tensorflow ==1.13.1
1. tf.data並行讀取tfrecord數據
def parse_exp(example):
features = {}
""" tfrecord解析代碼 """
return features
def input_fn(filenames = "./train_data/*.tfrecord", batch_size=128):
files = tf.data.Dataset.list_files(filenames)
dataset = files.apply(tf.contrib.data.parallel_interleave(lambda filename:
tf.data.TFRecordDataset(files), buffer_output_elements=batch_size*20, cycle_length=10))
dataset = dataset.shuffle(batch_size*4)
dataset = dataset.map(parse_exp, num_parallel_calls=8)
dataset = dataset.repeat().batch(batch_size).prefetch(1)
return dataset
2. 在進行分佈式訓練時,使用tf.fixed_size_partitioner參數分割,對於有較大Embedding計算的時候尤其有用,代碼如下:
def model_fn(features, mode, params):
""" 構建estimator模型 """
with tf.variable_scope("deviceID_embedding", partitioner=tf.fixed_size_partitioner(8, axis=0)):
deviceID_input = tf.feature_column.input_layer(features, params["deviceID"])
""" 構建自己的代碼邏輯 """
net = ...
output = tf.layers.dense(net, units=1)
return output
其中tf.fixed_size_partitioner(8, axis=0)的 8代表ps個數。