TensorFlow 的可训练权重 sequence_loss_by_example()

Question

我想要 seq2seq.sequence_loss_by_example() 的可训练权重，例如

w = tf.get_variable("w", [batch_size*num_steps])
loss = seq2seq.sequence_loss_by_example([logits_1],
            [tf.reshape(self._targets, [-1])],
            w,vocab_size_all)

但是，运行这段代码给我以下错误：

seq2seq.py, line 654, in sequence_loss_by_example
if len(targets) != len(logits) or len(weights) != len(logits):

根据 seq2seq.py 中此函数的文档字符串：

weights: list of 1D batch-sized float-Tensors of the same length as logits.

它需要一个 "Tensor"，但我想传递一个 tf.Variable。有没有办法在这个函数中有可训练的权重？

Answer 1

在 TensorFlow 中，需要 tf.Variable can be used anywhere a tf.Tensor（具有相同的元素类型和形状）。

因此，如果您想定义一个可训练的权重，您可以将 tf.Variable 个对象的列表作为 weights 参数传递给 seq2seq.sequence_loss_by_example()。例如，您可以执行以下操作：

# Defines a list of `num_steps` variables, each 1-D with length `batch_size`.
weights = [tf.get_variable("w", [batch_size]) for _ in range(num_steps)]

loss = seq2seq.sequence_loss_by_example([logits_1, ..., logits_n],
                                        [targets_1, ..., targets_n],
                                        weights,
                                        vocab_size_all)

TensorFlow 的可训练权重 sequence_loss_by_example()

Trainable weight for TensorFlow sequence_loss_by_example()

python

tensorflow