Neural network learns a constant function

Please find the ipython notebook in this link: https://drive.google.com/file/d/13Wtpd-Ur7tUO3mGpXQU2H08rNNyyqUyf/view?usp=sharing

The input to the model is 4D and output is 1D (regression problem). The code that I have shared is a simplification of the actual network while capturing the bug I am encountering.

Your code has a small issue with the shapes.
If you are using nn.MSELoss the shapes of your output and target should be the same.
However, in your current code the target is missing dim1, which will trigger some unwanted broadcasting inside the loss function.

train_x = torch.stack((torch.linspace(0, 10, 1000), torch.linspace(0.5, 10.5, 1000), torch.linspace(3, 14, 1000), torch.linspace(7.5, 19.5, 1000)),1)
train_y = (train_x[:,3]**2 - train_x[:,2])/10 + torch.cos(train_x[:,0] * (2 * math.pi)) + torch.sin(train_x[:,1] * (2 * math.pi)) + 0.1 * torch.randn(1000)
train_y.unsqueeze_(1)
print(train_x.size())
print(train_y.size())

After fixing this issue, I played around with your code and your model is able to learn an approximation of the curve using:

torch.manual_seed(2809)

class NN_Pred(nn.Module):
    def __init__(self):
        super(NN_Pred, self).__init__()
        self.base = nn.Sequential(
            nn.Linear(4, 64),
            nn.ReLU(),
            nn.Linear(64, 2),
            nn.ReLU()
        )

        self.base_linear = nn.Linear(2, 1)

        self.train()

    def forward(self, inputs):
        x = inputs
        hidden_base = self.base(x)
        return self.base_linear(hidden_base)

The model learned the approximation also using your current architecture, but was quite sensible to the random seed (a lot of solutions just learned the constant horizontal line). This modification is also not very robust, but might be a good starter to further experiment with your code.

Thanks for the help! I clearly overlooked the possibility of different dimensions for output and target. I also realized the reason why learning was not taking place was that the LR was set at 0.1. Changing it to 0.001 leads to stable training.