NumPy bridge (from_numpy, numpy) in PyTorch - Practice Problems & Coding Challenges

Challenge - 5 Problems

❓ Predict Output

intermediate

Output of PyTorch tensor from NumPy array modification

What is the output of the following code snippet?

PyTorch

import numpy as np
import torch
np_array = np.array([1, 2, 3])
torch_tensor = torch.from_numpy(np_array)
np_array[0] = 10
print(torch_tensor)

A tensor([10, 2, 3])

B Raises RuntimeError

C tensor([0, 2, 3])

D tensor([1, 2, 3])

💡 Hint

Remember that torch.from_numpy shares memory with the NumPy array.

The PyTorch tensor shares memory with the original NumPy array. Changing the NumPy array changes the tensor's values.

❓ Predict Output

intermediate

Effect of modifying PyTorch tensor on original NumPy array

What will be the output of the following code?

PyTorch

import numpy as np
import torch
np_array = np.array([4, 5, 6])
torch_tensor = torch.from_numpy(np_array)
torch_tensor[1] = 50
print(np_array)

A [0 50 0]

B [4 5 6]

C [ 4 50 6]

D Raises TypeError

💡 Hint

Think about whether the tensor and array share the same data.

Modifying the tensor changes the underlying NumPy array because they share memory.

❓ Model Choice

advanced

Choosing correct method to convert PyTorch tensor to NumPy array

Which method correctly converts a PyTorch tensor to a NumPy array without sharing memory?

A tensor.detach().numpy()

B tensor.clone().numpy()

C tensor.numpy()

D tensor.to_numpy()

💡 Hint

Consider how to avoid shared memory when converting.

Using clone() creates a copy of the tensor, so the resulting NumPy array does not share memory with the original tensor.

❓ Metrics

advanced

Understanding tensor and NumPy array memory sharing impact on metrics

If you modify a PyTorch tensor created from a NumPy array using torch.from_numpy, how does it affect the calculation of the mean of the original NumPy array?

A The mean changes because the NumPy array data changes with tensor modification.

B The mean calculation raises an error due to shared memory.

C The mean remains the same because tensors and arrays are independent.

D The mean changes only if tensor requires_grad is True.

💡 Hint

Think about how shared memory affects data values.

Since the tensor and NumPy array share memory, modifying the tensor changes the array's data, affecting the mean calculation.

🔧 Debug

expert

Debugging error when converting CUDA tensor to NumPy array

What error occurs when running this code and why? import torch t = torch.tensor([1, 2, 3]).cuda() np_array = t.numpy()

A RuntimeError: Can't call numpy() on Tensor that requires grad and is on GPU

B AttributeError: 'Tensor' object has no attribute 'numpy'

C TypeError: Cannot convert CUDA tensor to NumPy array directly

D RuntimeError: Can't call numpy() on a CUDA tensor. Move it to CPU first.

💡 Hint

Consider where the tensor is located (CPU or GPU) when calling numpy().

PyTorch tensors on GPU cannot be directly converted to NumPy arrays. They must be moved to CPU first.