Losses¶

For the losses, if y_pred is complex and y_true is floating, y_true will be automatically cast to complex by replicating its value on the imaginary part.

Complex Average Cross Entropy¶

Inspired on [CIT2018-CAO] Average Cross Entropy (ACE) loss function described on section 2.4.

This function applies Categorical Cross Entropy to both the real and imaginary part separately and then averages it. Mathematically this is

\[J^{ACE} = \frac{1}{2} \left[ J^{CCE}(\Re \hat{y}, y) + J^{CCE}(\Im \hat{y}, y) \right] \, ,\]

where \(J^{ACE}\) is the Complex Average Cross Entropy, \(J^{CCE}\) is the well known Categorical Cross Entropy. \(\hat{y}\) is the predicted labels with the corresponding ground truth \(y\). Finally \(\Re\) and \(\Im\) operators are the real and imaginary parts of the input respectively. For real-valued output \(J^{ACE} = J^{CCE}\).

Working example:

from cvnn.layers import ComplexDense, complex_input
from cvnn.losses import ComplexAverageCrossEntropy
import cvnn.dataset as dp
import tensorflow as tf

# Get dataset
m = 10000
n = 128
param_list = [
    [0.3, 1, 1],
    [-0.3, 1, 1]
]
dataset = dp.CorrelatedGaussianCoeffCorrel(m, n, param_list, debug=False)

# Build model
model = tf.keras.models.Sequential([
    complex_input(shape=(n)),
    ComplexDense(units=50, activation="cart_relu"),
    ComplexDense(2, activation="cart_softmax")
])

# Compile using ACE complex loss function
model.compile(loss=ComplexAverageCrossEntropy(), metrics=["accuracy"], optimizer="sgd")

model.fit(dataset.x, dataset.y, epochs=6)

Complex Weighted Average Cross Entropy¶

class ComplexWeightedAverageCrossEntropy¶

Assigns a weight to be passed as input to be multiplied to the result as

\[L={l_1, …, l_N}^⊤,\]

with

\[l_n =−w_n J^{ACE}_n\]

__init__(self, weights, **kwargs)¶

Parameters:	weights – List of weights to be applied. Must be same length as total classes.

Complex Average Cross Entropy Ignore Unlabeled¶

Complex Weighted Average Cross Entropy Ignore Unlabeled¶

Complex Mean Square Error¶

Performs the mean square error defined as:

\[\mathcal{L} = \Delta x^{2} + \Delta y^{2} \, ,\]

where \(\Delta x\) and \(\Delta y\) represents the real and imaginary part difference between the label and predicted respectively.

Working example:

import numpy as np
import tensorflow as tf
from cvnn.losses import ComplexMeanSquareError
y_true = np.random.randint(0, 2, size=(2, 3)).astype("float32")
y_pred = tf.complex(np.random.random(size=(2, 3)).astype("float32"),
                    np.random.random(size=(2, 3)).astype("float32"))
loss = ComplexMeanSquareError().call(y_true, y_pred)
expected_loss = np.mean(np.square(np.abs(tf.complex(y_true, y_true) - y_pred)), axis=-1)
assert np.all(loss == expected_loss)

[CIT2018-CAO]

Yice Cao, Yan Wu, Peng Zhang, Wenkai Liang and Ming Li “Pixel-Wise PolSAR Image Classification via a Novel Complex-Valued Deep Fully Convolutional Network” https://arxiv.org/abs/1909.13299 2019