frozencherry/Forgery-Localization-App

1import tensorflow as tf
2
3
4def dice_coefficient(y_true, y_pred, smooth=1e-5):
5    intersection = tf.reduce_sum(y_true * y_pred)
6    union = tf.reduce_sum(y_true) + tf.reduce_sum(y_pred)
7    dice = (2. * intersection + smooth) / (union + smooth)
8    return dice
9
10def dice_coef_loss(y_true, y_pred):
11    return 1-dice_coefficient(y_true, y_pred)
12
13def iou(y_true, y_pred, smooth=1):
14    intersection = tf.reduce_sum(y_true * y_pred)
15    union = tf.reduce_sum(y_true) + tf.reduce_sum(y_pred) - intersection
16    iou = (intersection + smooth) / (union + smooth)
17    return iou
18
19def iou_loss(y_true, y_pred):
20    return 1 - iou(y_true, y_pred)
21
22def accuracy(y_true, y_pred):
23    correct_pixels = tf.reduce_sum(tf.cast(tf.equal(y_true, tf.round(y_pred)), dtype=tf.float32))
24    total_pixels = tf.cast(tf.reduce_prod(tf.shape(y_true)), dtype=tf.float32)
25    accuracy = correct_pixels / total_pixels
26    return accuracy
27
28def weighted_dice_bce_loss(y_true, y_pred, bce_weight=0.5, smooth=1):
29    y_true = tf.cast(y_true, dtype=tf.float32)
30    y_pred = tf.cast(y_pred, dtype=tf.float32)
31    
32    intersection = tf.reduce_sum(y_true * y_pred)
33    dice = (2. * intersection + smooth) / (tf.reduce_sum(y_true) + tf.reduce_sum(y_pred) + smooth)
34    dice_loss = 1 - dice
35    
36    bce_loss = tf.reduce_mean(tf.keras.losses.binary_crossentropy(y_true, y_pred))
37    
38    total_loss = bce_weight * bce_loss + (1 - bce_weight) * dice_loss
39    return total_loss