append(depth_val)
return {'shape': indices_shp,
'dtype': on_value['dtype'],
'value': None}
[docs]class Gelu(PrimitiveWithInfer):
r"""
Gaussian Error Linear Units activation function.
GeLU is described in the paper `Gaussian Error Linear Units (GELUs) <https://arxiv.org/abs/1606.08415>`_.
And also please refer to `BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding.
<https://arxiv.org/abs/1810.04805>`_.
Defined as follows:
.. math::
\text{output} = 0.5 * x * (1 + erf(x / \sqrt{2})),
where :math:`erf` is the "Gauss error function" .
Inputs:
- **input_x** (Tensor) - Input to compute the Gelu.
Outputs:
Tensor, with the same type and shape as input.
Examples:
>>> tensor = Tensor(np.array([1.0, 2.0, 3.0]), mindspore.float32)
>>> gelu = Gelu()
>>> result = gelu(tensor)
"""
@prim_attr_register
def __init__(self):
"""init GeLU"""
self.init_prim_io_names(inputs=['x'], outputs=['output'])
def infer_shape(self, input_x):
return input_x
def infer_dtype(self, input_x):
validator.check_subclass("input_x", input_x, mstype.tensor)
validator.check_typename("input_x", input_x, (mstype.float16, mstype.float32))
return input_x
[docs]class GetNext(PrimitiveWithInfer):
"""
Returns the next element in the dataset queue.
Note:
GetNext op needs to be associated with network and also depends on the init_dataset interface,
it can't be used directly as a single op.
For details, please refer to `nn.cell_wrapper.DataWrapper` source code.
Args:
types (list[:class:`mindspore.dtype`]): The type of the outputs.
shapes (list[tuple[int]]): The dimensionality of the outputs.
output_num (int): The output number, length of `types` and `shapes`.
shared_name (str): The queue name of `init_dataset` interface.
Inputs:
No inputs.
Outputs:
tuple[Tensor], the output of Dataset. The shape is described in `shapes`
and the type is described is `types`.
Examples:
>>> get_next = GetNext([mindspore.float32, mindspore.int32], [[32, 1, 28, 28], [10]], 'shared_name')
>>> feature, label = get_next()
"""
@prim_attr_register
def __init__(self, types, shapes, output_num, shared_name):
validator.check_type("types", types, [list, tuple])
validator.check_type("shapes", shapes, [list, tuple])
validator.check("types length", len(types), "shapes length", len(shapes))
validator.check_type("output_num", output_num, [int])
def infer_shape(self):
return tuple(self.shapes)
def infer_dtype(self):
return tuple(self.types)
[docs]class PReLU(PrimitiveWithInfer):
r"""
Parametric Rectified Linear Unit activation function.
PReLU is described in the paper `Delving Deep into Rectifiers: Surpassing Human-Level Performance on
ImageNet Classification <https://arxiv.org/abs/1502.01852>`_. Defined as follows:
.. math::
prelu(x_i)= \max(0, x_i) + \min(0, w * x_i),
where :math:`x_i` is an element of an channel of the input.
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,
1 or the number of channels at input.
Outputs:
Tensor, with the same type as `input_x`.
Detailed information, please refer to `nn.PReLU`.
"""
@prim_attr_register
def __init__(self):
pass
def infer_shape(self, input_x_shape, weight_shape):
input_x_dim = len(input_x_shape)
weight_dim = len(weight_shape)
if weight_dim != 1:
raise ValueError(f'weight_dim must be 1, while weight_dim is {weight_dim}.')
if input_x_dim == 1 and weight_shape[0] != 1:
raise ValueError(f'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:
raise ValueError(f'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]}.')
return input_x_shape
def infer_dtype(self, input_x_dtype, weight_dtype):
validator.check_subclass("input_x_dtype", input_x_dtype, mstype.tensor)
validator.check_subclass("weight_dtype", weight_dtype, mstype.tensor)
validator.check_typename("input_x_dtype", input_x_dtype, (mstype.float16, mstype.float32))
validator.check_typename("weight_dtype", weight_dtype, (mstype.float16, mstype.float32))
return input_x_dtype
[docs]class LSTM(PrimitiveWithInfer):
"""
Performs the long short term memory(LSTM) on the input.
