问题描述
安装了pytorch最新版本1.6之后，在pycharm中编辑python代码时，输入torch.nn.看不到提示了，比如torch.nn.MSELoss()。而在1.4及以前的版本中，直接输入torch.nn.就会自动提示出很多torch.nn.modules中的API。
该问题的讨论在前几年有过不少（https://www.zhihu.com/question/279645242/answer/520263986
），但都是基于老版本，经过尝试，对于1.6版本是无效的。

原因分析
pycharm的自动提示是根据第三方包的每个文件夹下的__init__.pyi文件来显示的，只有__init__.pyi中import了的API才会被pycharm自动提示。

首先对pytorch.nn模块要知道，问题描述中提到的MSELoss等众多函数，真实位置是torch.nn.modules.MSELoss()，你直接调用这个真实位置是可以自动提示的。但是1.4及以前的版本中大家都熟悉了直接用nn.MSELoss()这样调用，如何让1.6版本也能像历史版本一样提示呢？

基于此，我对比了1.6和1.4的区别。
在torch 1.6版本包存放位置下，torch/nn/下是有__init__.pyi的，里面有一行from .modules import *，说明nn模块是可以直接调用子模块modules中的API的，所以直接调用nn.MSELoss()不会报错，只是不会自动提示。
然后在进入torch/nn/modules/发现，1.6版本中缺少__init__.pyi文件，所以在pycharm输入nn.的时候并不会提示子模块modules中的API。

解决方案
从pytorch 1.4版本中复制一份__init__.pyi文件到1.6版本的依赖包的相同目录下。具体位置是{你的第三方包存放位置}/Lib/site-packages/torch/nn/modules/init.pyi
然后就可以在pycharm中愉快使用nn.自动提示了。其他模块不自动提示的，解决方法类同。

补充
关于解决方案中第三方包存放位置不知道的，可以在pycharm左侧项目目录结构中看到一项External Libraries，点开它，你就能直接找到Lib/site-packages/torch/nn/modules/，从而不必去资源管理器找。

附件 init.pyi

from .module import Module as Module
from .activation import CELU as CELU, ELU as ELU, GLU as GLU, GELU as GELU, Hardshrink as Hardshrink, \Hardtanh as Hardtanh, LeakyReLU as LeakyReLU, LogSigmoid as LogSigmoid, LogSoftmax as LogSoftmax, PReLU as PReLU, \RReLU as RReLU, ReLU as ReLU, ReLU6 as ReLU6, SELU as SELU, Sigmoid as Sigmoid, Softmax as Softmax, \Softmax2d as Softmax2d, Softmin as Softmin, Softplus as Softplus, Softshrink as Softshrink, Softsign as Softsign, \Tanh as Tanh, Tanhshrink as Tanhshrink, Threshold as Threshold
from .adaptive import AdaptiveLogSoftmaxWithLoss as AdaptiveLogSoftmaxWithLoss
from .batchnorm import BatchNorm1d as BatchNorm1d, BatchNorm2d as BatchNorm2d, BatchNorm3d as BatchNorm3d, \SyncBatchNorm as SyncBatchNorm
from .container import Container as Container, ModuleDict as ModuleDict, ModuleList as ModuleList, \ParameterDict as ParameterDict, ParameterList as ParameterList, Sequential as Sequential
from .conv import Conv1d as Conv1d, Conv2d as Conv2d, Conv3d as Conv3d, ConvTranspose1d as ConvTranspose1d, \ConvTranspose2d as ConvTranspose2d, ConvTranspose3d as ConvTranspose3d
from .distance import CosineSimilarity as CosineSimilarity, PairwiseDistance as PairwiseDistance
from .dropout import AlphaDropout as AlphaDropout, Dropout as Dropout, Dropout2d as Dropout2d, Dropout3d as Dropout3d, \FeatureAlphaDropout as FeatureAlphaDropout
from .fold import Fold as Fold, Unfold as Unfold
from .instancenorm import InstanceNorm1d as InstanceNorm1d, InstanceNorm2d as InstanceNorm2d, \InstanceNorm3d as InstanceNorm3d
from .linear import Bilinear as Bilinear, Identity as Identity, Linear as Linear
from .loss import BCELoss as BCELoss, BCEWithLogitsLoss as BCEWithLogitsLoss, CTCLoss as CTCLoss, \CosineEmbeddingLoss as CosineEmbeddingLoss, CrossEntropyLoss as CrossEntropyLoss, \HingeEmbeddingLoss as HingeEmbeddingLoss, KLDivLoss as KLDivLoss, L1Loss as L1Loss, MSELoss as MSELoss, \MarginRankingLoss as MarginRankingLoss, MultiLabelMarginLoss as MultiLabelMarginLoss, \MultiLabelSoftMarginLoss as MultiLabelSoftMarginLoss, MultiMarginLoss as MultiMarginLoss, NLLLoss as NLLLoss, \NLLLoss2d as NLLLoss2d, PoissonNLLLoss as PoissonNLLLoss, SmoothL1Loss as SmoothL1Loss, \SoftMarginLoss as SoftMarginLoss, TripletMarginLoss as TripletMarginLoss
from .module import Module as Module
from .normalization import CrossMapLRN2d as CrossMapLRN2d, GroupNorm as GroupNorm, LayerNorm as LayerNorm, \LocalResponseNorm as LocalResponseNorm
from .padding import ConstantPad1d as ConstantPad1d, ConstantPad2d as ConstantPad2d, ConstantPad3d as ConstantPad3d, \ReflectionPad1d as ReflectionPad1d, ReflectionPad2d as ReflectionPad2d, ReplicationPad1d as ReplicationPad1d, \ReplicationPad2d as ReplicationPad2d, ReplicationPad3d as ReplicationPad3d, ZeroPad2d as ZeroPad2d
from .pixelshuffle import PixelShuffle as PixelShuffle
from .pooling import AdaptiveAvgPool1d as AdaptiveAvgPool1d, AdaptiveAvgPool2d as AdaptiveAvgPool2d, \AdaptiveAvgPool3d as AdaptiveAvgPool3d, AdaptiveMaxPool1d as AdaptiveMaxPool1d, \AdaptiveMaxPool2d as AdaptiveMaxPool2d, AdaptiveMaxPool3d as AdaptiveMaxPool3d, AvgPool1d as AvgPool1d, \AvgPool2d as AvgPool2d, AvgPool3d as AvgPool3d, FractionalMaxPool2d as FractionalMaxPool2d, \FractionalMaxPool3d as FractionalMaxPool3d, LPPool1d as LPPool1d, LPPool2d as LPPool2d, MaxPool1d as MaxPool1d, \MaxPool2d as MaxPool2d, MaxPool3d as MaxPool3d, MaxUnpool1d as MaxUnpool1d, MaxUnpool2d as MaxUnpool2d, \MaxUnpool3d as MaxUnpool3d
from .rnn import GRU as GRU, GRUCell as GRUCell, LSTM as LSTM, LSTMCell as LSTMCell, RNN as RNN, RNNBase as RNNBase, \RNNCell as RNNCell, RNNCellBase as RNNCellBase
from .sparse import Embedding as Embedding, EmbeddingBag as EmbeddingBag
from .upsampling import Upsample as Upsample, UpsamplingBilinear2d as UpsamplingBilinear2d, \UpsamplingNearest2d as UpsamplingNearest2d

转载：https://blog.csdn.net/winter2121/article/details/108799776