tf.losses模块下的tf.reduction

下面以常用的softmax_cross_entropy() 函数为例,做一个详细介绍。

tf.losses.softmax_cross_entropy(
    onehot_labels,  # 注意此处参数名就叫 onehot_labels
    logits,
    weights=1.0,
    label_smoothing=0,
    scope=None,
    loss_collection=tf.GraphKeys.LOSSES,
    reduction=Reduction.SUM_BY_NONZERO_WEIGHTS
)

Args:

  1. onehot_labels: [batch_size, num_classes] one_hot类型的label.
  2. logits: [batch_size, num_classes] 神经网络的logits输出. 这两个参数都没什么好解释的,最基本的知识。
  3. weights: 为可选参数,首先它是tensor 然后它可以是一个标量(此时作为loss的系数),也可以是一个[batch_size]的向量(看源码我个人理解就是对一个batch的样本进行加权,不知道有什么意义,因为一般来说训练样本的输入顺序是随机的,即每一个batch的样本全部都是随机的,这中情况下这个加权没有任何意义了)
  4. label_smoothing: 这个参数如果设置大于0,则对label进行平滑,平滑的公式为:
  5. new_onehot_labels = onehot_labels*(1-label_smoothing) + label_smoothing/num_classes
  6. scope: 命名空间
  7. loss_collection: 指定loss集合。
  8. reduction: 指定应用到这个loss的reduction类型.
  • NONE: Un-reduced weighted losses with the same shape as input.
  • SUM: Scalar sum of weighted losses.
  • MEAN: Scalar ‘SUM’ divided by sum of weights.
  • SUM_OVER_BATCH_SIZE: Scalar ‘SUM’ divided by number of elements in losses.
  • SUM_OVER_NONZERO_WEIGHTS: Scalar ‘SUM’ divided by number of non-zero weights.
  • SUM_BY_NONZERO_WEIGHTS: Same as ‘SUM_OVER_NONZERO_WEIGHTS’(注意:在损失函数中weights=1.0, 对应这里的weighted_sum_by_nonzero_weights,配合起来就是求和)

reduction一般都是使用SUM_BY_NONZERO_WEIGHTS,不过需要注意的是设置这个参数的时候是:reduction=tf.losses.Reduction.MEAN
 

将上述代码改写成 TensorFlow 2.x 版本的代码如下: ```python import tensorflow as tf def weight(shape, stddev, wd): initial = tf.random.truncated_normal(shape, stddev=stddev) var = tf.Variable(initial) if wd is not None: weight_decay = tf.multiply(tf.nn.l2_loss(var), wd, name='weight_loss') tf.add_to_collection('losses', weight_decay) return var class MyModel(tf.keras.Model): def __init__(self, in_nodes=784, h1_nodes=100, h2_nodes=100, h3_nodes=50): super(MyModel, self).__init__() self.w1 = weight([in_nodes, h1_nodes], 0.1, 0.005) self.b1 = tf.Variable(tf.zeros([h1_nodes])) self.w2 = weight([h1_nodes, h2_nodes], 0.1, 0.0) self.b2 = tf.Variable(tf.zeros([h2_nodes])) self.w3 = weight([h2_nodes, h3_nodes], 0.1, 0.0) self.b3 = tf.Variable(tf.zeros([h3_nodes])) self.w4 = weight([h3_nodes, 10], 0.1, 0.0) self.b4 = tf.Variable(tf.zeros([10])) def call(self, inputs, prob): x = inputs y_ = tf.cast(inputs, tf.float32) h1 = tf.nn.relu(tf.matmul(x, self.w1) + self.b1) h2 = tf.nn.relu(tf.matmul(h1, self.w2) + self.b2) h2_drop = tf.nn.dropout(h2, rate=prob) h3 = tf.nn.relu(tf.matmul(h2_drop, self.w3) + self.b3) h3_drop = tf.nn.dropout(h3, rate=prob) y = tf.nn.softmax(tf.matmul(h3_drop, self.w4) + self.b4) return y model = MyModel() x = tf.keras.Input(shape=(None, 784)) prob = tf.keras.Input(shape=()) y = model(x, prob) y_ = tf.keras.Input(shape=(None, 10)) # 定义损失函数 cross_entropy = tf.reduce_mean(-tf.reduce_sum(y_ * tf.math.log(y), reduction_indices=[1])) tf.add_to_collection('losses', cross_entropy) loss = tf.add_n(tf.get_collection('losses')) # 定义优化器 train_step = tf.train.AdamOptimizer(1e-4).minimize(loss) # 训练模型 with tf.Session() as sess: sess.run(tf.global_variables_initializer()) for i in range(1000): batch_xs, batch_ys = mnist.train.next_batch(100) sess.run(train_step, feed_dict={x: batch_xs, y_: batch_ys, prob: 0.5}) ``` 在 TensorFlow 2.x ,可以使用 `tf.reduce_mean` 和 `tf.reduce_sum` 函数来计算张量的平均值和总和;使用 `tf.math.log` 函数来计算张量的自然对数。此外,可以使用 `tf.train.AdamOptimizer` 来定义优化器,使用 `model.trainable_variables` 来获取所有可训练的变量。
评论
添加红包

请填写红包祝福语或标题

红包个数最小为10个

红包金额最低5元

当前余额3.43前往充值 >
需支付:10.00
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明:

1.余额是钱包充值的虚拟货币,按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载,可以购买VIP、付费专栏及课程。

余额充值