PyTorch中的torch.nn

PyTorch提供了torch.nn模块，以帮助我们创建和训练神经网络。我们将首先在MNIST数据集上训练基本神经网络，而无需使用这些模型的任何功能。我们将仅使用基本的PyTorch张量功能，然后一次从torch.nn中增量添加一个功能。

torch.nn为我们提供了更多的类和模块来实现和训练神经网络。

nn软件包包含以下模块和类：

S.No	Class and Module	Description
1.	torch.nn.Parameter	It is a type of tensor which is to be considered as a module parameter.
2.	Containers
	1) torch.nn.Module	It is a base class for all neural network module.
	2) torch.nn.Sequential	It is a sequential container in which Modules will be added in the same order as they are passed in the constructor.
	3) torch.nn.ModuleList	This will holds sub-modules in a list.
	4) torch.nn.ModuleDict	This will holds sub-modules in a directory.
	5) torch.nn.ParameterList	This will holds the parameters in a list.
	6) torch.nn.parameterDict	This will holds the parameters in a directory.
3.	Convolution layers
	1) torch.nn.Conv1d	This package will be used to apply a 1D convolution over an input signal composed of several input planes.
	2) torch.nn.Conv2d	This package will be used to apply a 2D convolution over an input signal composed of several input planes.
	3) torch.nn.Conv3d	This package will be used to apply a 3D convolution over an input signal composed of several input planes.
	4) torch.nn.ConvTranspose1d	This package will be used to apply a 1D transposed convolution operator over an input image composed of several input planes.
	5) torch.nn.ConvTranspose2d	This package will be used to apply a 2D transposed convolution operator over an input image composed of several input planes.
	6) torch.nn.ConvTranspose3d	This package will be used to apply a 3D transposed convolution operator over an input image composed of several input planes.
	7) torch.nn.Unfold	It is used to extracts sliding local blocks from a batched input tensor.
	8) torch.nn.Fold	It is used to combine an array of sliding local blocks into a large containing tensor.
4.	Pooling layers
	1) torch.nn.MaxPool1d	It is used to apply a 1D max pooling over an input signal composed of several input planes.
	2) torch.nn.MaxPool2d	It is used to apply a 2D max pooling over an input signal composed of several input planes.
	3) torch.nn.MaxPool3d	It is used to apply a 3D max pooling over an input signal composed of several input planes.
	4) torch.nn.MaxUnpool1d	It is used to compute the partial inverse of MaxPool1d.
	5) torch.nn.MaxUnpool2d	It is used to compute the partial inverse of MaxPool2d.
	6) torch.nn.MaxUnpool3d	It is used to compute the partial inverse of MaxPool3d.
	7) torch.nn.AvgPool1d	It is used to apply a 1D average pooling over an input signal composed of several input planes.
	8) torch.nn.AvgPool2d	It is used to apply a 2D average pooling over an input signal composed of several input planes.
	9) torch.nn.AvgPool3d	It is used to apply a 3D average pooling over an input signal composed of several input planes.
	10) torch.nn.FractionalMaxPool2d	It is used to apply a 2D fractional max pooling over an input signal composed of several input planes.
	11) torch.nn.LPPool1d	It is used to apply a 1D power-average pooling over an input signal composed of several input planes.
	12) torch.nn.LPPool2d	It is used to apply a 2D power-average pooling over an input signal composed of several input planes.
	13) torch.nn.AdavtiveMaxPool1d	It is used to apply a 1D adaptive max pooling over an input signal composed of several input planes.
	14) torch.nn.AdavtiveMaxPool2d	It is used to apply a 2D adaptive max pooling over an input signal composed of several input planes.
	15) torch.nn.AdavtiveMaxPool3d	It is used to apply a 3D adaptive max pooling over an input signal composed of several input planes.
	16) torch.nn.AdavtiveAvgPool1d	It is used to apply a 1D adaptive average pooling over an input signal composed of several input planes.
	17) torch.nn.AdavtiveAvgPool2d	It is used to apply a 2D adaptive average pooling over an input signal composed of several input planes.
	18) torch.nn.AdavtiveAvgPool3d	It is used to apply a 3D adaptive average pooling over an input signal composed of several input planes.
5.	Padding layers
	1) torch.nn.ReflectionPad1d	It will pad the input tensor using the reflection of the input boundary.
	2) torch.nn.ReflactionPad2d	It will pad the input tensor using the reflection of the input boundary.
	3) torch.nn.ReplicationPad1	It will pad the input tensor using the replication of the input boundary.
	4) torch.nn.ReplicationPad2d	It will pad the input tensor using the replication of the input boundary.
