传递三维输入的PyToch Skorch

Question

我正在尝试在使用skorch的同时在PyTorch中适应一个模型。我的问题是，我的模型使用了一个LSTM层，它需要3d输入，而我不知道如何正确地传递输入。

在向fit()传递2D数组时，对于预期的3d输入，我从PyTorch得到一个错误。如果传递一个3D数组，我会从fit()方法得到长度不一致的错误(这对我来说都很有意义)。

示例代码如下：

import numpy as np
import torch
from torch import nn
import skorch

class lstmNet(nn.Module):

    def __init__(self, input_size, hidden_size, num_layers):
        super(lstmNet, self).__init__()
        self.rnn = nn.LSTM(
            input_size,
            hidden_size,
            num_layers        )
        self.lin = nn.Linear(
            in_features=hidden_size,
            out_features=1
        )

    def forward(self, x):
        print(x.size())
        out, hn = self.rnn(x)
        out = self.lin(out)
        return out


input_feat = 5
hidden_size = 10
lstmLayers = 2

seq = 20
batch = 30
features = 5

net = skorch.NeuralNet(
    module=lstmNet(
        input_size=input_feat,
        hidden_size=hidden_size,
        num_layers=lstmLayers
    ),
    criterion=torch.nn.MSELoss,
    optimizer=torch.optim.SGD,
    lr=0.1,
    max_epochs=10
)

#inputArr2d = np.random.rand(seq * batch, features)
inputArr3d = np.random.rand(seq, batch, features)
print('input:\n {}\nshape: {}'.format(inputArr3d, inputArr3d.shape))
targetArr = np.random.rand((seq * batch))
#print('target:\n {}\nshape: {}'.format(targetArr, targetArr.shape))

net.fit(X=inputArr3d, y=targetArr)

这是调用net.fit(X=inputArr2d, y=targetArr)时的错误

Traceback (most recent call last):
File "C:\Spielplatz\Python\examples\playground.py", line 64, in <module>
net.fit(X=inputArr2d, y=targetArr)
File "C:\ProgramData\Anaconda3\lib\site-packages\skorch\net.py", line 686, in fit
self.partial_fit(X, y, **fit_params)
File "C:\ProgramData\Anaconda3\lib\site-packages\skorch\net.py", line 646, in partial_fit
self.fit_loop(X, y, **fit_params)
File "C:\ProgramData\Anaconda3\lib\site-packages\skorch\net.py", line 584, in fit_loop
step = self.train_step(Xi, yi, **fit_params)
File "C:\ProgramData\Anaconda3\lib\site-packages\skorch\net.py", line 507, in train_step
y_pred = self.infer(Xi, **fit_params)
File "C:\ProgramData\Anaconda3\lib\site-packages\skorch\net.py", line 810, in infer
return self.module_(x, **fit_params)
File "C:\ProgramData\Anaconda3\lib\site-packages\torch\nn\modules\module.py", line 491, in __call__
result = self.forward(*input, **kwargs)
File "C:\Spielplatz\Python\examples\playground.py", line 33, in forward
out, hn = self.rnn(x)
File "C:\ProgramData\Anaconda3\lib\site-packages\torch\nn\modules\module.py", line 491, in __call__
result = self.forward(*input, **kwargs)
File "C:\ProgramData\Anaconda3\lib\site-packages\torch\nn\modules\rnn.py", line 178, in forward
self.check_forward_args(input, hx, batch_sizes)
File "C:\ProgramData\Anaconda3\lib\site-packages\torch\nn\modules\rnn.py", line 126, in check_forward_args
expected_input_dim, input.dim()))
RuntimeError: input must have 3 dimensions, got 2

这是调用net.fit(X=inputArr3d, y=targetArr)时的错误

Traceback (most recent call last):
File "C:\Spielplatz\Python\examples\playground.py", line 64, in <module>
net.fit(X=inputArr3d, y=targetArr)
File "C:\ProgramData\Anaconda3\lib\site-packages\skorch\net.py", line 686, in fit
self.partial_fit(X, y, **fit_params)
File "C:\ProgramData\Anaconda3\lib\site-packages\skorch\net.py", line 646, in partial_fit
self.fit_loop(X, y, **fit_params)
File "C:\ProgramData\Anaconda3\lib\site-packages\skorch\net.py", line 573, in fit_loop
X, y, **fit_params)
File "C:\ProgramData\Anaconda3\lib\site-packages\skorch\net.py", line 1004, in get_split_datasets
dataset = self.get_dataset(X, y)
File "C:\ProgramData\Anaconda3\lib\site-packages\skorch\net.py", line 961, in get_dataset
return dataset(X, y, **kwargs)
File "C:\ProgramData\Anaconda3\lib\site-packages\skorch\dataset.py", line 104, in __init__
raise ValueError("X and y have inconsistent lengths.")
ValueError: X and y have inconsistent lengths.

Answer 1

正如您自己所说，传递2D数组是行不通的，因为您没有时间维度，所以在这种情况下错误消息是正确的(LSTM需要3维)。

配置部分是传递3D数组时的错误消息。正在发生的情况是，您正在将数据格式化为(sequence, batch, feature)，即"batch second“。还有一种称为"batch first“的替代格式，它将数据格式化为(batch, sequence, feature)，这在sklearn中是默认格式(因此在skorch中也是如此)。如果以这种方式重新格式化数据，并将LSTM配置为使用这种格式(传递batch_first=True)，则错误应该会消失：

    self.rnn = nn.LSTM(
        input_size,
        hidden_size,
        num_layers,
        batch_first=True,   
    )
# ...
inputArr3d = np.random.rand(batch, seq, features)

请注意，您的模块还需要将RNN的返回值转换为后续层可以处理的内容(例如，展平时间维度)，否则线性层将不知道如何处理额外的时间维度。