quacknet.activationFunctions

 1import numpy as np
 2
 3def relu(values, alpha = 0.01):
 4    """
 5    Applies Leaky Rectified Linear Unit (ReLU) activation function.
 6
 7    Args:
 8        values (ndarray): Input array to apply leaky ReLU to.
 9        alpha (float, optional): Slope for negative values. Default is 0.01.
10    
11    Returns:
12        ndarray: Array with Leaky ReLU applied to it.    
13    """
14    return np.maximum(values * alpha, values) 
15     
16def sigmoid(values):
17    """
18    Applies the sigmoid activation function.
19
20    Args:
21        values (ndarray): Input array to apply sigmoid.
22    
23    Returns:
24        ndarray: Array with sigmoid applied to it.    
25    """
26    return 1 / (1 + np.exp(-values))
27
28def tanH(values):
29    """
30    Applies the hyperbolic tangent (tanh) activation function.
31
32    Args:
33        values (ndarray): Input array to apply tanh.
34    
35    Returns:
36        ndarray: Array with tanh applied to it.    
37    """
38    return np.tanh(values)
39
40def linear(values): #Dont use too demanding on CPU
41    """
42    Applies the linear activation function.
43
44    Args:
45        values (ndarray): Input array.
46    
47    Returns:
48        ndarray: Output array (same as input).    
49    """
50    return values
51
52def softMax(values): 
53    """
54    Applies the softmax activation function.
55
56    Args:
57        values (ndarray): Input array to apply softmax.
58    
59    Returns:
60        ndarray: Array with softmax applied to it. 
61
62    Note:
63        The function handles overflow by subtracting the maximum value from inputs.
64    """
65    values = np.array(values, dtype=np.float64)
66    maxVal = np.max(values)
67    values = values - maxVal
68    summ = np.sum(np.exp(values))
69    out = np.exp(values) / summ
70    return out