Tensor shapes and dimensions in PyTorch - Practice Problems & Coding Challenges

Challenge - 5 Problems

❓ Predict Output

intermediate

What is the shape of the tensor after this operation?

Consider the following PyTorch code snippet. What is the shape of the tensor y after running this code?

PyTorch

import torch
x = torch.randn(4, 3, 2)
y = x.permute(1, 0, 2)

A (3, 4, 2)

B (4, 2, 3)

C (2, 3, 4)

D (3, 2, 4)

💡 Hint

Remember that permute rearranges the dimensions in the order you specify.

The original tensor x has shape (4, 3, 2). The permute(1, 0, 2) swaps the first and second dimensions, so the new shape is (3, 4, 2).

❓ Model Choice

intermediate

Choosing the correct input shape for a CNN

You want to build a convolutional neural network (CNN) in PyTorch to classify color images of size 64x64 pixels. Which input tensor shape should you use for a batch of 16 images?

A (64, 64, 3, 16)

B (16, 64, 64, 3)

C (16, 3, 64, 64)

D (3, 16, 64, 64)

💡 Hint

PyTorch expects image tensors in the format (batch_size, channels, height, width).

PyTorch convolution layers expect input tensors shaped as (batch_size, channels, height, width). For 16 images with 3 color channels and 64x64 pixels, the shape is (16, 3, 64, 64).

❓ Hyperparameter

advanced

Effect of batch size on tensor shape during training

If you change the batch size from 32 to 64 in a training loop, how does the shape of the input tensor to the model change?

A The tensor shape remains unchanged

B All dimensions double

C The last dimension doubles, others stay the same

D The first dimension doubles, other dimensions stay the same

💡 Hint

Batch size affects the number of samples processed at once, which is the first dimension.

The batch size corresponds to the first dimension of the input tensor. Increasing batch size from 32 to 64 doubles this dimension, while other dimensions (channels, height, width) remain unchanged.

🔧 Debug

advanced

Why does this tensor reshape cause an error?

Given a tensor t with shape (2, 3, 4), why does the following code raise an error?

t.view(3, 7)

PyTorch

import torch
t = torch.randn(2, 3, 4)
t_reshaped = t.view(3, 7)

A Because view cannot change the number of dimensions

B Because the total number of elements does not match between original and new shape

C Because 3 and 8 are not valid dimensions for any tensor

D Because the tensor must be contiguous before calling view

💡 Hint

Check the total number of elements before and after reshape.

The original tensor has 2*3*4 = 24 elements. The requested shape (3, 7) requires 3*7 = 21 elements, which does not match. PyTorch's `view()` requires the total number of elements to be the same.

🧠 Conceptual

expert

Understanding broadcasting with tensor shapes

You have two tensors: a with shape (5, 1, 4) and b with shape (1, 3, 4). What will be the shape of the result when you add a + b in PyTorch?

A (5, 3, 4)

B (5, 1, 4)

C (1, 3, 4)

D (5, 3)

💡 Hint

Broadcasting expands dimensions of size 1 to match the other tensor's size.

Broadcasting rules expand dimensions with size 1 to match the other tensor's size. Here, (5,1,4) and (1,3,4) broadcast to (5,3,4).