TensorFlow 中 tf.enable_eager_execution 使用实例_tf.enable_eager_execution( config=none, device

本文链接：https://blog.csdn.net/sdnuwjw/article/details/97276336

一、环境

TensorFlow API r1.14(rc)

CUDA 9.0 V9.0.176

Python 3.6.3

二、官方说明

在该程序的生命周期内启用“立即执行（eager execution）”

https://tensorflow.google.cn/versions/r1.14/api_docs/python/tf/enable_eager_execution

tf.enable_eager_execution(
config=None,
device_policy=None,
execution_mode=None
)

参数：

config：可选参数。通过 tf.compat.v1.ConfigProto 来配置操作执行的环境。需要注意的是 tf.compat.v1.ConfigProto 也用来配置图执行（通过 tf.compat.v1.Session），同时，当 eager execution 执行时，tf.compat.v1.ConfigProto 中的很多选项就不被实现（或无关）

device_policy：可选参数。控制在特定设备（例如，GPU 0）上需要输入的操作如何处理不同设备（例如，GPU 1或CPU）上的输入的策略。设置为“无”时，将自动选择适当的值。挑选的值可能会在TensorFlow版本之间发生变化。有效值有：

tf.contrib.eager.DEVICE_PLACEMENT_EXPLICIT
tf.contrib.eager.DEVICE_PLACEMENT_WARN
tf.contrib.eager.DEVICE_PLACEMENT_SILENT
tf.contrib.eager.DEVICE_PLACEMENT_SILENT_FOR_INT32

execution_mode：可选参数。控制调度操作实际如何执行的策略。当设置为 None 时，将自动挑选合适的数值。挑选的值可能会在TensorFlow版本之间发生变化。有效值有：

tf.contrib.eager.SYNC：同步执行每一项操作
tf.contrib.eager.ASYNC：异步执行每一项操作。这些操作可能返回 “non-ready” 的句柄

三、实例

>>> import tensorflow as tf
>>> import numpy as np
>>> tf.enable_eager_execution()
>>> data = np.random.random((5,30))
>>> data_tensor = tf.constant(value=data)
>>> split0,split1,split2 = tf.split(data_tensor,[4,15,11],-1)
>>> split0, split1, split2 = tf.split(data_tensor,[4,15,11],-1)
>>> tf.shape(split0)
<tf.Tensor: id=6, shape=(2,), dtype=int32, numpy=array([5, 4], dtype=int32)>
>>> tf.shape(split1)
<tf.Tensor: id=8, shape=(2,), dtype=int32, numpy=array([ 5, 15], dtype=int32)>
>>> tf.shape(split2)
<tf.Tensor: id=10, shape=(2,), dtype=int32, numpy=array([ 5, 11], dtype=int32)>

>>> split0.numpy()
array([[0.57453663, 0.09221047, 0.75910165, 0.42249415],
       [0.16336272, 0.89446457, 0.39257956, 0.72432271],
       [0.65180994, 0.36158292, 0.76486074, 0.90457114],
       [0.42410351, 0.33851704, 0.7050377 , 0.37142487],
       [0.22980119, 0.42904001, 0.29638316, 0.17583503]])
>>> split1.numpy()
array([[0.65376257, 0.28828267, 0.68491934, 0.0453603 , 0.52976963,
        0.51527368, 0.13887819, 0.66672668, 0.32432196, 0.53790642,
        0.73945528, 0.65519244, 0.00321005, 0.76134054, 0.43237806],
       [0.38422725, 0.85679974, 0.8189933 , 0.79522319, 0.28097536,
        0.01417454, 0.33259705, 0.58593418, 0.41862654, 0.92597912,
        0.62811071, 0.07483196, 0.57353834, 0.94820909, 0.05068218],
       [0.69485806, 0.5419911 , 0.53619689, 0.41984836, 0.17919892,
        0.90462194, 0.85142706, 0.88211552, 0.86519832, 0.46317626,
        0.32401062, 0.10729327, 0.94318461, 0.78615075, 0.31926406],
       [0.35100308, 0.73412197, 0.27498129, 0.56456188, 0.66821586,
        0.86167801, 0.99093966, 0.51488332, 0.54963875, 0.33069172,
        0.43082452, 0.28137253, 0.8633286 , 0.66494205, 0.85997601],
       [0.96145664, 0.05172849, 0.08726628, 0.16821706, 0.3590603 ,
        0.70607823, 0.83647191, 0.52196516, 0.45792938, 0.97573745,
        0.8225903 , 0.09839245, 0.41375104, 0.4079168 , 0.89302919]])
>>> split2.numpy()
array([[0.11837789, 0.52289792, 0.55807716, 0.96394212, 0.84602479,
        0.7968168 , 0.13631762, 0.99491611, 0.0810203 , 0.03436556,
        0.14814265],
       [0.5901035 , 0.99906073, 0.74242789, 0.63917107, 0.57637234,
        0.2895648 , 0.06453467, 0.75643109, 0.08641431, 0.8303831 ,
        0.50956032],
       [0.62365326, 0.84957791, 0.79885836, 0.90246727, 0.30868528,
        0.19903051, 0.78012872, 0.57343972, 0.28900716, 0.65201702,
        0.33714608],
       [0.17511412, 0.27725245, 0.6815811 , 0.89634696, 0.30047085,
        0.24422117, 0.13571881, 0.95206347, 0.36601743, 0.74425454,
        0.46046426],
       [0.38272897, 0.04541412, 0.68629762, 0.93640718, 0.42502042,
        0.70174166, 0.75355039, 0.15139218, 0.36311097, 0.55398162,
        0.09967336]])