	5) torch.nn.ReplicationPad3d	It will pad the input tensor using the replication of the input boundary.
	6) torch.nn.ZeroPad2d	It will pad the input tensor boundaries with zero.
	7) torch.nn.ConstantPad1d	It will pad the input tensor boundaries with a constant value.
	8) torch.nn.ConstantPad2d	It will pad the input tensor boundaries with a constant value.
	9) torch.nn.ConstantPad3d	It will pad the input tensor boundaries with a constant value.
6.	Non-linear activations (weighted sum, non-linearity)
	1) torch.nn.ELU	It will use to apply the element-wise function: ELU(x)=max(0,x)+min(0,α*(exp(x)-1))
	2) torch.nn.Hardshrink	It will use to apply the hard shrinkage function element-wise function:
	3) torch.nn.LeakyReLU	It will use to apply the element-wise function: LeakyReLu(x)=max(0,x) +negative_slope*min(0,x)
	4) torch.nn.LogSigmoid	It will use to apply the element-wise function:
	5) torch.nn.MultiheadAttention	It is used to allow the model to attend to information from different representation subspaces
	6) torch.nn.PReLU	It will be used to apply the element-wise function: PReLU(x)=max(0,x)+a*min(0,x)
	7) torch.nn.ReLU	It will use to apply the rectified linear unit function element-wise: ReLU(x)=max(0,x)
	8) torch.nn.ReLU6	It will be used to apply the element-wise function: ReLU6(x)=min(max(0,x),6)
	9) torch.nn.RReLU	It will use to apply the randomized leaky rectified linear unit function, element-wise, as described in the paper:
	10) torch.nn.SELU	It will use to apply the element-wise function as: SELU(x)=scale*(max(0,x)+ min(0,a*(exp(x)-1))) Here α= 1.6732632423543772848170429916717 and scale = 1.0507009873554804934193349852946.
	11) torch.nn.CELU	It will use to apply the element-wise function as:
	12) torch.nn.Sigmoid	It will use to apply the element-wise function as:
	13) torch.nn.Softplus	It will use to apply the element-wise function as:
	14) torch.nn.Softshrink	It will use to apply soft shrinkage function elementwise as:
	15) torch.nn.Softsign	It will use to apply the element-wise function as:
	16) torch.nn.Tanh	It will use to apply the element-wise function as:
	17) torch.nn.Tanhshrink	It will use to apply the element-wise function as: Tanhshrink(x)=x-Tanh(x)
	18) torch.nn.Threshold	It will use to thresholds each element of the input Tensor. Threshold is defined as:
7.	Non-linear activations (other)
	1) torch.nn.Softmin	It is used to apply the softmin function to an n-dimensional input Tensor to rescaling them. After that, the elements of the n-dimensional output Tensor lies in the range 0, 1, and sum to 1. Softmin is defined as:
	2) torch.nn.Softmax	It is used to apply the softmax function to an n-dimensional input Tensor to rescaling them. After that, the elements of the n-dimensional output Tensor lies in the range 0, 1, and sum to 1. Softmax is defined as:
	3) torch.nn.Softmax2d	It is used to apply SoftMax over features to each spatial location.
	4) torch.nn.LogSoftmax	It is used to apply LogSoftmax function to an n-dimensional input Tensor. The LofSoftmax function can be defined as:
	5) torch.nn.AdaptiveLogSoftmaxWithLoss	It is a strategy for training models with large output spaces. It is very effective when the label distribution is highly imbalanced
8.	Normalization layers
	1) torch.nn.BatchNorm1d	It is used to apply batch normalization over a 2D or 3D inputs.
	2) torch.nn.BatchNorm2d	It is used to apply batch normalization over a 4D.
	3) torch.nn.BatchNorm3d	It is used to apply batch normalization over 5D inputs.
	4) torch.nn.GroupNorm	It is used to apply group normalization over a mini-batch of inputs.
	5) torch.nn.SyncBatchNorm	It is used to apply batch normalization over n-dimensional inputs.
	6) torch.nn.InstanceNorm1d	It is used to apply an instance normalization over a 3D input.
	7) torch.nn.InstanceNorm2d	It is used to apply an instance normalization over a 4D input.
	8) torch.nn.InstanceNorm3d	It is used to apply an instance normalization over a 5D input.
	9) torch.nn.LayerNorm	It is used to apply layer normalization over a mini-batch of inputs.
	10) torch.nn.LocalResponseNorm	It is used to apply local response normalization over an input signal which is composed of several input planes, where the channel occupies the second dimension.
