Training epoch and Neural Net class of the tutorial inputs the info to wrong params

1

hi this is my first post, so if not suitable sorry.

basically I am developing a NN that uses images. However, I receive this error:

RuntimeError: Given groups=1, weight[5, 3, 436, 436], so expected input[5, 436, 436, 3] to have 3 channels, but got 436 channels instead

I got that legent of weight[batch,channel,h,w], but input has this data shift.

I copied the sample code from this.

only change I made is just changing self.conv1 = nn.Conv2d(3,6,5) to nn.Conv2d(3, 5, 436)

I can provide more info if necessary. Besides, can I use h!=w input images? thanks!

2

It looks like you provided an image with shapes [batch_size, w, h, channels] instead of [batch_size, channels, h, w].
How did you load your images?

You can use non square images without a problem. If you need a non square kernel, you can provide the kernel_size as a tuple: kernel_size=(kh, kw).

3

yeah after some research, I saw that it is cwh problem. currently I am trying to find the solution to reverse the order.

My code is like this:

    datasetTrain=FencerPicLoader('D:/x/x/x',transformations)
    train_loader=DataLoader(dataset=datasetTrain,
                            batch_size=5,  
                            shuffle=True,   
                            num_workers=2) 

    for epoch in range(2):  # loop over the dataset multiple times

        running_loss = 0.0
        for i, data in enumerate(train_loader, 0):
            # get the inputs
            inputs, labels = data

            # zero the parameter gradients
            optimizer.zero_grad()

            # forward + backward + optimize
            outputs = net(inputs)#####error jumps to Net()




class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.conv1 = nn.Conv2d(3, 5, 436) 
        self.pool = nn.MaxPool2d(2, 2)
        self.conv2 = nn.Conv2d(6, 16, 5) 
        self.fc1 = nn.Linear(16 * 5 * 5, 120)
        self.fc2 = nn.Linear(120, 84)
        self.fc3 = nn.Linear(84, 10)

Thanks for help and kernel advice! :slight_smile:

4

You can just permute the dimensions:

inputs = inputs.permute(0, 3, 2, 1)

or change your Dataset to return the input in the right shape.

1 Like
5

it gave this error:

RuntimeError: thnn_conv2d_forward is not implemented for type torch.ByteTensor

but probably the problem will be solved if I just change the array. :slight_smile:

edit: actually that solved the error. apparently forwarding is a problem… I believe I can solve probably.

6

nn.Conv2d is complaining about your input. Cast your input to input.float() and try it again.

7

Yeah, as I promised I solved that issue before the answer! :slight_smile:

ok last question (so so sorry I didnt want to), but before that: is there a good website that I can learn these kind of stuff easily? I think I will encounter more of them in the future and want to be less dependent on the pytorch forums etc. I checked stanford ML courses and fast.ai courses, but they hardly ever mentions these inner structures.

RuntimeError: Given input size: (5x1x1). Calculated output size: (5x0x0). Output size is too small at c:\programdata\miniconda3\conda-bld\pytorch_1524546371102\work\aten\src\thnn\generic/SpatialDilatedMaxPooling.c:67

so I got this error.

In conv2d(3,5,436), I get this error. If I say (3,5,218 or 109) I get RAM error like 71GB needed etc. If I write random number like 50, it works but it heavily slows the down computer.

THANKS