Chapter 18: Integration with NumPy¶

“Two libraries. One memory space. No nonsense.”

18.1 Why Integrate with NumPy?¶

NumPy is the OG of numerical computing in Python. Even if you're deep into PyTorch, you’ll often need to:

PyTorch makes this easy by letting you share memory between torch.Tensor and np.ndarray.

import numpy as np
import torch

arr = np.array([1.0, 2.0, 3.0])
t = torch.from_numpy(arr)

arr[0] = 99
print(t)  # tensor([99., 2., 3.])

This is fast and efficient — no memory copy.

t = torch.tensor([1.0, 2.0, 3.0])
arr = t.numpy()

Again, same memory — not a copy.

t = torch.tensor([1.0, 2.0]).to('cuda')
arr = t.cpu().numpy()  # Must move to CPU first!

import matplotlib.pyplot as plt
image = torch.randn(28, 28)
plt.imshow(image.numpy(), cmap='gray')

import pandas as pd
df = pd.read_csv('data.csv')
tensor = torch.from_numpy(df.values).float()

np.mean(tensor.numpy(), axis=0)

Use NumPy when you need broadcasting or functions that PyTorch lacks (e.g., np.percentile()).

The NumPy↔Torch conversion is zero-copy. That’s great, but watch for:

Gotcha	Solution
CUDA tensors can't `.numpy()`	Move to CPU first: t.cpu()`.numpy()`
Detached views may be unsafe	Use `.clone()` if unsure
Mixed float types (e.g. float64)	Use `.float()` before model usage
In-place ops affect both	Clone before modifying either

Avoid inside:

Use .numpy() only for:

preds = model(x)
np_preds = preds.detach().cpu().numpy()

Convert back to float32 if NumPy defaults to float64:
```
torch.from_numpy(arr.astype(np.float32))
```

Use .contiguous() if you hit weird shape bugs:

tensor = tensor.permute(1, 0).contiguous()

PyTorch and NumPy play together beautifully — just watch out for GPU vs CPU boundaries
These conversions are efficient, but must be used with care during training