TensorFlow LSTM模型中的NaN损耗

问题描述

下面的网络代码应该是您经典的简单LSTM语言模型，稍后将开始输出NaN丢失.在我的训练集上，这需要几个小时，而且我不能在较小的数据集上轻松地复制它。但这在严肃的训练中总是会发生的。

稀疏软最大值与交叉熵在数值上应该是稳定的，因此它不可能是原因.但除此之外，我没有在图中看到任何其他可能导致问题的节点。可能是什么问题？

class MyLM():
    def __init__(self, batch_size, embedding_size, hidden_size, vocab_size):
        self.x = tf.placeholder(tf.int32, [batch_size, None])  # [batch_size, seq-len]
        self.lengths = tf.placeholder(tf.int32, [batch_size])  # [batch_size]

        # remove padding. [batch_size * seq_len] -> [batch_size * sum(lengths)]
        mask = tf.sequence_mask(self.lengths)  # [batch_size, seq_len]
        mask = tf.cast(mask, tf.int32)  # [batch_size, seq_len]
        mask = tf.reshape(mask, [-1])  # [batch_size * seq_len]

        # remove padding + last token. [batch_size * seq_len] -> [batch_size * sum(lengths-1)]
        mask_m1 = tf.cast(tf.sequence_mask(self.lengths - 1, maxlen=tf.reduce_max(self.lengths)), tf.int32)  # [batch_size, seq_len]
        mask_m1 = tf.reshape(mask_m1, [-1])  # [batch_size * seq_len]

        # remove padding + first token.  [batch_size * seq_len] -> [batch_size * sum(lengths-1)]
        m1_mask = tf.cast(tf.sequence_mask(self.lengths - 1), tf.int32)  # [batch_size, seq_len-1]
        m1_mask = tf.concat([tf.cast(tf.zeros([batch_size, 1]), tf.int32), m1_mask], axis=1)  # [batch_size, seq_len]
        m1_mask = tf.reshape(m1_mask, [-1])  # [batch_size * seq_len]

        embedding = tf.get_variable("TokenEmbedding", shape=[vocab_size, embedding_size])
        x_embed = tf.nn.embedding_lookup(embedding, self.x)  # [batch_size, seq_len, embedding_size]

        lstm = tf.nn.rnn_cell.LSTMCell(hidden_size, use_peepholes=True)

        # outputs shape: [batch_size, seq_len, hidden_size]
        outputs, final_state = tf.nn.dynamic_rnn(lstm, x_embed, dtype=tf.float32,
                                                 sequence_length=self.lengths)
        outputs = tf.reshape(outputs, [-1, hidden_size])  # [batch_size * seq_len, hidden_size]

        w = tf.get_variable("w_out", shape=[hidden_size, vocab_size])
        b = tf.get_variable("b_out", shape=[vocab_size])
        logits_padded = tf.matmul(outputs, w) + b  # [batch_size * seq_len, vocab_size]
        self.logits = tf.dynamic_partition(logits_padded, mask_m1, 2)[1]  # [batch_size * sum(lengths-1), vocab_size]

        predict = tf.argmax(logits_padded, axis=1)  # [batch_size * seq_len]
        self.predict = tf.dynamic_partition(predict, mask, 2)[1]  # [batch_size * sum(lengths)]

        flat_y = tf.dynamic_partition(tf.reshape(self.x, [-1]), m1_mask, 2)[1]  # [batch_size * sum(lengths-1)]

        self.cross_entropy = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=self.logits, labels=flat_y)
        self.cost = tf.reduce_mean(self.cross_entropy)
        self.train_step = tf.train.AdamOptimizer(learning_rate=0.01).minimize(self.cost)

---

解决方案

可能是exploding gradients的情况，其中梯度可能在LSTM中的反向传播期间爆炸，从而导致数字溢出。处理爆炸梯度的常用技术是执行Gradient Clipping。