def f1_reweight_loss(logits,
label_ids,
positive_idx,
negative_idx,
correct_class_weight=None,
wrong_confusion_matrix=None,
label_size=19):
softmax_layer = MaskedSoftmaxLayer("softmax_layer")
probs = softmax_layer(logits)
batch_idx = tf.range(tf.shape(probs)[0])
label_with_idx = tf.concat(
[tf.expand_dims(t, 1) for t in [batch_idx, label_ids]], 1)
golden_prob = tf.gather_nd(probs, label_with_idx)
m = tf.reduce_sum(positive_idx)
n = tf.reduce_sum(negative_idx)
p1 = tf.reduce_sum(positive_idx * golden_prob)
p2 = tf.reduce_sum(negative_idx * golden_prob)
beta2 = 1
neg_weight = p1 / ((beta2 * m) + n - p2 + 1e-8)
all_one = tf.ones(tf.shape(golden_prob))
loss_weight = all_one * positive_idx + all_one * neg_weight * negative_idx
loss = -loss_weight * tf.log(golden_prob + 1e-8)
return loss
def f1_confusion_loss(logits,label_ids,positive_idx,negative_idx,correct_class_weight,wrong_confusion_matrix, label_size = 19):
softmax_layer = MaskedSoftmaxLayer("softmax_layer")
probs = softmax_layer(logits)
log_probs = tf.log(probs + 1e-8)
log_one_minus_prob = tf.log(1-probs + 1e-8)
batch_idx = tf.range(tf.shape(probs)[0])
label_with_idx = tf.concat([tf.expand_dims(t, 1) for t in [batch_idx,label_ids]], 1)
golden_prob = tf.gather_nd(probs,label_with_idx)
golden_log_prob = tf.gather_nd(log_probs,label_with_idx)
positvie_confusion_weight = tf.gather(correct_class_weight,label_ids)
negative_confusion_weight = tf.gather(wrong_confusion_matrix,label_ids) #B*label_size
positive_cost = positvie_confusion_weight * golden_log_prob
negative_cost = tf.reduce_sum(negative_confusion_weight * log_one_minus_prob,axis = 1)
cost = positive_cost + negative_cost
m = tf.reduce_sum(positive_idx)
n = tf.reduce_sum(negative_idx)
p1 = tf.reduce_sum(positive_idx * golden_prob)
p2 = tf.reduce_sum(negative_idx * golden_prob)
neg_weight = p1 / (m+n-p2 + 1e-8)
all_one = tf.ones(tf.shape(golden_prob))
balanced_weight = all_one * positive_idx + all_one * neg_weight * negative_idx
loss = - balanced_weight * cost
return loss
def f1_entropy_loss(logits,label_ids,positive_idx = None,negative_idx = None,correct_class_weight =None,wrong_confusion_matrix = None, label_size = 19):
wrong_confusion_matrix = 1.0- tf.diag(tf.Variable([1.0] * label_size,dtype = tf.float32,trainable = False))
softmax_layer = MaskedSoftmaxLayer("softmax_layer")
probs = softmax_layer(logits)
log_probs = tf.log(probs + 1e-8)
negative_entropy = probs * tf.log(probs + 1e-8)
batch_idx = tf.range(tf.shape(probs)[0])
label_with_idx = tf.concat([tf.expand_dims(t, 1) for t in [batch_idx,label_ids]], 1)
golden_prob = tf.gather_nd(probs,label_with_idx)
golden_log_prob = tf.gather_nd(log_probs,label_with_idx)
negative_confusion_weight = tf.gather(wrong_confusion_matrix,label_ids) #B*label_size
positive_cost = golden_log_prob
negative_cost = tf.reduce_sum(- negative_confusion_weight * negative_entropy,axis = 1)
cost = positive_cost + negative_cost
m = tf.reduce_sum(positive_idx)
n = tf.reduce_sum(negative_idx)
p1 = tf.reduce_sum(positive_idx * golden_prob)
p2 = tf.reduce_sum(negative_idx * golden_prob)
neg_weight = p1 / (m+n-p2 + 1e-8)
all_one = tf.ones(tf.shape(golden_prob))
balanced_weight = all_one * positive_idx + all_one * neg_weight * negative_idx
loss = - balanced_weight * cost
return loss
def likelihood_loss(logits,label_ids,positive_idx = None,negative_idx = None,correct_class_weight =None,wrong_confusion_matrix = None, label_size = 19):
softmax_layer = MaskedSoftmaxLayer("softmax_layer")
probs = softmax_layer(logits)
batch_idx = tf.range(tf.shape(probs)[0])
label_with_idx = tf.concat([tf.expand_dims(t, 1) for t in [batch_idx,label_ids]], 1)
golden_prob = tf.gather_nd(probs,label_with_idx)
loss = - tf.log(golden_prob +1e-8)
return loss
def weight_loss(logits,label_ids,positive_idx = None,negative_idx = None,correct_class_weight =None,wrong_confusion_matrix = None, label_size = 19):
class_weight = [1.0] * label_size
class_weight[9] = 0.1
class_weight = tf.Variable(class_weight,dtype = tf.float32,name = "cls_weight",trainable = False)
softmax_layer = MaskedSoftmaxLayer("softmax_layer")
probs = softmax_layer(logits)
batch_idx = tf.range(tf.shape(probs)[0])
label_with_idx = tf.concat([tf.expand_dims(t, 1) for t in [batch_idx,label_ids]], 1)
golden_prob = tf.gather_nd(probs,label_with_idx)
loss_weight = tf.gather(class_weight,label_ids)
loss = - loss_weight * tf.log(golden_prob + 1e-8)
return loss