python - RuntimeError: shape '[784]' is invalid for input of size 7840

Question

Below is the code:

import torch

import torchvision

from torchvision import transforms, datasets

#Establishing the batch size

Batch_size = 10

#Downloading the train data

train_mnist = datasets.MNIST(root="./data", train=True, download=True,
                   transform=transforms.Compose([transforms.ToTensor()]))

#Passing the train data into a Dataloader

train_set = torch.utils.data.DataLoader(train_mnist, batch_size=Batch_size, shuffle=True)

#Downloading the test data

test_mnist = datasets.MNIST(root="./data", train=True, download=True,
                   transform=transforms.Compose([transforms.ToTensor()]))

#Passing the test data into a Dataloader

test_set = torch.utils.data.DataLoader(test_mnist, batch_size=Batch_size, shuffle=True)

#Building the network

import torch.nn as nn
import torch.nn.functional as f
class Netwk(nn.Module):

def __init__(self):
    super().__init__()
    self.fc1 = nn.Linear(28*28, 64)
    self.fc2 = nn.Linear(64, 64)
    self.fc3 = nn.Linear(64, 64)
    self.fc4 = nn.Linear(64, 10)

def Fpropagation(self, x):
    x = F.relu(self.fc1(x))
    x = F.relu(self.fc2(x))
    x = F.relu(self.fc3(x))
    x = self.fc4(x)

return F.log_softmax(x, dim=1)

return x

net = Netwk()

print(net)

#Creating optimization for the loss

import torch.optim as optim

Optimizer = optim.Adam(net.parameters(), lr=0.001)

EPOCHS = 3
for epoch in range(EPOCHS):
    for data in train_set:
        X, y = data
        net.zero_grad()
        output = net(X.view(28*28))
        loss = F.nll_loss(output, y)
        loss.backward()
        optimizer.step()
        print(loss)

#The error i get

RuntimeError                              Traceback (most recent call last)

<ipython-input-58-b92f3c4f7059> in <module>()
     10     X, y = data
     11     net.zero_grad()
---> 12     output = net(X.view(28*28))
     13     loss = F.nll_loss(output, y)
     14     loss.backward()

RuntimeError: shape '[784]' is invalid for input of size 7840

Have tried to get around it, i seem not able to understand whats wrong. From what have tried googling seems my dimensions have an issue and i dont know how to get the right dimensions If at all its the issue.

Answer 1

You are forgetting about "Batch" here. See, when you're in the for loop and destructuring your data with the line X, y = data. Here, what you may've thought is X is an image from trainset, but actually X is a batch of images (10 specifically, as you've set 10 as your batch_size) and y similarly is a batch of labels. Means, your X has a shape of something like 10x28x28x1. Now, you may realize the line net(X.view(28*28)) is actually squashing your whole batch of 10 images into a single 28x28=>784 size vector, which is wrong, right? So rather what you might want to do here is something like net(X.view(-1, 28*28)), now -1 will ensure that whatever your batch size is, it remains intact as you're sending data into the model, so now you could easily deduce the shape of X would be 10x784 which is what you want.

Another thing I've noticed, a typo maybe, your test_set and train_set are actually the same. To download the proper test_set, make sure to send train=False as the parameter.