Skip to content

Utilities

The npfl138 package also provides a few utilities for startup initialization, parameter initialization override, and for version management.

npfl138.startup 

startup(
    seed: int | None = None,
    threads: int | None = None,
    forkserver_instead_of_fork: bool = False,
) -> None

Initialize the environment.

  • Allow using TF32 for matrix multiplication.
  • Set the random seed if given.
  • Set the number of threads if given.
  • Use forkserver instead of fork if requested.

Parameters:

  • seed (int | None, default: None ) –

    If not None, set the Python, Numpy, and PyTorch random seeds to this value.

  • threads (int | None, default: None ) –

    If not None of 0, set the number of threads to this value. Otherwise, use as many threads as cores.

  • forkserver_instead_of_fork (bool, default: False ) –

    If True, use forkserver instead of fork as the default multiprocessing method. This will be the default in Python 3.14.

Source code in npfl138/startup.py 
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
def startup(seed: int | None = None, threads: int | None = None, forkserver_instead_of_fork: bool = False) -> None:
    """Initialize the environment.

    - Allow using TF32 for matrix multiplication.
    - Set the random seed if given.
    - Set the number of threads if given.
    - Use `forkserver` instead of `fork` if requested.

    Parameters:
      seed: If not `None`, set the Python, Numpy, and PyTorch random seeds to this value.
      threads: If not `None` of 0, set the number of threads to this value.
        Otherwise, use as many threads as cores.
      forkserver_instead_of_fork: If `True`, use `forkserver` instead of `fork` as the
        default multiprocessing method. This will be the default in Python 3.14.
    """

    # Allow TF32 when available.
    torch.backends.cuda.matmul.allow_tf32 = True

    # Set random seed if not None.
    if seed is not None:
        random.seed(seed)
        np.random.seed(seed)
        torch.manual_seed(seed)

    # Set number of threads if > 0; otherwise, use as many threads as cores.
    if threads is not None and threads > 0:
        torch.set_num_threads(threads)
        torch.set_num_interop_threads(threads)

    # If instructed, use `forkserver` instead of `fork` (which will be the default in Python 3.14).
    if "fork" in torch.multiprocessing.get_all_start_methods():
        if os.environ.get("FORCE_FORK_METHOD") == "1":
            torch.multiprocessing.set_start_method("fork")
        elif forkserver_instead_of_fork or os.environ.get("FORCE_FORKSERVER_METHOD") == "1":
            torch.multiprocessing.set_start_method("forkserver")

npfl138.global_keras_initializers 

global_keras_initializers(
    parameter_initialization: bool = True,
    batchnorm_momentum_override: bool = True,
) -> None

Change default PyTorch initializers to Keras defaults.

The following initializers are used:

  • Linear, Conv1d, Conv2d, Conv3d, ConvTranspose1d, ConvTranspose2d, ConvTranspose3d, Bilinear: Xavier uniform for weights, zeros for biases.
  • Embedding, EmbeddingBag: Uniform [-0.05, 0.05] for weights.
  • RNN, RNNCell, LSTM, LSTMCell, GRU, GRUCell: Xavier uniform for input weights, orthogonal for recurrent weights, zeros for biases (with LSTM forget gate bias set to 1).

Furthermore, for batch normalization layers, the default momentum value is changed from 0.1 to 0.01.

Parameters:

  • parameter_initialization (bool, default: True ) –

    If True, override the default PyTorch initializers with Keras defaults.

  • batchnorm_momentum_override (bool, default: True ) –

    If True, override the default momentum value of 0.1 in batch normalization layers to 0.01.

Source code in npfl138/initializers_override.py 
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
def global_keras_initializers(
    parameter_initialization: bool = True,
    batchnorm_momentum_override: bool = True,
) -> None:
    """Change default PyTorch initializers to Keras defaults.

    The following initializers are used:

    - `Linear`, `Conv1d`, `Conv2d`, `Conv3d`, `ConvTranspose1d`, `ConvTranspose2d`, `ConvTranspose3d`, `Bilinear`:
      Xavier uniform for weights, zeros for biases.
    - `Embedding`, `EmbeddingBag`: Uniform [-0.05, 0.05] for weights.
    - `RNN`, `RNNCell`, `LSTM`, `LSTMCell`, `GRU`, `GRUCell`: Xavier uniform for input weights,
      orthogonal for recurrent weights, zeros for biases (with LSTM forget gate bias set to 1).

    Furthermore, for batch normalization layers, the default momentum value is changed from 0.1 to 0.01.

    Parameters:
     parameter_initialization: If True, override the default PyTorch initializers with Keras defaults.
     batchnorm_momentum_override: If True, override the default momentum value of 0.1 in
       batch normalization layers to 0.01.
    """
    if parameter_initialization:
        for class_, reset_parameters_method in KerasParameterInitialization.overrides.items():
            class_.reset_parameters = reset_parameters_method

    if batchnorm_momentum_override:
        for batch_norm_super in KerasBatchNormMomentum.batch_norms:
            for batch_norm in [batch_norm_super] + batch_norm_super.__subclasses__():
                KerasBatchNormMomentum.override_default_argument_value(batch_norm.__init__, "momentum", 0.01)

npfl138.require_version 

require_version(required_version: str) -> None

Make sure the installed version is at least required_version.

If not, show a nice error message with the instructions how to upgrade.

Parameters:

  • required_version (str) –

    The required version of the npfl138 package, as a string in the format "semester.week" or "semester.week.patch".

Source code in npfl138/version.py 
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
def require_version(required_version: str) -> None:
    """Make sure the installed version is at least `required_version`.

    If not, show a nice error message with the instructions how to upgrade.

    Parameters:
      required_version: The required version of the npfl138 package, as
        a string in the format "semester.week" or "semester.week.patch".
    """

    required = required_version.split(".")
    assert len(required) <= 3, "Expected at most 3 version components"

    required = list(map(int, required))
    current = list(map(int, __version__.split(".")))

    assert current[:len(required)] >= required, (
        f"The npfl138>={required_version} is required, but found only {__version__}.\n"
        f"Please update the npfl138 package by running either:\n"
        f"- `VENV_DIR/bin/pip install --upgrade npfl138` when using a venv, or\n"
        f"- `python3 -m pip install --user --upgrade npfl138` otherwise.")

npfl138.__version__ module-attribute 

__version__ = '2425.10.0'