Linear layer input neurons number calculation after conv2d

vision
1

I will be really thankful to those who can explain me this. I know the formula for calculation but after some iterations, I haven’t got to the answer yet.

The formula for output neuron:
Output = ((I-K+2P)/S + 1), where
I - a size of input neuron,
K - kernel size,
P - padding,
S - stride.

Input tensor shape:

torch.Size([36, 200, 150, 3])

I have the following model:

class Flatten(torch.nn.Module):
    def forward(self, x):
        return x.view(x.size()[0], -1)

model = torch.nn.Sequential(
        torch.nn.Conv2d(32, 64, kernel_size=(3, 3)),
        torch.nn.ReLU(),
        torch.nn.Conv2d(64, 128, kernel_size=(3, 3)),
        torch.nn.ReLU(),
        torch.nn.MaxPool2d(kernel_size=(2, 2)),
        torch.nn.Dropout(0.25),
        Flatten(),
        torch.nn.Linear(128, 128),
        torch.nn.ReLU(),
        torch.nn.Linear(128, 2),
        torch.nn.Softmax()
        )

I can’t calculate the number of neurons in Linear layer.
Could you kindly help and explain me the calculations of Linear input neuron size behind this network?

1 Like
2

Your input shape seems to be a bit wrong, as it looks like the channels are in the last dimension.
In PyTorch, image data is expected to have the shape [batch_size, channel, height, width].
Based on your shape, I guess 36 is the batch_size, while 3 seems to be the number channels.

However, as your model expects 32 input channels, your input won’t work at all currently.

Let’s just assume we are using an input of [1, 32, 200, 150] and walk through the model and the shapes.
Since your nn.Conv2d layers don’t use padding and a default stride of 1, your activation will lose one pixel in both spatial dimensions.
After the first conv layer your activation will be [1, 64, 198, 148], after the second [1, 128, 196, 146].
nnMaxPool2d(2) will halve the activation to [1, 128, 98, 73].

If you set the number of in_features for the first linear layer to 128*98*73 your model will work for my input.

I also recommend to just print out the shape of your activation before the linear layer, if the shape calculation is too cumbersome, and set the input features according to this.
For your Sequential model you can just create a print layer with:

class Print(nn.Module):
    def forward(self, x):
        print(x.size())
        return x
11 Likes
3

Thank you very much ptrblck. It was really helpful.
It seems that I manage to make a function out off it for those who struggle with the issue.

def count_input_neuron(model, image_dim):
    return model(torch.rand(1, *(image_dim))).data.view(1, -1).size(1)

So, I have to change it to 3? The number of features or number of color channels?

4

The in_channels of the first conv layer correspond to the channels of your input.
In case you are using a color image tensor, i.e. 3 channels, you would have to set in_channels=3.

1 Like
5

First of all there is a problem with your input shape. The shape should be BATCH_SIZE * CHANNEL * HEIGHT *WIDTH. So lets correct your size and I assume you BATCH_SIZE = 36, CHANNEL = 3, HEIGHT = 200 , WIDTH = 150.

images = image.permute(0,3,1,2)

Next lets change your first Conv2d code. IT should be

torch.nn.Conv2d(3, 64, kernel_size=(3, 3))

So after the first convolution using your formular, we will have

[3, 64, 198, 148]

After the second Conv2d operation, we will have

[3, 128, 196, 146].

The maxpooling which halves the activations we will have

[3, 128, 98, 73]

And finally the input of the fully connected layer will be 128×98×73 = 915712

2 Likes
6

Hi, I have some questions here:
"
So after the first convolution using your formular, we will have

[3, 64, 198, 148]

After the second Conv2d operation, we will have

[3, 128, 196, 146].

"
Since we assume batch_size=36, maybe it should be that after the first convolution layer, we will have

[36, 64, 198, 148].

And after the second convolution layer, we will have

[36, 128, 196, 146].
7

First of all welcome to the Pytorch Community Rui_Li :blush:
The first dimension of Pytorch Convolution should always be the the number of channels (3) for the input image , While on the other and the first dimension of the inputed image should be the batch_size(36). So do not let this confuse you . Furthermore the convolution operation will be done 36 number of times for all of the 36(batch size) images in parallel.

8

Speaking about the topic of calculating the number of neurons for your linear layer input, what about the rest of your linear layers? Is there a specific theory or formula we can use to determine the number of layers to use and the number to put for our input and output for the linear layers?

I 've read quite abit online for this problem, but it seems like most people does not have a specific solution or answer to this. It is mostly based on empirical methods and taking that as a reference to increase or decrease layers. What are your thoughts about this? Thank you!

9

I would try to start with a known good baseline and try to adapt the architecture to your use case.
Also, the number of neurons defines e.g. if your model has a bottleneck, which might be useful for certain use cases.
That being said, I think a lot of architectures are developed empirically and the reasoning about why the model is working find is created afterwards. :wink:

10

You could also use torch.nn.LazyLinear, a module where in_features is inferred.
Please refer to Inferring shape via flatten operator - #20 by iacob

2 Likes