9.	Recurrent layers
	1) torch.nn.RNN	It is used to apply a multi-layer Elman RNN with tanh or ReLU non-linearity to an input sequence. Each layer computes the following function for each element in the input sequence: ht=tanh(Wih xt+bih+Whh tt-1+bhh)
	2) torch.nn.LSTM	It is used to apply a multi-layer long short-term memory (LSTM) RNN to an input sequence. Each layer computes the following function for each element in the input sequence:
	3) torch.nn.GRU	It is used to apply a multi-layer gated recurrent unit (GRU) RNN to an input sequence. Each layer computes the following function for each element in the input sequence:
	4) torch.nn.RNNCell	It is used to apply an Elman RNN cell with tanh or ReLU non-linearity to an input sequence. Each layer computes the following function for each element in the input sequence: h’=tanh(Wih x+bih+Whh h+bhh) ReLU is used in place of tanh
	5) torch.nn.LSTMCell	It is used to apply a long short-term memory (LSTM) cell to an input sequence. Each layer computes the following function for each element in the input sequence: Where σ is the sigmoid function, and * is the Hadamard product.
	6) torch.nn.GRUCell	It is used to apply a gated recurrent unit (GRU) cell to an input sequence. Each layer computes the following function for each element in the input sequence:
10.	Linear layers
	1) torch.nn.Identity	It is a placeholder identity operator which is argument-insensitive.
	2) torch.nn.Linear	It is used to apply a linear transformation to the incoming data: y=xAT+b
	3) torch.nn.Bilinear	It is used to apply a bilinear transformation to the incoming data: y=x1 Ax2+b
11.	Dropout layers
	1) torch.nn.Dropout	It is used for regularization and prevention of co-adaptation of neurons. A factor ofduring training scales the output. That means the module computes an identity function during the evaluation.
	2) torch.nn.Dropout2d	If adjacent pixels within feature maps are correlated, then torch.nn.Dropout will not regularize the activations, and it will decrease the effective learning rate. In this case, torch.nn.Dropout2d() is used to promote independence between feature maps.
	3) torch.nn.Dropout3d	If adjacent pixels within feature maps are correlated, then torch.nn.Dropout will not regularize the activations, and it will decrease the effective learning rate. In this case, torch.nn.Dropout2d () is used to promote independence between feature maps.
	4) torch.nn.AlphaDropout	It is used to apply Alpha Dropout over the input. Alpha Dropout is a type of Dropout which maintains the self-normalizing property.
12.	Sparse layers
	1) torch.nn.Embedding	It is used to store word embedding’s and retrieve them using indices. The input for the module is a list of indices, and the output is the corresponding word embedding.
	2) torch.nn.EmbeddingBag	It is used to compute sums or mean of ‘bags’ of embedding without instantiating the Intermediate embedding.
13.	Distance Function
	1) torch.nn.CosineSimilarity	It will return the cosine similarity between x1 and x2, computed along dim.
	2) torch.nn.PairwiseDistance	It computes the batch-wise pairwise distance between vectors v1, v2 using the p-norm:
14.	Loss function
	1) torch.nn.L1Loss	It is used to a criterion which measures the mean absolute error between each element in the input x and target y. The unreduced loss can be described as:       l(x,y)=L={l1,…,ln },ln=|xn-yn |, Where N is the batch size.
	2) torch.nn.MSELoss	It is used to a criterion which measures the mean squared error between each element in the input x and target y. The unreduced loss can be described as: l(x,y)=L={l1,…,ln },ln=(xn-yn)2, Where N is the batch size.
	3) torch.nn.CrossEntropyLoss	This criterion combines nn.LogSoftmax() and nn.NLLLoss() in one single class. It is helpful when we train a classification problem with C classes.
	4) torch.nn.CTCLoss	The Connectionist Temporal Classification loss calculates loss between a continuous time series and a target sequence.
	5) torch.nn.NLLLoss	The Negative Log-Likelihood loss is used to train a classification problem with C classes.
	6) torch.nn.PoissonNLLLoss	The Negative log-likelihood loss with the Poisson distribution of t target~Poisson(input)loss(input,target)=input-target*log(target!)he target.
	7) torch.nn.KLDivLoss	It is a useful distance measure for continuous distribution, and it is also useful when we perform direct regression over the space of continuous output distribution.
	8) torch.nn.BCELoss	It is used to create a criterion which measures the Binary Cross Entropy between the target and the output. The unreduced loss can be described as: l(x,y)=L={l1,…,ln },ln=-wn [yn*logxn+ (1-yn )*log(1-xn)], Where N is the batch size.