Detailed information, please refer to `nn.LSTM`.
"""
@prim_attr_register
def __init__(self, input_size, hidden_size, num_layers, has_bias, bidirectional, dropout):
self.input_size = check_int_positive(input_size)
self.hidden_size = check_int_positive(hidden_size)
self.num_layers = check_int_positive(num_layers)
self.has_bias = check_bool(has_bias)
self.bidirectional = check_bool(bidirectional)
self.dropout = validator.check_type("dropout", dropout, [float])
self.dropout = validator.check_number_range('dropout', dropout, 0, 1, Rel.INC_BOTH)
if bidirectional:
self.num_directions = 2
else:
self.num_directions = 1
def infer_shape(self, x_shape, h_shape, c_shape, w_shape):
# (batch, seq, feature)
validator.check_integer("x_shape", len(x_shape), 3, Rel.EQ)
# h and c should be same shape
validator.check_integer("h_shape", len(h_shape), 3, Rel.EQ)
validator.check_integer("h_shape", len(h_shape), len(c_shape), Rel.EQ)
validator.check_integer("h_shape", h_shape[0], c_shape[0], Rel.EQ)
validator.check_integer("h_shape", h_shape[1], c_shape[1], Rel.EQ)
validator.check_integer("h_shape", h_shape[2], c_shape[2], Rel.EQ)
# (num_layers * num_directions, batch, hidden_size)
validator.check_integer("h[0]", h_shape[0], self.num_layers * self.num_directions, Rel.EQ)
validator.check_integer("h[1]", h_shape[1], x_shape[1], Rel.EQ)
validator.check_integer("h[2]", h_shape[2], self.hidden_size, Rel.EQ)
y_shape = (x_shape[0], x_shape[1], self.hidden_size * self.num_directions)
# set arbitrary shape for reserved space
reserved_shape = (1, 1)
state_shape = (1, 1)
return (y_shape, h_shape, c_shape, reserved_shape, state_shape)
def infer_dtype(self, x_dtype, h_dtype, c_dtype, w_dtype):
validator.check_typename("x_dtype", x_dtype, (mstype.float32, mstype.float16))
validator.check_typename("h_dtype", h_dtype, (mstype.float32, mstype.float16))
validator.check_typename("c_dtype", c_dtype, (mstype.float32, mstype.float16))
validator.check_typename("w_dtype", w_dtype, (mstype.float32, mstype.float16))
validator.check_typename("datatype", x_dtype, (h_dtype.element_type(),))
validator.check_typename("datatype", x_dtype, (c_dtype.element_type(),))
validator.check_typename("datatype", x_dtype, (w_dtype.element_type(),))
return (x_dtype, x_dtype, x_dtype, x_dtype, x_dtype)
[docs]class SigmoidCrossEntropyWithLogits(PrimitiveWithInfer):
r"""
Uses the given logits to compute sigmoid cross entropy.
Note:
Sets input logits as `X`, input label as `Y`, output as `loss`. Then,
.. math::
p_{ij} = sigmoid(X_{ij}) = \frac{1}{1 + e^{-X_{ij}}}
.. math::
loss_{ij} = -[Y_{ij} * ln(p_{ij}) + (1 - Y_{ij})ln(1 - p_{ij})]
Inputs:
- **logits** (Tensor) - Input logits.
- **label** (Tensor) - Ground truth label.
Outputs:
Tensor, with the same shape and type as input `logits`.
"""
@prim_attr_register
def __init__(self):
"""Init SigmoidCrossEntropyWithLogits"""
self.init_prim_io_names(inputs=['predict', 'target'], outputs=['loss'])
def infer_shape(self, x_shape, y_shape):
validator.check_param_equal("x_shape", x_shape, "y_shape", y_shape)
return x_shape
def infer_dtype(self, x_dtype, y_dtype):
args = {"x_dtype": x_dtype, "y_dtype": y_dtype}
validator.check_type_same(args, mstype.number_type)
return x_dtype
[docs]class Pad(PrimitiveWithInfer):
"""
Pads input tensor according to the paddings.
