我使用tensorflow / slim，使用具有l2正则化的conv2d层，如下所示：
net = slim.conv2d(inputs, n_filters, filter_size, weights_regularizer=slim.l2_regularizer(0.001), activation_fn=None)
为了通过交叉熵和正则化损失对总损失进行正则化，我编写了以下脚本：
loss_cross = tf.nn.softmax_cross_entropy_with_logits_v2(logits=network, labels=net_output)
beta = 0.01
regularization_losses = tf.losses.get_regularization_losses()
losses = loss_cross + beta * regularization_losses
loss = tf.reduce_mean(losses)
opt = tf.train.AdamOptimizer(lr).minimize(0.0001, var_list=[var for var in tf.trainable_variables()])
出现错误：
Traceback (most recent call last):
File "main_orgsettings_losses.py", line 288, in <module>
losses = loss_cross + beta * regularization_losses
TypeError: can't multiply sequence by non-int of type 'float'
当我移除beta时，出现错误消息：
InvalidArgumentError: Incompatible shapes: [1,512,512] vs. [1]
输入大小中的[512，512]和[1]是唯一具有正则化的层。
当我在reduce_mean之后将损失添加正则化时：
loss = tf.reduce_mean(losses) + regularization_losses
它可以工作，但是我认为不是beta未被使用，我认为应该在{{1}之前将reguilarization_loss添加到losses }