-
Notifications
You must be signed in to change notification settings - Fork 0
/
loss.py
71 lines (58 loc) · 2.86 KB
/
loss.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
import tensorflow as tf
from tensorflow import keras
import numpy as np
#!!!ADD LOSS MODULES HERE !!!
# based on the results of this paper: https://www.sciencedirect.com/science/article/pii/S2666827021001092
def berhu_loss(y_true, y_pred, threshold=0.2):
bounds = threshold * keras.backend.max(keras.backend.abs(y_pred - y_true))
l1 = keras.backend.abs(y_pred - y_true)
l2 = (keras.backend.square(y_pred - y_true) + (bounds ** 2)) / (2 * bounds)
l1_mask = tf.cast((l1 <= bounds), tf.float32)
res = l1 * l1_mask + l2 * tf.cast(l1_mask == 0, tf.float32)
return keras.backend.mean(res)
def ssim_loss(y_true, y_pred, sharpen=False):
if sharpen:
edge_kernel = np.array([[0, -1, 0], [-1, 5, -1], [0, -1, 0]])
edge_kernel = np.expand_dims(edge_kernel, 2)
edge_kernel = np.expand_dims(edge_kernel, 3)
edge_kernel = tf.constant(edge_kernel, dtype=tf.float32)
sharp_img = tf.nn.conv2d(y_true, edge_kernel, strides=[1,1,1,1], padding="SAME")
y_true = sharp_img
l_ssim = 1 - tf.image.ssim(y_true, y_pred, max_val=1.0)
return tf.reduce_mean(l_ssim)
def sobel_loss(y_true, y_pred):
sobel_x = np.array([[1, 0, -1], [2, 0, -2], [1, 0, -1]])
sobel_x = np.expand_dims(sobel_x, 2)
sobel_x = np.expand_dims(sobel_x, 3)
sobel_x = tf.constant(sobel_x, dtype=tf.float32)
sobel_y = np.array([[-1, -2, -1], [0, 0, 0], [1, 2, 1]])
sobel_y = np.expand_dims(sobel_y, 2)
sobel_y = np.expand_dims(sobel_y, 3)
sobel_y = tf.constant(sobel_y, dtype=tf.float32)
y_true_x = tf.nn.conv2d(y_true, sobel_x, strides=[1,1,1,1], padding="SAME")
y_true_y = tf.nn.conv2d(y_true, sobel_y, strides=[1,1,1,1], padding="SAME")
y_pred_x = tf.nn.conv2d(y_pred, sobel_x, strides=[1,1,1,1], padding="SAME")
y_pred_y = tf.nn.conv2d(y_pred, sobel_y, strides=[1,1,1,1], padding="SAME")
diff_x = tf.math.abs(y_pred_x - y_true_x)
diff_y = tf.math.abs(y_pred_y - y_true_y)
return keras.backend.mean(diff_x + diff_y)
def edge_loss(y_true, y_pred):
dy_true, dx_true = tf.image.image_gradients(y_true)
dy_pred, dx_pred = tf.image.image_gradients(y_pred)
l_edges = tf.keras.backend.mean(tf.keras.backend.abs(dy_pred - dy_true) + tf.keras.backend.abs(dx_pred - dx_true), axis=-1)
l_edges = tf.reduce_mean(l_edges)
return l_edges
def depth_smoothness_loss(y_true, y_pred):
# Edges
dy_true, dx_true = tf.image.image_gradients(y_true)
dy_pred, dx_pred = tf.image.image_gradients(y_pred)
weights_x = tf.exp(tf.reduce_mean(tf.abs(dx_true)))
weights_y = tf.exp(tf.reduce_mean(tf.abs(dy_true)))
smoothness_x = dx_pred * weights_x
smoothness_y = dy_pred * weights_y
depth_smoothness_loss = tf.reduce_mean(abs(smoothness_x)) + tf.reduce_mean(
abs(smoothness_y)
)
return depth_smoothness_loss
def l1_loss(y_true, y_pred):
return tf.reduce_mean(tf.abs(y_true - y_pred))