	9) torch.nn.BCEWithLogitsLoss	It combines a Sigmoid layer and the BCELoss in one single class. We can take advantage of the log-sum-exp trick for numerical stability by combining the operation into one layer.
	10) torch.nn.MarginRankingLoss	It creates a criterion which measures the loss of given inputs x1, x2, two 1D mini-batch Tensors, and a label 1D mini-batch tensor y which contain 1 or -1. The loss function for each sample in the mini-batch is as follows:       loss(x,y)=max(0,-y*(x1-x2 )+margin
	11) torch.nn.HingeEmbeddingLoss	HingeEmbeddingLoss measures the loss of given an input tensor x and a labels tensor y which contain 1 or -1. It is used for measuring whether two inputs are similar or dissimilar. The loss function is defined as:
	12) torch.nn.MultiLabelMarginLoss	It is used to create a criterion which optimizes a multi-class multi-classification hinge loss between input x and output y.
	13) torch.nn.SmoothL1Loss	It is used to create a criterion which uses a squared term if the absolute element-wise error falls below 1 and an L1 term otherwise. It is also known as Huber loss:
	14) torch.nn.SoftMarginLoss	It is used to create a criterion which optimizes the two-class classification logistic loss between input tensor x and target tensor y which contain 1 or -1.
	15) torch.nn.MultiLabelSoftMarginLoss	It is used to create a criterion which optimizes the multi-label one-versus-all loss based on max-entropy between input x and target y of size (N, C).
	16) torch.nn.CosineEmbeddingLoss	It is used to create a criterion which measures the loss of given input tensors x1, x2 and a tensor label y with values 1 or -1. It is used for measuring whether two inputs are similar or dissimilar, using the cosine distance.
	17) torch.nn.MultiMarginLoss	It is used to create a criterion which optimizes a multi-class classification hinge loss between input x and output y.
	18) torch.nn.TripletMarginLoss	It is used to create a criterion which measures the triplet loss of given an input tensors x1, x2, x3 and a margin with a value greater than 0. It is used for measuring a relative similarity between samples. A triplet is composed of an anchor, positive example, and a negative example. L(a,p,n)=max{d(ai,pi )-d(ai,ni )+margin,0}
15.	Vision layers
	1) torch.nn.PixelShuffle	It is used to re-arrange the elements in a tensor of shape(*,C×r2,H,W) to a tensor of shape (*,C,H×r,W,r)
	2) torch.nn.Upsample	It is used to upsample a given multi-channel 1D, 2D or 3D data.
	3) torch.nn.upsamplingNearest2d	It is used to apply 2D nearest neighbor upsampling to an input signal which is composed with multiple input channel.
	4) torch.nn.UpsamplingBilinear2d	It is used to apply 2D bilinear upsampling to an input signal which is composed with, multiple input channel.
16.	DataParallel layers(multi-GPU, distributed)
	1) torch.nn.DataParallel	It is used to implement data parallelism at the module level.
	2) torch.nn.DistributedDataParallel	It is used to implement distributed data parallelism, which is based on the torch.distributed package at the module level.
	3) torch.nn.DistributedDataParallelCPU	It is used to implement distributed data parallelism for the CPU at the module level.
17.	Utilities
	1) torch.nn.clip_grad_norm_	It is used to clip the gradient norm of an iterable of parameters.
	2) torch.nn.clip_grad_value_	It is used to clip the gradient norm of an iterable of parameters at the specified value.
	3) torch.nn.parameters_to_vector	It is used to convert parameters to one vector.
	4) torch.nn.vector_to_parameters	It is used to convert one vector to the parameters.
	5) torch.nn.weight_norm	It is used to apply weight normalization to a parameter in the given module.
	6) torch.nn.remove_weight_norm	It is used to remove the weight normalization and re-parameterization from a module.
	7) torch.nn.spectral_norm	It is used to apply spectral normalization to a parameter in the given module.
	8) torch.nn.PackedSequence	It will use to hold the data and list of batch_sizes of a packed sequence.
	9) torch.nn.pack_padded_sequence	It is used to pack a Tensor containing padded sequences of variable length.
	10) torch.nn.pad_packed_sequence	It is used to pads a packed batch of variable-length sequences.
	11) torch.nn.pad_sequence	It is used to pad a list of variable length Tensors with padding value.
	12) torch.nn.pack_sequence	It is used to packs a list of variable length Tensors
	13) torch.nn.remove_spectral_norm	It is used to removes the spectral normalization and re-parameterization from a module.

参考：

https://pytorch.org/docs/stable/nn.html