# Built-in imports
import warnings, os
# External imports
import numpy as np
# Medusa imports
from medusa import tensorflow_integration
# Extras
if os.environ.get("MEDUSA_EXTRAS_GPU_TF") == "1":
import tensorflow as tf
def __aux_symm_triu_gpu(W):
N = tf.shape(W)
aux = tf.subtract(tf.linalg.band_part(tf.ones(N), 0, -1),tf.linalg.band_part(tf.ones(N), 0, 0))
W = tf.math.multiply(W,tf.cast(aux,tf.float64))
W = tf.math.add(W,tf.transpose(W))
W = tf.math.divide(tf.math.reduce_sum(W,axis=0),2)
return W
def __aux_no_match_gpu(W):
in_degree = tf.math.reduce_sum(W,axis=0)
out_degree = tf.math.reduce_sum(W,axis=1)
W = tf.math.add(in_degree, out_degree)
return W
def __degree_gpu(W):
"""
Calculates node degree (also called strength in weighted networks) using GPU
Parameters
----------
W : numpy 2D matrix
Graph matrix. ChannelsXChannels.
Returns
-------
nodal_degree : numpy array
Nodal degree.
"""
W = tf.math.divide(tf.math.round(tf.math.multiply(W,10000000000)),10000000000)
W = W - tf.cast(tf.eye(tf.shape(W)[0]),tf.float64)
check_symmetry = tf.reduce_all(tf.math.equal(W,tf.transpose(W)))
check_symmetry = tf.cond(
tf.math.reduce_sum(tf.subtract(tf.linalg.band_part(W, -1, 0),tf.linalg.band_part(W, 0, 0))) == 0,
lambda: 1, lambda: check_symmetry)
check_symmetry = tf.cond(
tf.reduce_all(tf.math.equal(W,-tf.transpose(W))),
lambda: 2, lambda: check_symmetry)
nodal_degree = tf.switch_case(tf.cast(check_symmetry,tf.int32),
branch_fns={0: lambda: __aux_no_match_gpu(W), 1: lambda: __aux_symm_triu_gpu(W), 2: lambda: -tf.math.reduce_sum(W,axis=0)})
return nodal_degree
def __aux_symm_triu_cpu(W):
N = np.shape(W)[0]
aux = np.ones((N,N))
aux = np.triu(aux,k=1)
W = W * aux
W = W + np.transpose(W)
W = np.sum(W,axis=0) / 2
return W
def __aux_no_match_cpu(W):
in_degree = np.sum(W,axis=0)
out_degree = np.sum(W,axis=1)
W = in_degree + out_degree
return W
def __degree_cpu(W):
"""
Calculates node degree (also called strength in weighted networks) using CPU
Parameters
----------
W : numpy 2D matrix
Graph matrix. ChannelsXChannels.
Returns
-------
nodal_degree : numpy array
Nodal degree.
"""
W = np.divide(np.round(W * 10000000000),10000000000)
W = W - np.diag(np.diag(W))
check_symmetry = (W.transpose() == W).all() # if symmetric
if (W == np.triu(W)).all(): # if upper triangular
check_symmetry = 1
if (W.transpose() == -W).all(): # if anti-symmetric
check_symmetry = 2
if check_symmetry == 0:
nodal_degree = __aux_no_match_cpu(W)
elif check_symmetry == 1:
nodal_degree = __aux_symm_triu_cpu(W)
elif check_symmetry == 2:
nodal_degree = -np.sum(W,axis=0)
return nodal_degree
[docs]def degree(W, mode):
"""
Calculates the graph degree.
Parameters
----------
W : numpy 2D matrix
Graph matrix. ChannelsXChannels.
Returns
-------
nodal_degree : numpy array
Nodal degree.
"""
if W.shape[0] is not W.shape[1]:
raise ValueError('W matrix must be square')
if not np.issubdtype(W.dtype, np.number):
raise ValueError('W matrix contains non-numeric values')
if mode == 'GPU' and os.environ.get("MEDUSA_EXTRAS_GPU_TF") == "1" and \
tensorflow_integration.check_tf_config(autoconfig=True):
nodal_degree = __degree_gpu(W)
else:
nodal_degree = __degree_cpu(W)
return nodal_degree