Args:
paddings (tuple): The shape of parameter `paddings` is (N, 2). N is the rank of input data. All elements of
paddings are int type. For `D` th dimension of input, paddings[D, 0] indicates how many sizes to be
extended ahead of the `D` th dimension of the input tensor, and paddings[D, 1] indicates how many sizes to
be extended behind of the `D` th dimension of the input tensor.
Inputs:
- **input_x** (Tensor) - The input tensor.
Outputs:
Tensor, the tensor after padding.
Examples:
>>> input_tensor = Tensor(np.array([[-0.1, 0.3, 3.6], [0.4, 0.5, -3.2]]), mindspore.float32)
>>> pad_op = Pad(((1, 2), (2, 1)))
>>> output_tensor = pad_op(input_tensor)
>>> assert output_tensor == Tensor(np.array([[ 0. , 0. , 0. , 0. , 0. , 0. ],
>>> [ 0. , 0. , -0.1, 0.3, 3.6, 0. ],
>>> [ 0. , 0. , 0.4, 0.5, -3.2, 0. ],
>>> [ 0. , 0. , 0. , 0. , 0. , 0. ],
>>> [ 0. , 0. , 0. , 0. , 0. , 0. ]]), mindspore.float32)
"""
@prim_attr_register
def __init__(self, paddings):
"""Init Pad"""
self.init_prim_io_names(inputs=['x'], outputs=['y'])
if not isinstance(paddings, tuple):
raise TypeError('Paddings must be tuple type.')
for item in paddings:
if len(item) != 2:
raise ValueError('The shape of paddings must be (n, 2).')
def infer_shape(self, x):
paddings = np.array(self.paddings)
validator.check_integer('paddings.shape', paddings.size, len(x) * 2, Rel.EQ)
if not np.all(paddings >= 0):
raise ValueError('All elements of paddings must be >= 0.')
y_shape = ()
for i in range(int(paddings.size / 2)):
y_shape += ((x[i] + paddings[i, 0] + paddings[i, 1]),)
return y_shape
def infer_dtype(self, x):
return x
[docs]class ROIAlign(PrimitiveWithInfer):
"""
Computes Region of Interest (RoI) Align operator.
The operator computes the value of each sampling point by bilinear interpolation from the nearby grid points on the
feature map. No quantization is performed on any coordinates involved in the RoI, its bins, or the sampling
points. The details of (RoI) Align operator are described in `Mask R-CNN <https://arxiv.org/abs/1703.06870>`_.
Args:
pooled_height (int): The output features' height.
pooled_width (int): The output features' width.
spatial_scale (float): A scaling factor that maps the raw image coordinates to the input
feature map coordinates. Suppose the height of a RoI is `ori_h` in the raw image and `fea_h` in the
input feature map, the `spatial_scale` should be `fea_h / ori_h`.
sample_num (int): Number of sampling points. Default: 2.
Inputs:
- **features** (Tensor) - The input features, whose shape should be `(N, C, H, W)`.
- **rois** (Tensor) - The shape is `(rois_n, 5)`. `rois_n` represents the number of RoI. The size of
the second dimension should be `5` and the `5` colunms are
`(image_index, top_left_x, top_left_y, bottom_right_x, bottom_right_y)`. `image_index` represents the
index of image. `top_left_x` and `top_left_y` represent the `x, y` coordinates of the top left corner
of corresponding RoI, respectively. `bottom_right_x` and `bottom_right_y` represent the `x, y`
coordinates of the bottom right corner of corresponding RoI, respectively.
Outputs:
Tensor, the shape is `(rois_n, C, pooled_height, pooled_width)`.
