tf.keras.initializers.STFT

Initializer of Conv kernels for Short-term Fourier Transformation (STFT).

Inherits From: Initializer

tf.keras.initializers.STFT(
    side='real',
    window='hann',
    scaling='density',
    periodic=False
)

Since the formula involves complex numbers, this class compute either the real or the imaginary components of the final output.

Additionally, this initializer supports windowing functions across the time dimension as commonly used in STFT. Windowing functions from the module scipy.signal.windows are supported, including the common hann and hamming windowing functions. This layer supports periodic windows and scaling-based normalization.

This is primarily intended for use in the STFTSpectrogram layer.

Examples:

# Standalone usage:
initializer = STFTInitializer("real", "hann", "density", False)
values = initializer(shape=(128, 1, 513))

side String, "real" or "imag" deciding if the kernel will compute the real side or the imaginary side of the output. Defaults to "real".
window String for the name of the windowing function in the scipy.signal.windows module, or array_like for the window values, or None for no windowing.
scaling String, "density" or "spectrum" for scaling of the window for normalization, either L2 or L1 normalization. None for no scaling.
periodic Boolean, if True, the window function will be treated as periodic. Defaults to False.

Methods

clone

View source

clone()

from_config

View source

@classmethod
from_config(
    config
)

Instantiates an initializer from a configuration dictionary.

Example:

initializer = RandomUniform(-1, 1)
config = initializer.get_config()
initializer = RandomUniform.from_config(config)

Args
config A Python dictionary, the output of get_config().

Returns
An Initializer instance.

get_config

View source

get_config()

Returns the initializer's configuration as a JSON-serializable dict.

Returns
A JSON-serializable Python dict.

__call__

View source

__call__(
    shape, dtype=None
)

Returns a tensor object initialized as specified by the initializer.

The shape is assumed to be (T, 1, F // 2 + 1), where T is the size of the given window, and F is the number of frequency bands. Only half the frequency bands are used, which is a common practice in STFT, because the second half are the conjugates of the first half in a reversed order.

Args
shape Shape of the tensor.
dtype Optional dtype of the tensor. Only numeric or boolean dtypes are supported. If not specified, keras.backend.floatx() is used, which default to float32 unless you configured it otherwise (via keras.backend.set_floatx(float_dtype)).