Source code for mindspore.nn.layer.basic

# 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.
# ============================================================================
"""basic"""
import numpy as np
import mindspore.common.dtype as mstype
from mindspore.common.tensor import Tensor
from mindspore.common.initializer import initializer
from mindspore._checkparam import check_int_positive, check_bool
from mindspore.ops import operations as P
from mindspore.ops import functional as F
from mindspore.ops.functional import identity
from mindspore.ops.operations import _inner_ops as inner
from mindspore.ops.primitive import constexpr
from mindspore.common.parameter import Parameter
from mindspore._extends import cell_attr_register
from mindspore.common.api import ms_function
from mindspore import context
from mindspore.ops import _selected_ops
from ..cell import Cell
from .activation import get_activation
from ..._checkparam import Validator as validator


__all__ = ['Dropout', 'Flatten', 'Dense', 'ClipByNorm', 'Norm', 'OneHot', 'Pad', 'Unfold']

[docs]class Dropout(Cell): r""" Dropout layer for the input. Randomly set some elements of the input tensor to zero with probability :math:`1 - keep\_prob` during training using samples from a Bernoulli distribution. Note: Each channel will be zeroed out independently on every construct call. The outputs are scaled by a factor of :math:`\frac{1}{keep\_prob}` during training so that the output layer remains at a similar scale. During inference, this layer returns the same tensor as the input. This technique is proposed in paper `Dropout: A Simple Way to Prevent Neural Networks from Overfitting <http://www.cs.toronto.edu/~rsalakhu/papers/srivastava14a.pdf>`_ and proved to be effective to reduce over-fitting and prevents neurons from co-adaptation. See more details in `Improving neural networks by preventing co-adaptation of feature detectors <https://arxiv.org/pdf/1207.0580.pdf>`_. Args: keep_prob (float): The keep rate, greater than 0 and less equal than 1. E.g. rate=0.9, dropping out 10% of input units. Default: 0.5. seed0 (int): The first random seed. Default: 0. seed1 (int): The second random seed. Default: 0. dtype (:class:`mindspore.dtype`): Data type of input. Default: mindspore.float32. Raises: ValueError: If keep_prob is not in range (0, 1). Inputs: - **input** (Tensor) - An N-D Tensor. Outputs: Tensor, output tensor with the same shape as the input. Examples: >>> x = Tensor(np.ones([20, 16, 50]), mindspore.float32) >>> net = nn.Dropout(keep_prob=0.8) >>> net(x) """ def __init__(self, keep_prob=0.5, seed0=0, seed1=0, dtype=mstype.float32): super(Dropout, self).__init__() if keep_prob <= 0 or keep_prob > 1: raise ValueError("dropout probability should be a number in range (0, 1], but got {}".format(keep_prob)) validator.check_subclass("dtype", dtype, mstype.number_type, self.cls_name) self.keep_prob = keep_prob self.seed0 = seed0 self.seed1 = seed1 self.dtype = dtype self.get_shape = P.Shape() self.dropout_gen_mask = P.DropoutGenMask(Seed0=seed0, Seed1=seed1) self.dropout_do_mask = P.DropoutDoMask() self.cast = P.Cast() self.is_gpu = context.get_context('device_target') in ["GPU"] if self.is_gpu: self.dropout = P.Dropout(keep_prob) def construct(self, x): if not self.training: return x if self.is_gpu: out, _ = self.dropout(x) return out if self.keep_prob == 1: return x shape = self.get_shape(x) dtype = P.DType()(x) keep_prob = self.cast(self.keep_prob, dtype) output = self.dropout_gen_mask(shape, keep_prob) return self.dropout_do_mask(x, output, keep_prob) def extend_repr(self): str_info = 'keep_prob={}, Seed0={}, Seed1={}, dtype={}' \ .format(self.keep_prob, self.seed0, self.seed1, self.dtype) return str_info
[docs]class Flatten(Cell): r""" Flatten layer for the input. Flattens a tensor without changing dimension of batch size on the 0-th axis. Inputs: - **input** (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 dimensions. Examples: >>> net = nn.Flatten() >>> input = Tensor(np.array([[[1.2, 1.2], [2.1, 2.1]], [[2.2, 2.2], [3.2, 3.2]]]), mindspore.float32) >>> input.shape (2, 2, 2) >>> net(input) [[1.2 1.2 2.1 2.1] [2.2 2.2 3.2 3.2]] """ def __init__(self): super(Flatten, self).__init__() def construct(self, x): return F.reshape(x, (F.shape(x)[0], -1))