Examples:
>>> input_tensor = Tensor(np.array([[[[1., 2.], [3., 4.]]]]), mindspore.float32)
>>> rois = Tensor(np.array([[0, 0.2, 0.3, 0.2, 0.3]]), mindspore.float32)
>>> roi_align = P.ROIAlign(1, 1, 0.5, 2)
>>> output_tensor = roi_align(input_tensor, rois)
>>> assert output_tensor == Tensor(np.array([[[[2.15]]]]), mindspore.float32)
"""
@prim_attr_register
def __init__(self, pooled_height, pooled_width, spatial_scale, sample_num=2):
"""init ROIAlign"""
validator.check_type("pooled_height", pooled_height, [int])
validator.check_type("pooled_width", pooled_width, [int])
validator.check_type("spatial_scale", spatial_scale, [float])
validator.check_type("sample_num", sample_num, [int])
self.pooled_height = pooled_height
self.pooled_width = pooled_width
self.spatial_scale = spatial_scale
self.sample_num = sample_num
def infer_shape(self, inputs_shape, rois_shape):
return [rois_shape[0], inputs_shape[1], self.pooled_height, self.pooled_width]
def infer_dtype(self, inputs_type, rois_type):
return inputs_type
[docs]class Adam(PrimitiveWithInfer):
r"""
Updates gradients by Adaptive Moment Estimation (Adam) algorithm.
The Adam algorithm is proposed in `Adam: A Method for Stochastic Optimization <https://arxiv.org/abs/1412.6980>`_.
The updating formulas are as follows,
.. math::
\begin{array}{ll} \\
m = \beta_1 * m + (1 - \beta_1) * g \\
v = \beta_2 * v + (1 - \beta_2) * g * g \\
l = \alpha * \frac{\sqrt{1-\beta_2^t}}{1-\beta_1^t} \\
w = w - l * \frac{m}{\sqrt{v} + \epsilon}
\end{array}
:math:`m` represents the 1st moment vector, :math:`v` represents the 2nd moment vector, :math:`g` represents
`gradient`, :math:`l` represents scaling factor `lr`, :math:`\beta_1, \beta_2` represent `beta1` and `beta2`,
:math:`t` represents updating step while :math:`beta_1^t` and :math:`beta_2^t` represent `beta1_power` and
`beta2_power`, :math:`\alpha` represents `learning_rate`, :math:`w` represents `var`, :math:`\epsilon` represents
`epsilon`.
Args:
use_locking (bool): Whether to enable a lock to protect updating variable tensors.
If True, updating of the var, m, and v tensors will be protected by a lock.
If False, the result is unpredictable. Default: False.
use_nesterov (bool): Whether to use Nesterov Accelerated Gradient (NAG) algorithm to update the gradients.
If True, updates the gradients using NAG.
If False, updates the gradients without using NAG. Default: False.
Inputs:
- **var** (Tensor) - Weights to be updated.
- **m** (Tensor) - The 1st moment vector in the updating formula.
- **v** (Tensor) - the 2nd moment vector in the updating formula.
- **beta1_power** (float) - :math:`beta_1^t` in the updating formula.
- **beta2_power** (float) - :math:`beta_2^t` in the updating formula.
- **lr** (float) - :math:`l` in the updating formula.
- **beta1** (float) - The exponential decay rate for the 1st moment estimates.
- **beta2** (float) - The exponential decay rate for the 2nd moment estimates.
- **epsilon** (float) - Term added to the denominator to improve numerical stability.
- **gradient** (Tensor) - Gradients.
Outputs:
Tensor, has the same shape and data type as `var`.
