NumPy bridge (from_numpy, numpy) in PyTorch - ML Experiment: Train & Evaluate

Experiment - NumPy bridge (from_numpy, numpy)

Problem:You want to efficiently convert data between NumPy arrays and PyTorch tensors to use PyTorch models with existing NumPy data.

Current Metrics:Currently, you manually convert data using separate steps, which is slow and error-prone.

Issue:Manual conversions cause extra memory usage and slow down data processing.

Your Task

Use PyTorch's from_numpy and numpy() functions to convert data back and forth efficiently without extra memory copies.

You must use PyTorch and NumPy only.

Do not create new arrays or tensors unnecessarily.

Ensure changes in one reflect in the other.

Hint 1

Hint 2

Hint 3

Solution

PyTorch

import numpy as np
import torch

# Create a NumPy array
np_array = np.array([1, 2, 3, 4], dtype=np.float32)

# Convert NumPy array to PyTorch tensor (shares memory)
tensor = torch.from_numpy(np_array)

# Modify the tensor
tensor[0] = 10

# Check if NumPy array changed
print(f"NumPy array after tensor change: {np_array}")  # Should show [10.  2.  3.  4.]

# Convert tensor back to NumPy array (shares memory)
np_array2 = tensor.numpy()

# Modify the new NumPy array
np_array2[1] = 20

# Check if tensor changed
print(f"Tensor after NumPy array change: {tensor}")  # Should show tensor([10., 20.,  3.,  4.])

Used torch.from_numpy() to create tensor sharing memory with NumPy array.

Used tensor.numpy() to convert tensor back to NumPy array sharing memory.

Demonstrated that modifying one changes the other, confirming no extra copies.

Results Interpretation

Before: Manual conversions created copies, causing extra memory use and slower data handling.

After: Using from_numpy and numpy() shares memory between NumPy arrays and tensors, making conversions fast and memory efficient.

Sharing memory between NumPy arrays and PyTorch tensors avoids unnecessary data copying, improving performance and reducing memory use.

Bonus Experiment

Try modifying a tensor created from a NumPy array with requires_grad=True and observe how gradients behave.

💡 Hint

Remember that tensors sharing memory with NumPy arrays cannot have requires_grad=True directly; consider cloning or detaching.