[docs]class Dense(Cell): r""" The fully connected layer. Applies dense-connected layer for the input. This layer implements the operation as: .. math:: \text{outputs} = \text{activation}(\text{inputs} * \text{kernel} + \text{bias}), where :math:`\text{activation}` is the activation function passed as the activation argument (if passed in), :math:`\text{activation}` is a weight matrix with the same data type as the inputs created by the layer, and :math:`\text{bias}` is a bias vector with the same data type as the inputs created by the layer (only if has_bias is True). Args: in_channels (int): The number of channels in the input space. out_channels (int): The number of channels in the output space. weight_init (Union[Tensor, str, Initializer, numbers.Number]): The trainable weight_init parameter. The dtype is same as input x. The values of str refer to the function `initializer`. Default: 'normal'. bias_init (Union[Tensor, str, Initializer, numbers.Number]): The trainable bias_init parameter. The dtype is same as input x. The values of str refer to the function `initializer`. Default: 'zeros'. has_bias (bool): Specifies whether the layer uses a bias vector. Default: True. activation (str): Regularizer function applied to the output of the layer, eg. 'relu'. Default: None. Raises: ValueError: If weight_init or bias_init shape is incorrect. Inputs: - **input** (Tensor) - Tensor of shape :math:`(N, in\_channels)`. Outputs: Tensor of shape :math:`(N, out\_channels)`. Examples: >>> net = nn.Dense(3, 4) >>> input = Tensor(np.random.randint(0, 255, [2, 3]), mindspore.float32) >>> net(input) [[ 2.5246444 2.2738023 0.5711005 -3.9399147 ] [ 1.0739875 4.0155234 0.94188046 -5.459526 ]] """ @cell_attr_register(attrs=['has_bias', 'activation']) def __init__(self, in_channels, out_channels, weight_init='normal', bias_init='zeros', has_bias=True, activation=None): super(Dense, self).__init__() self.in_channels = check_int_positive(in_channels) self.out_channels = check_int_positive(out_channels) self.has_bias = check_bool(has_bias) if isinstance(weight_init, Tensor): if weight_init.dim() != 2 or weight_init.shape[0] != out_channels or \ weight_init.shape[1] != in_channels: raise ValueError("weight_init shape error") self.weight = Parameter(initializer(weight_init, [out_channels, in_channels]), name="weight") if self.has_bias: if isinstance(bias_init, Tensor): if bias_init.dim() != 1 or bias_init.shape[0] != out_channels: raise ValueError("bias_init shape error") self.bias = Parameter(initializer(bias_init, [out_channels]), name="bias") self.matmul = P.MatMul(transpose_b=True) self.bias_add = _selected_ops.BiasAdd() self.activation = get_activation(activation) self.activation_flag = self.activation is not None def construct(self, x): output = self.matmul(x, self.weight) if self.has_bias: output = self.bias_add(output, self.bias) if self.activation_flag: return self.activation(output) return output def extend_repr(self): str_info = 'in_channels={}, out_channels={}, weight={}, has_bias={}' \ .format(self.in_channels, self.out_channels, self.weight, self.has_bias) if self.has_bias: str_info = str_info + ', bias={}'.format(self.bias) if self.activation_flag: str_info = str_info + ', activation={}'.format(self.activation) return str_info
@constexpr def _is_equal_one(x): if x is None: return False return bool(x.asnumpy().mean() == 1.0)
[docs]class ClipByNorm(Cell): r""" Clips tensor values to a maximum :math:`L_2`-norm. The output of this layer remains the same if the :math:`L_2`-norm of the input tensor is not greater than the argument clip_norm. Otherwise the tensor will be normalized as: .. math:: \text{output}(X) = \frac{\text{clip_norm} * X}{L_2(X)}, where :math:`L_2(X)` is the :math:`L_2`-norm of :math:`X`. Inputs: - **input** (Tensor) - Tensor of shape N-D. - **clip_norm** (Tensor) - A scalar Tensor of shape :math:`()` or :math:`(1)` and of the same type as the input Tensor. Outputs: Tensor, clipped tensor with the same shape as the input. Examples: >>> net = nn.ClipByNorm() >>> input = Tensor(np.random.randint(0, 10, [4, 16]), mindspore.float32) >>> clip_norm = Tensor(np.array([100]).astype(np.float32)) >>> net(input, clip_norm) """ def __init__(self): super(ClipByNorm, self).__init__() self.reduce_sum = P.ReduceSum(keep_dims=True) self.select_ = P.Select() self.greater_ = P.Greater() self.axis = () self.cast = P.Cast() self.zero = Tensor(np.array([0.0]).astype(np.float32)) self.sqrt = P.Sqrt() self.max_op = P.Maximum() self.shape = P.Shape() self.reshape = P.Reshape() self.fill = P.Fill() self.expand_dims = P.ExpandDims() self.dtype = P.DType()
