Skip to content

Latest commit

 

History

History
95 lines (83 loc) · 4.3 KB

kv variable_api.md

File metadata and controls

95 lines (83 loc) · 4.3 KB
Raw

Kv Variable

| English | 中文 |

We have implemented an entirely new Variable, named KvVariable. Its internal data type utilizes key-value format for storage (such as using a hash map structure), and supports the following features:

  • KvVariable only need embedding_dimension to be created. New features will be dynamically added to the variable
  • KvVariable is primarily used for embedding calculations, it support forward and backward computations of embedding_lookup, and support parameter updates of all optimizers.
  • KvVariable supports PartitionedVariable for sharding large embeddings.
  • KvVariable supports save/restore with checkpoint, savedmodel formats, and supports automatic clipping of sparse vectors when exported.

User Interface

The tf.get_kv_variable is a low-level API supporting tf variable_scope and partitioner.

   def get_kv_variable(name,
                       embedding_dim=None,
                       key_dtype=dtypes.int64,
                       value_dtype=dtypes.float32,
                       initializer=None,
                       trainable=None,
                       collections=None,
                       partitioner=None,
                       constraint=None):

  • name: name of the variable.

  • embedding_dim: dimension of the variable.

  • key_dtype: type of the variable key (default is tf.int64; accept int, int32).

  • value_dtype: type of the variable value (default is tf.float32).

  • initializer: Initializer or Tensor. If it is a Tensor, its shape must be defined.

  • trainable: a collection list that the variable will be added to, default is [GraphKeys.GLOBAL_VARIABLES] (see tf.Variable).

  • partitioner: optional parameter, accepts a partition function, features in the variable will be evenly distributed into several partitions.

    tf.get_kv_variable low level api, with tf.variable_scope and partitioner support

       def get_kv_variable(name,
                       embedding_dim=None,
                       key_dtype=dtypes.int64,
                       value_dtype=dtypes.float32,
                       initializer=None,
                       trainable=None,
                       collections=None,
                       partitioner=None,
                       constraint=None):

embedding ops：Embedding ops: support the following three types of lookup. Same with the original TensorFlow interfaces.

low level api example

import tensorflow as tf
import os

ckp_dir = './checkpoint'
ckp_path = os.path.join(ckp_dir, 'model.ckpt')

num_shards = 5
with tf.variable_scope('test', reuse=tf.AUTO_REUSE):
var = tf.get_kv_variable("kv_embedding",
                           embedding_dim=64,
                           key_dtype=tf.int64,
                           initializer=tf.compat.v1.ones_initializer(),
                        partitioner=tf.compat.v1.fixed_size_partitioner(num_shards=num_shards))

emb = tf.nn.embedding_lookup(var, tf.convert_to_tensor([1, 2, 3,6,8,9], dtype=tf.int64))
emb1 = tf.nn.embedding_lookup(var, tf.convert_to_tensor([1000000000000000], dtype=tf.int64))
fun = tf.add(emb, 1.0, name='add')
loss = tf.reduce_sum(fun)
opt = tf.train.FtrlOptimizer(0.005,
                           l1_regularization_strength=0.025,
                           l2_regularization_strength=0.00001)
g_v = opt.compute_gradients(loss)
train_op = opt.apply_gradients(g_v)
init = tf.global_variables_initializer()
saver = tf.train.Saver()

with tf.Session() as sess:
sess.run(init)
print(sess.run({'emb':emb, 'fun':fun, 'train': train_op}))
print(sess.run(emb1))
save_path = saver.save(sess, ckp_path)
print'model saved in file %s' % save_path)

with tf.Session() as sess:
saver.restore(sess, ckp_path)
print(sess.run(emb1))
print(sess.run({'emb':emb, 'fun':fun}))
print(sess.run(emb1))