运行内存CPU被撑爆

[qujinqiang]

我在 bilstm层上加crf 报一下问题：
/data/software/anaconda2/envs/py3_tf4/lib/python3.5/site-packages/tensorflow/python/ops/gradients_impl.py:96: UserWarning: Converting sparse IndexedSlices to a dense Tensor of unknown shape. This may consume a large amount of memory.
"Converting sparse IndexedSlices to a dense Tensor of unknown shape. "

训练代码大概如下所示：
def trainModel(self):
with tf.variable_scope('Inputs'):
self.X_inputs = tf.placeholder(tf.int32, [None, self.timestep_size], name='X_input') # 输入
self.y_inputs = tf.placeholder(tf.int32, [None, self.timestep_size], name='y_input') # 对应的标记

    bilstm_output = self.bi_lstm(self.X_inputs) # 返回的是隐藏层状态的输出

    with tf.variable_scope('outputs'):
        softmax_w = self.weight_variable([self.hidden_size * 2, self.class_num])
        softmax_b = self.bias_variable([self.class_num])
        self.y_pred = tf.matmul(bilstm_output, softmax_w) + softmax_b  # 隐藏层状态加上softmax才是实际的y预测值
        print(self.y_pred.shape)
        self.scores = tf.reshape(self.y_pred, [-1, self.timestep_size,
                                          self.class_num])  # 算好分数后，再重新reshape成[batchsize, timesteps, num_class]
        print(self.scores.shape)

    log_likelihood, self.transition_params = crf.crf_log_likelihood(self.scores, self.y_inputs, self.length)
    self.loss = tf.reduce_mean(-log_likelihood)

    optimizer = tf.train.AdamOptimizer(learning_rate=self.lr)
    self.train_op = optimizer.minimize(self.loss)

[Slyne]

Hi 不确定你的bilistm_output的输出shape是怎么样子的
另外crf的参数如下:
tf.contrib.crf.crf_log_likelihood
crf_log_likelihood(
inputs,
tag_indices,
sequence_lengths,
transition_params=None
)
inputs: A [batch_size, max_seq_len, num_tags] tensor of unary potentials to use as input to the CRF layer.
tag_indices: A [batch_size, max_seq_len] matrix of tag indices for which we compute the log-likelihood.
sequence_lengths: A [batch_size] vector of true sequence lengths.
transition_params: A [num_tags, num_tags] transition matrix, if available.

[qujinqiang]

@Slyne
代码上基本没有变化，但是可以跑模型了；基于2014的人民日报语料训练出来准确率大概92%左右
不过bi_lstm 加上 crf 层后依然存在着 "Converting sparse IndexedSlices to a dense Tensor of unknown shape. " 这种警告；
你那边训练的时候会存在相关警告吗？