[docs] @ms_function def construct(self, x, clip_norm): """add ms_function decorator for pynative mode""" mul_x = F.square(x) l2sum = self.cast(self.reduce_sum(mul_x, self.axis), mstype.float32) cond = self.greater_(l2sum, self.zero) ones_ = self.fill(self.dtype(cond), self.shape(cond), 1.0) l2sum_safe = self.select_(cond, l2sum, self.cast(ones_, self.dtype(l2sum))) l2norm = self.select_(cond, self.sqrt(l2sum_safe), l2sum) if _is_equal_one(clip_norm): intermediate = x else: intermediate = x * clip_norm max_norm = self.max_op(l2norm, clip_norm) values_clip = self.cast(intermediate, mstype.float32) / self.expand_dims(max_norm, -1) values_clip = self.reshape(values_clip, self.shape(x)) values_clip = identity(values_clip) return values_clip
[docs]class Norm(Cell): """ Computes the norm of vectors, currently including Euclidean norm, i.e., :math:`L_2`-norm. Args: axis (tuple): The axis over which to compute vector norms. Default: (). keep_dims (bool): If True, the axis indicated in `axis` are kept with size 1. Otherwise, the dimensions in `axis` are removed from the output shape. Default: False. Inputs: - **input** (Tensor) - Tensor which is not empty. Outputs: Tensor, output tensor with dimensions in 'axis' reduced to 1 will be returned if 'keep_dims' is True; otherwise a Tensor with dimensions in 'axis' removed is returned. Examples: >>> net = nn.Norm(axis=0) >>> input = Tensor(np.random.randint(0, 10, [4, 16]), mindspore.float32) >>> net(input) """ def __init__(self, axis=(), keep_dims=False): super(Norm, self).__init__() self.axis = axis self.keep_dims = keep_dims self.reduce_sum = P.ReduceSum(True) self.sqrt = P.Sqrt() self.squeeze = P.Squeeze(self.axis) def construct(self, x): x = self.sqrt(self.reduce_sum(F.square(x), self.axis)) if not self.keep_dims: x = self.squeeze(x) return x def extend_repr(self): str_info = 'axis={}, keep_dims={}'.format(self.axis, self.keep_dims) return str_info
[docs]class OneHot(Cell): """ Returns a one-hot tensor. The locations represented by indices in argument 'indices' take value on_value, while all other locations take value off_value. Note: If the input indices is rank :math:`N`, the output will have rank :math:`N+1`. The new axis is created at dimension `axis`. Args: axis (int): Features x depth if axis == -1, depth x features if axis == 0. Default: -1. depth (int): A scalar defining the depth of the one hot dimension. Default: 1. on_value (float): A scalar defining the value to fill in output[i][j] when indices[j] = i. Default: 1.0. off_value (float): A scalar defining the value to fill in output[i][j] when indices[j] != i. Default: 0.0. dtype (:class:`mindspore.dtype`): Data type of 'on_value' and 'off_value', not the data type of indices. Default: mindspore.float32. Inputs: - **indices** (Tensor) - A tensor of indices of data type mindspore.int32 and arbitrary shape. Outputs: Tensor, the one-hot tensor of data type 'dtype' with dimension at 'axis' expanded to 'depth' and filled with on_value and off_value. Examples: >>> net = nn.OneHot(depth=4, axis=1) >>> indices = Tensor([[1, 3], [0, 2]], dtype=mindspore.int32) >>> net(indices) [[[0. 0.] [1. 0.] [0. 0.] [0. 1.]] [[1. 0.] [0. 0.] [0. 1.] [0. 0.]]] """ def __init__(self, axis=-1, depth=1, on_value=1.0, off_value=0.0, dtype=mstype.float32): super(OneHot, self).__init__() self.onehot = P.OneHot(axis) self.depth = depth self.on_value = Tensor(on_value, dtype) self.off_value = Tensor(off_value, dtype) def construct(self, indices): return self.onehot(indices, self.depth, self.on_value, self.off_value)