"""
@prim_attr_register
def __init__(self, use_locking=False, use_nesterov=False):
validator.check_type("use_locking", use_locking, [bool])
validator.check_type("use_nesterov", use_nesterov, [bool])
def infer_shape(self, var_shape, m_shape, v_shape, beta1_power_shape, beta2_power_shape, lr_shape,
beta1_shape, beta2_shape, epsilon_shape, grad_shape):
validator.check_param_equal("var_shape", var_shape, "m_shape", m_shape)
validator.check_param_equal("var_shape", var_shape, "v_shape", v_shape)
validator.check_param_equal("var_shape", var_shape, "grad_shape", grad_shape)
return var_shape
def infer_dtype(self, var_dtype, m_dtype, v_dtype, beta1_power_dtype, beta2_power_dtype, lr_dtype,
beta1_dtype, beta2_dtype, epsilon_dtype, grad_dtype):
args = {"var_dtype": var_dtype, "m_dtype": m_dtype, "v_dtype": v_dtype, "grad_dtype": grad_dtype}
validator.check_type_same(args, mstype.number_type)
args = {"beta1_power_dtype": beta1_power_dtype, "beta2_power_dtype": beta2_power_dtype, 'lr_dtype': lr_dtype,
"beta1_dtype": beta1_dtype, "beta2_dtype": beta2_dtype, "epsilon_dtype": epsilon_dtype}
validator.check_type_same(args, [mstype.float16, mstype.float32])
return var_dtype
[docs]class BinaryCrossEntropy(PrimitiveWithInfer):
r"""
Computes the Binary Cross Entropy between the target and the output.
Note:
Sets input as :math:`x`, input label as :math:`y`, output as :math:`\ell(x, y)`.
Let,
.. math::
L = \{l_1,\dots,l_N\}^\top, \quad
l_n = - w_n \left[ y_n \cdot \log x_n + (1 - y_n) \cdot \log (1 - x_n) \right]
Then,
.. math::
\ell(x, y) = \begin{cases}
L, & \text{if reduction} = \text{'none';}\\
\operatorname{mean}(L), & \text{if reduction} = \text{'mean';}\\
\operatorname{sum}(L), & \text{if reduction} = \text{'sum'.}
\end{cases}
Args:
reduction (str): Specifies the reduction to apply to the output.
Its value should be one of 'none', 'mean', 'sum'. Default: 'mean'.
Inputs:
- **input_x** (Tensor) - The input Tensor.
- **input_y** (Tensor) - The label Tensor which has same shape as `input_x`.
- **weight** (Tensor, optional) - A rescaling weight applied to the loss of each batch element.
And it should have same shape as `input_x`. Default: None.
Outputs:
Tensor or Scalar, if `reduction` is 'none', then output is a tensor and same shape as `input_x`.
Otherwise it is a scalar.
"""
@prim_attr_register
def __init__(self, reduction='mean'):
self.reduction = validator.check_string('reduction', reduction, ['none', 'mean', 'sum'])
def infer_shape(self, x_shape, y_shape, weight_shape):
validator.check_param_equal('x_shape', x_shape, 'y_shape', y_shape)
if weight_shape:
validator.check_param_equal('y_shape', y_shape, 'weight_shape', weight_shape)
if self.reduction in ('mean', 'sum'):
shape = []
else:
shape = x_shape
return shape
def infer_dtype(self, x_type, y_type, weight_type):
args = {'x_type': x_type, 'y_type': y_type}
validator.check_type_same(args, (mstype.float16, mstype.float32))
if weight_type:
validator.check_two_types_same('x_type', x_type, 'weight_type', weight_type)
return x_type
[docs]class SparseApplyAdagrad(PrimitiveWithInfer):
r"""
Update relevant entries according to the adagrad scheme.
.. math::
accum += grad * grad
.. math::
var -= lr * grad * (1 / sqrt(accum))
Args:
lr (float): Learning rate.
use_locking (bool): If True, updating of the var and accum tensors will be protected. Default: False.
Inputs:
- **var** (Tensor) - Variable to be updated. The type must be float32.
- **accum** (Tensor) - Accum to be updated. The shape must be the same as `var`'s shape,
the type must be float32.
- **grad** (Tensor) - Gradient. The shape must be the same as `var`'s shape
except first dimension, the type must be float32.
- **indices** (Tensor) - A vector of indices into the first dimension of `var` and `accum`.
The shape of `indices` must be the same as `grad` in first dimension, the type must be int32.
Outputs:
Tensor, has the same shape and type as `var`.
