RuntimeError: mat1 and mat2 shapes cannot be multiplied (28x28 and 784x128)

vision
1

Hello all,
I am newbie to torch …I am getting this error while passing the new data to the model

"Model "
Sequential(
Linear(in_features=784, out_features=128, bias=True)
ReLU()
Linear(in_features=128, out_features=64, bias=True)
Tanh()
Linear(in_features=64, out_features=10, bias=True)
LogSoftmax(dim=1)
)

2

Hi Vimal!

The first layer of your model is a Linear with in_features = 784. It.
therefore expects to be passed a batch of 1-d vectors of length 784.
(More precisely, a Linear expects a tensor whose last dimension is
in_features and that has any number of leading “batch” dimensions,
including zero. So you can think of this as being a potentially
multidimensional “batch” of 1-d vectors, if you will.)

It looks like you are passing in a image with width = height = 28.
You need to flatten() your 28x28 image into a length-784 vector (or,
more generally, flatten() the last two dimensions of your batch of
images) before passing it to your Linear.

Note, you can include the flattening operation as a “layer” in your
model by using the Module (class or function-object) version, Flatten.

Consider:

>>> import torch
>>> torch.__version__
'1.9.0'
>>> batch_of_3_images = torch.randn (3, 28, 28)
>>> batch_of_3_images.shape
torch.Size([3, 28, 28])
>>> lin = torch.nn.Linear (in_features = 784, out_features = 128, bias = True)
>>> flat = torch.nn.Flatten (start_dim = -2)
>>> flat_batch = flat (batch_of_3_images)
>>> flat_batch.shape
torch.Size([3, 784])
>>> lin (flat_batch).shape
torch.Size([3, 128])

Best.

K. Frank

1 Like
3

Thanks much Frank !..I will try