tensorflow 在加载大型的embedding模型参数时，会遇到cannot be larger than 2GB

这种问题是，对于每一个变量 variable 由于是基于protobuf存在这大小限制(2G),这个时候，我们需要将embedding拆开，拆分成N等分，来使得每一个

variable都在2G以下;

　　

 # !/usr/bin/env/python
 # coding=utf-8
 import tensorflow as tf
 import numpy as np

 input_ids = tf.placeholder(dtype=tf.int32, shape=[None,None])

 num_shards = 3
 weights = []
 weights_shape = np.arange(27).reshape(9, 3)
 # assert weights_shape[0] % num_shards == 0
 num_shards_len = (weights_shape.shape[0]) / num_shards
 assert  (weights_shape.shape[0]) % num_shards ==0
 begin_ = 0
 ends_ = num_shards_len
 for i in range(0, num_shards):
     if (i + 1) * num_shards_len < weights_shape.shape[0]:
         begin_ = i * num_shards_len
         if i + 1 == num_shards:
             ends_ = weights_shape.shape[0]
         else:
             ends_ = (i + 1) * num_shards_len
     else:
         begin_ = i * num_shards_len
         ends_ = weights_shape.shape[0]
     weights_i = tf.get_variable("words-%02d" % i,
                                 initializer=tf.constant(weights_shape[begin_: ends_, ]))
     weights.append(weights_i)

 input_embedding = tf.nn.embedding_lookup(weights, input_ids,partition_strategy="div")

 sess = tf.InteractiveSession()
 sess.run(tf.global_variables_initializer())
 print(sess.run(weights))

 print(sess.run(input_embedding, feed_dict={input_ids: [[1, 2], [3, 0], [8, 2], [5, 1]]}))

结果为:

[array([[0, 1, 2],
       [3, 4, 5],
       [6, 7, 8]]), array([[ 9, 10, 11],
       [12, 13, 14],
       [15, 16, 17]]), array([[18, 19, 20],
       [21, 22, 23],
       [24, 25, 26]])]
[[[ 3  4  5]
  [ 6  7  8]]

 [[ 9 10 11]
  [ 0  1  2]]

 [[24 25 26]
  [ 6  7  8]]

 [[15 16 17]
  [ 3  4  5]]]