"""
@prim_attr_register
def __init__(self, lr, use_locking=False):
self.lr = validator.check_type("lr", lr, [float])
self.use_locking = validator.check_type("use_locking", use_locking, [bool])
def infer_shape(self, var_shape, accum_shape, grad_shape, indices_shape):
validator.check_param_equal('var shape', var_shape, 'accum shape', accum_shape)
validator.check_param_equal('len of var shape', len(var_shape), 'len of grad shape', len(grad_shape))
if len(var_shape) > 1:
validator.check_param_equal('var_shape', var_shape[1:], 'grad_shape', grad_shape[1:])
validator.check_integer("len of indices shape", len(indices_shape), 1, Rel.EQ)
validator.check('the first dimension of grad', grad_shape[0],
'the shape of indices', indices_shape[0], Rel.EQ)
return var_shape
def infer_dtype(self, var_type, accum_type, grad_type, indices_type):
validator.check_subclass("var_type", var_type, mstype.tensor)
validator.check_subclass("accum_type", accum_type, mstype.tensor)
validator.check_subclass("grad_type", grad_type, mstype.tensor)
validator.check_subclass("indices_type", indices_type, mstype.tensor)
args = {'var_type': var_type, 'accum_type': accum_type, 'grad_type': grad_type}
validator.check_type_same(args, (mstype.float32,))
validator.check_typename('indices_type', indices_type, [mstype.int32])
return var_type
[docs]class LARSUpdate(PrimitiveWithInfer):
"""
Conduct lars (layer-wise adaptive rate scaling) update on the square sum of gradient.
Args:
epsilon (float): Term added to the denominator to improve numerical stability. Default: 1e-05.
hyperpara (float): Trust coefficient for calculating the local learning rate. Default: 0.001.
use_clip (bool): Whether to use clip operation for calculating the local learning rate. Default: False.
Inputs:
- **weight** (Tensor) - The weight to be updated.
- **gradient** (Tensor) - The gradient of weight, which has the same shape and dtype with weight.
- **norm_weight** (Tensor) - A scalar tensor, representing the square sum of weight.
- **norm_gradient** (Tensor) - A scalar tensor, representing the square sum of gradient.
- **weight_decay** (Union[Number, Tensor]) - Weight decay. It should be a scalar tensor or number.
- **learning_rate** (Union[Number, Tensor]) - Learning rate. It should be a scalar tensor or number.
Outputs:
Tensor, representing the new gradient.
"""
@prim_attr_register
def __init__(self, epsilon=1e-05, hyperpara=0.001, use_clip=False):
"""init"""
validator.check_type("epsilon", epsilon, [float])
validator.check_type("hyperpara", hyperpara, [float])
validator.check_type("use_clip", use_clip, [bool])
def infer_shape(self, weight_shape, gradient_shape, norm_weight_shape, norm_gradient_shape, weight_decay_shape,
learning_rate_shape):
validator.check_param_equal("Weight shape", weight_shape, "gradient shape", gradient_shape)
validator.check_param_equal("Norm weight shape", norm_weight_shape, "norm gradient shape", norm_gradient_shape)
shp_len = len(weight_decay_shape)
validator.check_shape_length("Weight decay's shape", shp_len, 1, Rel.LE)
if shp_len == 1:
validator.check_integer("Weight decay's shape", weight_decay_shape[0], 1, Rel.EQ)
shp_len = len(learning_rate_shape)
validator.check_shape_length("Learning rate's shape", shp_len, 1, Rel.LE)
if shp_len == 1:
validator.check_integer("Learning rate's shape", learning_rate_shape[0], 1, Rel.EQ)
return weight_shape
def infer_dtype(self, weight_dtype, gradient_dtype, norm_weight_dtype, norm_gradient_dtype,
weight_decay_dtype, learning_rate_dtype):
args = {"Weight dtype": weight_dtype, "gradient dtype": gradient_dtype, "norm weight dtype": norm_weight_dtype,
"norm gradient dtype": norm_gradient_dtype}
validator.check_type_same(args, [mstype.float16, mstype.float32, mstype.int16, mstype.int32])
validator.check_args_tensor(args)
validator.check_typename("weight_decay_dtype", weight_decay_dtype,
[mstype.float16, mstype.float32, mstype.float64])
validator.check_typename("learning_rate_dtype", learning_rate_dtype,
[mstype.float16, mstype.float32, mstype.float64])
return weight_dtype