Error in Binary Classification in NeuralNetBinaryClassifier.

[kaihhe]

When I use BinaryClassifer, and there is always an error ValueError: The target data should be 1-dimensional..

I change my target data into 1-dimensional with array.ravel(), it show ValueError: Target size (torch.Size([128])) must be the same as input size (torch.Size([128, 1])).

When I use set criterion=nn.BCEWithLogitsLoss, it run successfully.

I thinks there is wrong in classifier.check_data().

[BenjaminBossan]

Can you tell us what criterion you are using if not BCEWithLogitsLoss?

As a workaround, you could override check_data and see if that works.

[kaihhe]

Yes I use BCEWithLogitsLoss. I use BCELoss and it works. I wish you could fix that problem about BCEWithLogitsLoss.

[BenjaminBossan]

I cannot reproduce the error. Can you please post what kind of data you use, i.e.

• type
• dtype
• shape

of X and y? Also, the code for the module would help.

[kaihhe]

I write a case for the error.

import numpy as np
import torch as torch
import torch.nn as nn
import torch.nn.functional as F
from skorch import NeuralNetBinaryClassifier
torch.set_default_dtype(torch.float64)


class MyNet(nn.Module):
    def __init__(self):
        super(MyNet, self).__init__()
        self.linear = nn.Linear(100, 50)
        self.output = nn.Linear(50, 1)

    def forward(self, input):
        out = input
        out = self.linear(out)
        out = F.relu(out)
        out = self.output(out)
        return out


X = np.random.normal(size=(200, 100))
y = np.zeros((200,1))
y[100:] = 1
net = NeuralNetBinaryClassifier(MyNet)
net.fit(X, y) # ValueError: The target data should be 1-dimensional.
net.fit(X, y.ravel()) # ValueError: X and y have inconsistent lengths.

I change the loss into BCELoss and add sigmoid function in MyNet. It works.

out = torch.sigmoid(out)

net = NeuralNetBinaryClassifier(MyNet, criterion=nn.BCELoss)
net.fit(X, y.ravel())

[thomasjpfan]

The output of the Module needs to be 1-D:

import torch.nn as nn
import numpy as np
from skorch import NeuralNetBinaryClassifier
from torch.nn.functional import relu

class MyNet(nn.Module):
    def __init__(self):
        super(MyNet, self).__init__()
        self.linear = nn.Linear(100, 50)
        self.output = nn.Linear(50, 1)

    def forward(self, input):
        out = input
        out = self.linear(out)
        out = relu(out)
        out = self.output(out)
        return out.squeeze(-1)


X = np.random.normal(size=(200, 100)).astype('float32')
y = np.zeros(200).astype('float32')
y[:100] = 1
net = NeuralNetBinaryClassifier(MyNet)

[kaihhe]

Ok, thank you.