Utilities (utils)

Data loading

utils.Dataset

Abstract base class for datasets. Subclass it to wrap custom data sources:

from numpygrad.utils import Dataset
import numpygrad as npg

class MyDataset(Dataset):
    def __init__(self, X, y):
        self.X = X
        self.y = y

    def __len__(self) -> int:
        return len(self.X)

    def __getitem__(self, index):
        return npg.array(self.X[index]), npg.array(self.y[index])

utils.TensorDataset

Wraps two arrays (data and targets) as a dataset. Internally stores them as Array objects and uses slice-based batching for efficiency:

from numpygrad.utils import TensorDataset
import numpy as np

X = np.random.randn(500, 8).astype(np.float32)
y = np.random.randint(0, 4, size=(500,))

dataset = TensorDataset(X, y)
print(len(dataset))         # 500
x_i, y_i = dataset[0]      # single sample

utils.DataLoader

Iterates over a dataset in batches:

from numpygrad.utils import DataLoader

loader = DataLoader(dataset, batch_size=32, shuffle=True, drop_last=False)

for x_batch, y_batch in loader:
    ...  # x_batch: Array (32, 8), y_batch: Array (32,)

Parameters:

• batch_size — number of samples per batch

• shuffle — whether to shuffle at the start of each epoch (default True)

• drop_last — discard the final batch if it is smaller than batch_size (default False)

For TensorDataset, batching is done with a single slice operation. For custom Dataset subclasses, samples are fetched one at a time and stacked.

Visualisation

utils.draw_computation_graph(root, save_path=None)

Renders the computation graph rooted at root as a PNG using Graphviz:

from numpygrad.utils import draw_computation_graph
from pathlib import Path

x = npg.array([1.0, 2.0], requires_grad=True)
y = npg.array([3.0, 4.0], requires_grad=True)
z = (x * y).sum()

draw_computation_graph(z, save_path=Path("graph.png"))

Requires the optional graphviz package:

pip install graphviz

If save_path is None, the graph is displayed inline (in a Jupyter notebook) or saved to a temporary file.

Seeding

For npg.manual_seed and the full npg.random namespace see Random (npg.random).