Error: RNN Expected input Batch_size

1

Hello everyone!
first of all this forum helped me so much in the past few days… Thank you very much for all the good posts and answers!

Now I have a problem I couldn’t find in a post…

I’m currently working on a RNN network for classification of images. I tried CNN but the images are signals not “normal” Images. I need to track them over time. (but that problem comes later)

the Images/Signals are 128x64

# init
num_classes = 2  
num_epochs = 10
batch_size = 4
learning_rate = 0.001

# input size and sequence length are the Image dimension
input_size = 128 # tried both ways same error
sequence_length = 64
hidden_size = 128
num_layers = 2

the data is loaded over

trainset = torchvision.datasets.ImageFolder(root="/PATH",
                                            transform=transforms.ToTensor())
validation = torchvision.datasets.ImageFolder(root="/PATH",
                                              transform=transforms.ToTensor())

trainloader = torch.utils.data.DataLoader(dataset = trainset,
                                          batch_size = batch_size,
                                          shuffle=False)

validationloader = torch.utils.data.DataLoader(dataset = validation,
                                               batch_size = batch_size,
                                               shuffle=False)

My class

class RNN(nn.Module):
    def __init__(self, input_size, hidden_size, num_layers, num_classes):
        super(RNN, self).__init__()
        self.num_layers = num_layers
        self.hidden_size = hidden_size
        self.rnn = nn.RNN(input_size, hidden_size, num_layers, batch_first=True)
        # -> x : (batch_size, seq, input_size)

        self.fc = nn.Linear(hidden_size, num_classes)


    def forward(self, x):
        h0 = torch.zeros(self.num_layers, x.size(0), self.hidden_size).to(device)
        print(x.size(0)) # outputs: 12

        # Forward propagate RNN
        out, _ = self.rnn(x, h0)
       

        # out: tensor of shape (batch_size, seq_length, hidden_size)
        print(x.shape) # outputs: torch.Size([12, 64, 128])

        out = out[:, -1, :]

        out = self.fc(out)
        return out

the Error i’m getting is:

ValueError: Expected input batch_size (12) to match target batch_size (4).

I know the problem is that the size from x is 12 and the batch_size is 4… but if i change batch_size to 12 the Error still remains just with 36… Somehow my batch_size is 3 times to big for the input (if I see this right?)

training loop

net = RNN(input_size, hidden_size, num_layers, num_classes).to(device)

criterion = nn.CrossEntropyLoss()
optimizer = optim.Adam(net.parameters(), lr=learning_rate)

# train the model
for epoch in range(10):  # loop over the dataset multiple times
    print("epoche: ", epoch)
    running_loss = 0.0
    for i, (images, labels) in enumerate(trainloader):
        # origin shape: [N, 1, 64, 128]
        # resized: [N, 64, 128]
        images = images.reshape(-1, sequence_length, input_size).to(device)
        labels = labels.to(device)

        # zero the parameter gradients
        optimizer.zero_grad()

        # forward + backward + optimize
        outputs = net(images)
        loss = criterion(outputs, labels)
        loss.backward()
        optimizer.step()

        # print statistics
        running_loss += loss.item()
        if i % 100 == 99:    # print every 200 mini-batches
            print('[%d, %5d] loss: %.3f' % (epoch + 1, i + 1, running_loss / 2000))
            running_loss = 0.0

print('Finished Training')

EDIT: If I try the whole RNN with a MNIST (input_size: 28 / sequence_length: 28)dataset it works just fine (28x28)

Thanks all!

2

I cannot reproduce the issue using your code snippet and get a valid output with a batch size of 4, which should be expected.
I’m using this code snippet based on the posted code:

# init
num_classes = 2  
num_epochs = 10
batch_size = 4
learning_rate = 0.001

# input size and sequence length are the Image dimension
input_size = 128 # tried both ways same error
sequence_length = 64
hidden_size = 128
num_layers = 2

class RNN(nn.Module):
    def __init__(self, input_size, hidden_size, num_layers, num_classes):
        super(RNN, self).__init__()
        self.num_layers = num_layers
        self.hidden_size = hidden_size
        self.rnn = nn.RNN(input_size, hidden_size, num_layers, batch_first=True)
        # -> x : (batch_size, seq, input_size)

        self.fc = nn.Linear(hidden_size, num_classes)


    def forward(self, x):
        h0 = torch.zeros(self.num_layers, x.size(0), self.hidden_size)
        print(x.size(0)) # outputs: 12

        # Forward propagate RNN
        out, _ = self.rnn(x, h0)
       

        # out: tensor of shape (batch_size, seq_length, hidden_size)
        print(x.shape) # outputs: torch.Size([12, 64, 128])

        out = out[:, -1, :]

        out = self.fc(out)
        return out

net = RNN(input_size, hidden_size, num_layers, num_classes)
# origin shape: [N, 1, 64, 128]
# resized: [N, 64, 128]
images = torch.randn(batch_size, 1, 64, 128)
images = images.reshape(-1, sequence_length, input_size)
print(images.shape)
> torch.Size([4, 64, 128])

outputs = net(images)
print(outputs.shape)
> torch.Size([4, 2])
3

Thank you for your answer and for running the code!
I think the images are somehow the problem… I’ll test more, if I find the error I’ll post it!
Just to be sure does the amount of images i have in the folder matter?

4

Since it’s a shape mismatch I guess that some reshaping operation is wrong in the code.
Could you post the shape of the data you are getting and compare it to my code?

5

Thanks for your answer.
My original error didn’t pop up any more after I upgraded my system. Not a single change in the code.
If I print the shape of the data with batch_size = 1 its [3,64, 128]
with batch_size = 4 the [3, 68, 522] so not exactly 4 times more. And the fact that its a 3 channel not just 1 channel. But that I will try to change directly from the source of the data.

What I’m getting now is the error:

RuntimeError: The size of tensor a (2) must match the size of tensor b (4) at non-singleton dimension 1

when I change the batch_size to 2 i will train (extremely poorly loss starts at 0.011) but at the end of the program I get the same error again with tensor a (6) and tensor b (2)

RuntimeError: The size of tensor a (6) must match the size of tensor b (2) at non-singleton dimension 0