-
Notifications
You must be signed in to change notification settings - Fork 0
/
DNN_misc.py
120 lines (101 loc) · 4.43 KB
/
DNN_misc.py
1
2
3
4
5
6
7
8
9
10
11
12
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
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
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
119
120
import glob
import os
from pathlib import Path
import numpy as np
import tensorflow as tf
import tensorflow.keras
from tensorflow.keras import backend as k
from tensorflow.keras.layers import Dense, Dropout, Activation, Flatten, BatchNormalization, Input, ReLU
from Disco_tf import distance_corr
def manageOutputdir(outputdir):
# check if dir exists
if(os.path.exists(outputdir)):
i=2
while(os.path.exists(outputdir+"__"+str(i))):
i+=1
outputdir += "__"+str(i)
Path(outputdir).mkdir(parents=True)
Path(outputdir+"/plots/").mkdir(parents=True)
Path(outputdir+"/data/").mkdir(parents=True)
Path(outputdir+"/network/").mkdir(parents=True)
return outputdir
# custom metrics
def acc_all():
m = tf.keras.metrics.Accuracy()
def accuracy(labels, predictions):
weights = tf.reshape(labels[:,2],[-1, 1])
m.update_state(labels[:,0], predictions, sample_weight=weights)
return m.result()
return accuracy
# partial accuracys for weighted
def acc_sig(labels, predictions):
msk = k.equal(labels[:,0], 1)
bools = k.equal(tf.boolean_mask(labels[:,0], msk), k.round(tf.boolean_mask(predictions, msk)))
bools = tf.cast(bools, dtype=tf.float32)
bools = tf.math.multiply(bools, tf.boolean_mask(labels[:,2], msk))
return k.mean(bools)
def acc_bkg(labels, predictions):
msk = k.equal(labels[:,0], 0)
bools = k.equal(tf.boolean_mask(labels[:,0], msk), k.round(tf.boolean_mask(predictions, msk)))
bools = tf.cast(bools, dtype=tf.float32)
bools = tf.math.multiply(bools, tf.boolean_mask(labels[:,2], msk))
return k.mean(bools)
def nominalLoss_wrapper():
def nominalLoss(y_true, y_pred):
bce = tf.keras.losses.MeanSquaredError()
#bce = tf.keras.losses.BinaryCrossentropy()
weights = tf.reshape(y_true[:,2],[-1])
true = tf.reshape(y_true[:,0],[-1, 1])
pred = tf.reshape(y_pred,[-1, 1])
loss_bce = bce(true, pred, sample_weight=weights)
return loss_bce
return nominalLoss
def DisCoLoss_wrapper(param):
def DisCoLoss(y_true, y_pred):
weights = tf.reshape(y_true[:,2],[-1])
weights_normed = tf.reshape(tf.linalg.normalize(weights)[0],[-1])
weights_normed = tf.math.multiply(weights_normed, tf.cast(tf.size(weights_normed), tf.dtypes.float32)/tf.math.reduce_sum(weights_normed))
st = tf.reshape(y_true[:,1],[-1])
pred = tf.reshape(y_pred,[-1])
loss_DisCo = param * distance_corr(st, pred, weights_normed, 2)
return loss_DisCo
return DisCoLoss
# loss for DisCo
def cross_entropy_DisCo(param):
def loss(y_true, y_pred):
# calculating weighted BCE
bce = tf.keras.losses.BinaryCrossentropy()
weights = tf.reshape(y_true[:,2],[-1, 1])
loss_bce = bce(y_true[:,0], y_pred, sample_weight=weights)
# calculating DisCo
# ensuring weights sum up to sample count
weights_normed = tf.reshape(tf.linalg.normalize(weights)[0],[-1])
weights_normed = tf.math.multiply(weights_normed, tf.cast(tf.size(weights_normed), tf.dtypes.float32)/tf.math.reduce_sum(weights_normed))
st = tf.reshape(y_true[:,1],[-1])
pred = tf.reshape(y_pred,[-1])
loss_DisCo = param * distance_corr(st, pred, weights_normed, 2)
return loss_bce + loss_DisCo
return loss
def mean_squared_error_DisCo(param):
def loss(y_true, y_pred):
# calculating weighted BCE
mse = tf.keras.losses.MeanSquaredError()
weights = tf.reshape(y_true[:,2],[-1])
pred = tf.reshape(y_pred,[-1, 1])
true = tf.reshape(y_true[:,0],[-1, 1])
loss_mse = mse(true, pred, sample_weight=weights)
# calculating DisCo
# ensuring weights sum up to sample count
weights_normed = tf.reshape(tf.linalg.normalize(weights)[0],[-1])
weights_normed = tf.math.multiply(weights_normed, tf.cast(tf.size(weights_normed), tf.dtypes.float32)/tf.math.reduce_sum(weights_normed))
st = tf.reshape(y_true[:,1],[-1])
pred = tf.reshape(y_pred,[-1])
loss_DisCo = param * distance_corr(st, pred, weights_normed, 2)
return loss_mse + loss_DisCo
return loss
def layerblock(L, nodes, kernel_init, dropout_val=0, batchNorm=False):
if(batchNorm): L = BatchNormalization()(L)
L = Dense(nodes, kernel_initializer=kernel_init, bias_initializer='zeros', activation='tanh')(L)
if(dropout_val > 0): L = Dropout(dropout_val)(L)
L = ReLU()(L)
return L