Negative output from neural network?

#1

I’m currently trying to build a network for Human Activity Recognition using pyTorch and DSADS dataset.

This is my neural network definition.

import torch.nn as nn

class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.relu = nn.ReLU()
        self.pool = nn.MaxPool2d((1, 2), stride=2)

        self.conv1 = nn.Conv2d(6,16, (1, 6))
        self.conv2 = nn.Conv2d(16,32, (1, 6))

        self.fc1 = nn.Linear(32 * 1 * 3, 70)
        self.fc2 = nn.Linear(70, 30)
        self.fc3 = nn.Linear(30, 19)

    def forward(self, x):
        x = self.conv1(x)
        x = self.relu(x)
        x = self.pool(x)
        x = self.conv2(x)
        x = self.relu(x)
        x = self.pool(x)
        x = x.view(x.size()[0], -1)
        x = self.fc1(x)
        x = self.relu(x)
        x = self.fc2(x)
        x = self.relu(x)
        x = self.fc3(x)
        return x

criterion = nn.CrossEntropyLoss()
optimizer = optim.SGD(net.parameters(), lr=0.0001, momentum=0.9, weight_decay=0.005)

This is how I load train and test data.

class data_loader(Dataset):
    def __init__(self, samples, labels, t):
        self.samples = samples
        self.labels = labels
        self.T = t
        
    def __getitem__(self, index):
        sample, target = self.samples[index], self.labels[index]
        if self.T:
            return self.T(sample), target
        else:
            return sample, target

    def __len__(self):
        return len(self.samples)

# train_data_reshaped: numpy array of dimension (57000, 6, 1, 30)
# test_data_reshaped: numpy array of dimension (22800, 6, 1, 30)
train_set = data_loader(train_data_reshaped, train_labels, transform)
    test_set = data_loader(test_data_reshaped, test_labels, transform)
    train_loader = DataLoader(train_set, batch_size=64, shuffle=True, drop_last=True)
    test_loader = DataLoader(test_set, batch_size=64, shuffle=False)

Training and testing:

for epoch in range(30):
    for (inputs, labels) in train_loader:
        inputs, labels = inputs.to(device).float(), labels.to(device)
        inputs = inputs.view(-1, 6, 1, 30)
        labels = labels[:, -1].long()
        labels = labels - 1        
        optimizer.zero_grad()
        outputs = net(inputs)
        print("outputs: ", outputs)
        loss = criterion(outputs, labels)
        loss.backward()
        optimizer.step()

    for (inputs, labels) in test_loader:
        inputs, labels = inputs.to(device).float(), labels.to(device)
        inputs = inputs.view(-1, 6, 1, 30)        
        labels = labels[:, -1].long()
        labels = labels - 1
        optimizer.zero_grad()
        outputs = net(inputs)
        loss = criterion(outputs, labels)

I expect the net will produce non-negative outputs each of them represents probability. However, my net produces negative output like this:

outputs:  tensor([[-0.2586, -0.0550,  0.2156,  ...,  0.0132, -0.1516,  0.0052],
        [-0.2586, -0.0952,  0.2352,  ...,  0.0199, -0.1925,  0.0107],
        [-0.4870, -0.0813,  0.4605,  ...,  0.0353, -0.1569,  0.0322],
        ...,
        [-0.2445, -0.1426,  0.1625,  ...,  0.0872, -0.1863,  0.0826],
        [-0.3659, -0.1266,  0.3297,  ...,  0.0860, -0.1500,  0.0035],
        [-0.3237, -0.0675,  0.2883,  ...,  0.0384, -0.2692,  0.0562]],
       grad_fn=<AddmmBackward>)

Any idea of what I did wrong? Thanks in advance.

#2

It’s okay that the net produces negative output, to get the actual prediction you can use argmax, or if you prefer the probabilities use softmax after net output

#3

Thank you for your reply. However, my accuracy and loss did not change much in the whole learning process. I have tried changing the optimizer to Adam optimizer, learning rate, but the accuracy and loss still do not change.
Any idea where I can start inspecting?

This is how I calculated accuracy and loss

for (inputs, labels) in test_loader:
        inputs, labels = inputs.to(device).float(), labels.to(device)
        inputs = inputs.view(-1, 6, 1, 30)        
        labels = labels[:, -1].long()
        labels = labels - 1
        optimizer.zero_grad()
        outputs = net(inputs)
        loss = criterion(outputs, labels)
        sum_loss += loss.item()
        predicted = torch.argmax(outputs, dim=1)
        sum_total += labels.size(0)
        sum_correct += (predicted == labels).sum().item()
        print("test  mean loss={}, accuracy={}"
            .format(sum_loss*BATCH_SIZE/len(test_loader.dataset), float(sum_correct/sum_total)))
    test_loss_value.append(sum_loss*BATCH_SIZE/len(test_loader.dataset))
    test_acc_value.append(float(sum_correct/sum_total))

Accuracy:

Screen Shot 2020-10-13 at 19.21.04
Screen Shot 2020-10-13 at 19.21.04828×778 32 KB

#4

Loss looks like this:

Screen Shot 2020-10-13 at 20.02.35
Screen Shot 2020-10-13 at 20.02.35828×778 34.5 KB

#5

Try to scale down the use case and overfit a small dataset, e.g. just 10 samples by playing around with the hyperparameters.
If that doesn’t work at all there might be an issue in the code we haven’t seen yet.

#6

Hi @SarahTeoh. I’m facing a problem almost exactly like the one you have described here. Have you found out what was wrong?