目前我正在尝试使用Tensorflow附加我的代码来实现三元组丢失。
但我发现培训总是分歧。任何人都可以帮我找出问题所在吗？
def compute_triplet_loss(anchor_features, positive_features, negative_features, margin=0.1):
with tf.name_scope("triplet_loss"):
anchor_features_norm = compute_norm(anchor_features)
positive_features_norm = compute_norm(positive_features)
negative_features_norm = compute_norm(negative_features)
denom1 = tf.multiply(anchor_features_norm, positive_features_norm)
denom2 = tf.multiply(anchor_features_norm, negative_features_norm)
a_p_product = compute_dot_product(anchor_features, positive_features)
a_n_product = compute_dot_product(anchor_features, negative_features)
a_p_vec = tf.divide(a_p_product, denom1)
a_n_vec = tf.divide(a_n_product, denom2)
loss = tf.maximum(0., tf.add(tf.subtract(a_n_vec, a_p_vec), margin))
return loss
def DNN(input_tensor, is_training, Reuse=True):
tf.add_to_collection('input_tensor', input_tensor)
with tf.name_scope('uttr_net', [input_tensor]):
with tf.variable_scope('uttr_vars', reuse=Reuse):
x = slim.fully_connected(input_tensor, 1024, scope='fc_1')
x = tf.layers.batch_normalization(x, training=is_training, momentum=0.995, name='bn_1')
x = slim.fully_connected(x, 1024, scope='fc_2')
x = tf.layers.batch_normalization(x, training=is_training, momentum=0.995, name='bn_2')
x = slim.fully_connected(x, 1024, scope='fc_3')
x = tf.layers.batch_normalization(x, training=is_training, momentum=0.995, name='bn_3')
x = slim.fully_connected(x, 80, scope='d_vec')
x = tf.layers.batch_normalization(x, training=is_training, momentum=0.995, name='bn_4')
# pooling
x = tf.expand_dims(x, 1)
x = tf.nn.avg_pool(x, ksize=[1,1,2,1], strides=[1,1,1,1], padding='VALID')
x = tf.squeeze(x, [1])
x = slim.dropout(x, keep_prob=0.65, is_training=is_training, scope='final_drop')
x = tf.reduce_mean(x, axis=1, name='uttr_mean')
tf.add_to_collection('d-vec', x)
return x
def build_graph(input_shape, learning_rate=0.001):
input_positive_0 = tf.placeholder(tf.float32, shape=input_shape, name='input_positive_0')
input_positive_1 = tf.placeholder(tf.float32, shape=input_shape, name='input_positive_1')
input_negative = tf.placeholder(tf.float32, shape=input_shape, name='input_negative')
is_training = tf.placeholder(tf.bool, [])
tf.add_to_collection('input_positive_0', input_positive_0)
tf.add_to_collection('input_positive_1', input_positive_1)
tf.add_to_collection('input_negative', input_negative)
tf.add_to_collection('is_training', is_training)
DNN_0 = DNN(input_positive_0, is_training, False)
DNN_1 = DNN(input_positive_1, is_training)
DNN_2 = DNN(input_negative, is_training)
loss = compute_triplet_loss(DNN_0, DNN_1, DNN_2)
loss = tf.reduce_sum(loss)
optimizer = tf.train.AdamOptimizer(learning_rate).minimize(loss)
return optimizer, loss, input_positive_0, input_positive_1, input_negative, is_training