How to visualize fully connected layers as images

1

quick question, how can i visualize aka plt.imshow() the fully connected layers, when i try to do so it says that it has invalid dimensions for images. any help is greatly appreciated

2

What object are you calling plt.imshow() on? If it’s a Tensor, it shouldn’t be a problem. E.g.,

a = torch.Tensor(10, 1).uniform_(0, 1) 
plt.imshow(a)

should work. If it’s a variable, use the .data attribute. E.g.,

b = torch.autograd.Variable(a)
plt.imshow(b.data)

The same should apply if you have a model from the module class, e.g.:

class MyNet(torch.nn.Module):

    def __init__(self, num_features, num_classes):
        super(MyNet, self).__init__()
        self.linear_1 = torch.nn.Linear(num_features, num_hidden_1)
...

# do training
...
 
plt.imshow(model.linear_1.weight.data)

Maybe one thing to check would be calling .size() on your object, maybe it’s missing an axis or has too many axes

2 Likes
3

its a 1*800 that is its size, does that help? ill try the methods that you have put forth, thank you!

4

Hm, imshow should work with 2D arrays, not sure why you are getting that error. It also works with 3D arrays if the last axis has 1, 3, or 4 elements (i.e., color channels).