How to save and load NumPy arrays with NPZ

NPZ archives keep related NumPy arrays together in one file, which suits small checkpoints, paired training data, and intermediate outputs that should be reloaded as a set. Unlike CSV or plain text exports, each member keeps NumPy shape and dtype metadata.

np.savez() writes an uncompressed archive, and np.savez_compressed() uses ZIP compression for smaller files when write speed is less important than file size. Passing keyword arguments gives each array a stable key, while positional arrays are loaded later as arr_0, arr_1, and similar generated names.

Loading an NPZ file returns a dictionary-like NpzFile object rather than one array. Use a context manager or close the loaded object after reading so the archive file handle is released, and keep allow_pickle=False for archives that do not need trusted object arrays.

1. Create a script that writes named arrays to a compressed NPZ archive.

   array-save-load-npz.py

   import numpy as np
   from pathlib import Path
    
   path = Path("model-batch.npz")
   features = np.arange(12, dtype=np.float32).reshape(3, 4)
   labels = np.array([0, 1, 1], dtype=np.int64)
   sample_ids = np.array(["train-001", "train-002", "train-003"])
    
   np.savez_compressed(
       path,
       features=features,
       labels=labels,
       sample_ids=sample_ids,
   )
    
   with np.load(path) as data:
       print("keys:", sorted(data.files))
       print("features shape:", data["features"].shape)
       print("features dtype:", data["features"].dtype)
       print("labels:", data["labels"].tolist())
       print("sample ids:", data["sample_ids"].tolist())
       print("features match:", np.array_equal(features, data["features"]))
       print("labels match:", np.array_equal(labels, data["labels"]))
    
   print("file exists:", path.exists())

   Keyword names passed to np.savez_compressed() become archive keys. Use np.savez() with the same keyword pattern when write speed matters more than file size.

2. Run the script and check the loaded keys, dtype, and equality checks.

   $ python array-save-load-npz.py
   keys: ['features', 'labels', 'sample_ids']
   features shape: (3, 4)
   features dtype: float32
   labels: [0, 1, 1]
   sample ids: ['train-001', 'train-002', 'train-003']
   features match: True
   labels match: True
   file exists: True

   The with np.load(path) block closes the NpzFile handle after the arrays are read.

3. Remove the sample script and archive.

   $ rm model-batch.npz array-save-load-npz.py

   Keep the NPZ file instead when it is the checkpoint another job should consume.