RuntimeError: mat1 and mat2 shapes cannot be multiplied (4x21632 and 1152x120)

vision
1

I ammfacing this error from a long time, anyone can tell where is the problem attached the code below
class NetModel(nn.Module):

def __init__(self):

    super().__init__()

    self.conv_1 = nn.Conv2d(3, 32, 3, padding=1)

    self.conv_2 = nn.Conv2d(32, 64, 3)

    self.maxpool_1 = nn.MaxPool2d(2,2)

    self.conv_3 = nn.Conv2d(64, 128, 3)

    self.bn = nn.BatchNorm2d(128)

    self.fc1 = nn.Linear(128*3*3, 120)

    self.fc2 = nn.Linear(120, 84)

    self.fc3 = nn.Linear(84, 10)

def forward(self, x):

    x = F.relu(self.conv_1(x))

    x = F.relu(self.conv_2(x))

    x = self.maxpool_1(x)

    x = F.relu(self.conv_3(x))

    x = self.bn(x)

    x = torch.flatten(x, 1)

    x = F.relu(self.fc1(x))

    x = F.relu(self.fc2(x))

    x = self.fc3(x)

    return x
2

Based on the error message, I guess the activation input to self.fc1 has 21632 features, while self.fc1 is using in_features=1152.
Use:

self.fc1 = nn.Linear(21632, 120)

and it should work.

3

Did you calculate dimensions of your convnet? Because it’s probably wrong.

General formula with input vector dimensions = HxWxC is floor((W−K+2*P) / S) + 1
P - padding
S- stride
K- kernel size
If W <> H then calculations need to be done for both dimensions: Height and Width.

Plus there is 2x2 maxpool which will reduce the size of each dimension twice.

What is the dimension of your input?

4

The dimension of the input image os 32x32x3

5

What you need to do is to flatten your data after the last convolutional layer - look at the script below - and also correct the dimension of your first linear layer

import torch
from torch.nn import Conv2d, MaxPool2d, BatchNorm2d, Linear

input = torch.rand((1, 3, 32, 32)) # Batch size, number of channels, Heigth, Width

conv_1 = Conv2d(3, 32, 3, padding=1)
t1 = conv_1(input)
t1.size()
>>> torch.Size([1, 32, 32, 32])

conv_2 = Conv2d(32, 64, 3)
t2 = conv_2(t1)
t2.size()
>>> torch.Size([1, 64, 30, 30])

maxpool_1 = MaxPool2d(2,2)
t3 = maxpool_1(t2)
t3.size()
>>> torch.Size([1, 64, 15, 15])

conv_3 = Conv2d(64, 128, 3)
t4 = conv_3(t3)
t4.size()
>>> torch.Size([1, 128, 13, 13])

bn = BatchNorm2d(128) # doesn't change the size
t5 = bn(t4)
t5.size()
>>> torch.Size([1, 128, 13, 13])

t6 = t5.reshape(1, 128*13*13) # flat the output for the input to nn
t6.size()
>>> torch.Size([1, 21632])

fc1 = Linear(128*13*13, 120)
t7 = fc1(t6)
# Looks fine :-)