Mat1 and mat2 shapes cannot be multiplied (256x320 and 640x8)

astra1234567 · October 24, 2022, 2:17pm

Hi ! I have problem with my CNN model. I don’t understand why my model works for summary when I give (1, 64, 16) and why it doesn’t work for real input ( it is batch, so it is tensor of inputs (256, 1, 64, 16)). I use torch.utils.data.DataLoader.
Summary:

My model:

Error:
RuntimeError: mat1 and mat2 shapes cannot be multiplied (256x320 and 640x8)
and it goes ahead in this line: self.fc1 = nn.Linear(64 * 5 * 2 , 8)

Can someone help me with this? Could you explain me, how solve this problem ? Thank you !

InnovArul · October 25, 2022, 6:13am

It doesn’t raise any error for me in pytorch 1.12.1. Not sure why it didn’t work for you.

gist.github.com

https://gist.github.com/InnovArul/7af1684ccc50fa245954da2cb4a5eca1

cnn_sample.py

from turtle import forward
import torch
import torch.nn as nn, torch.nn.functional as F

class CNN(nn.Module):
    def __init__(self) -> None:
        super().__init__()

        self.conv1 = nn.Conv2d(in_channels=1, out_channels=8, kernel_size=6, stride=1, padding=2)
        self.RL1 = nn.ReLU()

This file has been truncated. show original

P.S: If you paste the code directly here instead of an image, it’s easier for others to execute the code and you might get a faster reply

astra1234567 · October 25, 2022, 7:29pm

@InnovArul Thank you for your help. I found mistake. It was in pre-processing data. In your example, you used d = torch.randn(256, 1, 64, 16) but what if I will use d = torch.randn(1, 64, 16), I know it will be probably error. It is possible use single tensor (1, 64, 16) if I trained model with batch (256, 1, 64, 16) ?

InnovArul · October 26, 2022, 4:47am

The nn modules are written to work on batches of data. i.e., they expect a batch dimension at 0’th position. To pass one data point, you can use batch dimension with size 1. i.e., data of size (1, 1, 64, 16).