[docs]class Pad(Cell): """ Pads the input tensor according to the paddings and mode. 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. mode (str): Specifies padding mode. The optional values are "CONSTANT", "REFLECT", "SYMMETRIC". Default: "CONSTANT". Inputs: - **input_x** (Tensor) - The input tensor. Outputs: Tensor, the tensor after padding. - If `mode` is "CONSTANT", it fills the edge with 0, regardless of the values of the `input_x`. If the `input_x` is [[1,2,3],[4,5,6],[7,8,9]] and `paddings` is [[1,1],[2,2]], then the Outputs is [[0,0,0,0,0,0,0],[0,0,1,2,3,0,0],[0,0,4,5,6,0,0],[0,0,7,8,9,0,0],[0,0,0,0,0,0,0]]. - If `mode` is "REFLECT", it uses a way of symmetrical copying throught the axis of symmetry to fill in. If the `input_x` is [[1,2,3],[4,5,6],[7,8,9]] and `paddings` is [[1,1],[2,2]], then the Outputs is [[6,5,4,5,6,5,4],[3,2,1,2,3,2,1],[6,5,4,5,6,5,4],[9,8,7,8,9,8,7],[6,5,4,5,6,5,4]]. - If `mode` is "SYMMETRIC", the filling method is similar to the "REFLECT". It is also copied according to the symmetry axis, except that it includes the symmetry axis. If the `input_x` is [[1,2,3],[4,5,6],[7,8,9]] and `paddings` is [[1,1],[2,2]], then the Outputs is [[2,1,1,2,3,3,2],[2,1,1,2,3,3,2],[5,4,4,5,6,6,5],[8,7,7,8,9,9,8],[8,7,7,8,9,9,8]]. Examples: >>> from mindspore import Tensor >>> from mindspore.ops import operations as P >>> import mindspore.nn as nn >>> import numpy as np >>> class Net(nn.Cell): >>> def __init__(self): >>> super(Net, self).__init__() >>> self.pad = nn.Pad(paddings=((1,1),(2,2)), mode="CONSTANT") >>> def construct(self, x): >>> return self.pad(x) >>> x = np.random.random(size=(2, 3)).astype(np.float32) >>> pad = Net() >>> ms_output = pad(Tensor(x)) """ def __init__(self, paddings, mode="CONSTANT"): super(Pad, self).__init__() self.mode = mode self.paddings = paddings validator.check_string('mode', self.mode, ["CONSTANT", "REFLECT", "SYMMETRIC"], self.cls_name) 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).') if mode == "CONSTANT": self.pad = P.Pad(self.paddings) else: self.paddings = Tensor(np.array(self.paddings)) self.pad = P.MirrorPad(mode=mode) def construct(self, x): if self.mode == "CONSTANT": x = self.pad(x) else: x = self.pad(x, self.paddings) return x
[docs]class Unfold(Cell): """ Extract patches from images. The input tensor must be a 4-D tensor and the data format is NCHW. Args: ksizes (Union[tuple[int], list[int]]): The size of sliding window, should be a tuple or list of int, and the format is [1, ksize_row, ksize_col, 1]. strides (Union[tuple[int], list[int]]): Distance between the centers of the two consecutive patches, should be a tuple or list of int, and the format is [1, stride_row, stride_col, 1]. rates (Union[tuple[int], list[int]]): In each extracted patch, the gap between the corresponding dim pixel positions, should be a tuple or list of int, and the format is [1, rate_row, rate_col, 1]. padding (str): The type of padding algorithm, is a string whose value is "same" or "valid", not case sensitive. Default: "valid". - same: Means that the patch can take the part beyond the original image, and this part is filled with 0. - valid: Means that the patch area taken must be completely contained in the original image. Inputs: - **input_x** (Tensor) - A 4-D tensor whose shape is [in_batch, in_depth, in_row, in_col] and data type is number. Outputs: Tensor, a 4-D tensor whose data type is same as 'input_x', and the shape is [out_batch, out_depth, out_row, out_col], the out_batch is same as the in_batch. Examples: >>> net = Unfold(ksizes=[1, 2, 2, 1], strides=[1, 1, 1, 1], rates=[1, 1, 1, 1]) >>> image = Tensor(np.ones([1, 1, 3, 3]), dtype=mstype.float16) >>> net(image) Tensor ([[[[1, 1] [1, 1]] [[1, 1], [1, 1]] [[1, 1] [1, 1]], [[1, 1], [1, 1]]]], shape=(1, 4, 2, 2), dtype=mstype.float16) """ def __init__(self, ksizes, strides, rates, padding="valid"): super(Unfold, self).__init__() self.extract_image_patches = inner.ExtractImagePatches(ksizes, strides, rates, padding) self.transpose = P.Transpose() self.format_NHWC = (0, 2, 3, 1) self.format_NCHW = (0, 3, 1, 2) def construct(self, input_x): x_transpose = self.transpose(input_x, self.format_NHWC) ret = self.extract_image_patches(x_transpose) ret_transpose = self.transpose(ret, self.format_NCHW) return ret_transpose