TensorFlow通过raw_ops调用底层op计算

tf底层的算子定义在tensorflow-1.15.0\tensorflow\core\ops\ops.pbtxt

raw_ops调用必须采用a = tf.raw_ops.Add(x=x, y=y)的形式，x=不能不写，可以采用下面这种方式解决：

name和data关联可以采用**arges

 inputs = dict(zip(input_names, input_data))
result_tf = tf_op(**inputs, **attr_seting)
 

example 1 简单加法

import tensorflow as tf
import numpy


x = 10
y = 20

a = tf.raw_ops.Add(x=x, y=y)

# Start training
with tf.Session() as sess:

    # Run the initializer
    print(sess.run(a))

注意要使用关键字参数，即形如Add(x=x, y=y)，具体的关键字参数是什么，要去算子介绍页面查看，例如tensorflow::ops::Add介绍页。

example 2 调用反向梯度优化op

import tensorflow as tf
import numpy as np


shape = (1,8)

var = tf.Variable(tf.ones(shape=shape),name='var')


alpha = 0.5
delta = tf.Variable(tf.ones(shape=shape),name='delta')


output = tf.raw_ops.ApplyGradientDescent(  var=var,  alpha=alpha,  delta=delta)

with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())

    # Run the initializer
    print(sess.run(output))
 
    

example 3 调用ApplyMomentum 反向梯度momentum优化op

import tensorflow as tf
import numpy as np


shape = (1,8)
lr = 1.0
momentum = 0.9

var = tf.Variable(tf.ones(shape=shape),name='var')
grad = tf.Variable(tf.ones(shape=shape),name='delta')
accum = tf.Variable(tf.ones(shape=shape),name='delta')

output = tf.raw_ops.ApplyMomentum(
         var=var,
         accum=accum,
         lr=lr,
         grad=grad,
         momentum=momentum)


with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())

    # Run the initializer
    print(sess.run(output))
    print(sess.run(var))
    print(sess.run(accum))
 
    

另一种简单的方法就是在TensorFlow的python包中找到算子定义包并导入调用，例如：

from tensorflow.python.training.gen_training_ops import apply_proximal_adagrad

example4

import tensorflow as tf
import numpy as np

tf.reset_default_graph()

pld1 = tf.placeholder(dtype="float32", shape=[None, 32])
pld2 = tf.placeholder(dtype="float32", shape=[None, 64])
pld3 = tf.placeholder(dtype="float32", shape=[None, 32])
pld4 = tf.placeholder(dtype="float32", shape=[None, 128])

batch = 1

exp1 = tf.raw_ops.Exp(x=pld1)
exp2 = tf.raw_ops.Exp(x=pld2)
exp3 = tf.raw_ops.Exp(x=pld3)
exp4 = tf.raw_ops.Exp(x=pld4)

values = [
    exp1,
    exp2,
    exp3,
    exp4,
]

input1_np = np.random.randn(*[batch, 32])
input2_np = np.random.randn(*[batch, 64])
input3_np = np.random.randn(*[batch, 32])
input4_np = np.random.randn(*[batch, 128])

feed_dict = dict(zip([pld1, pld2, pld3, pld4, ], [input1_np, input2_np, input3_np, input4_np]))

concat1 = tf.raw_ops.ConcatV2(values=values, axis=-1)

size_splits = tf.constant(
    [32,64,32,128],
    dtype="int32",
)

split1 = tf.raw_ops.SplitV(
    value=concat1, size_splits=size_splits, axis=-1, num_split=4)

with tf.Session() as sess:
    # sess.run(tf.global_variables_initializer())
    print(sess.run(split1, feed_dict=feed_dict))

# dump graph as pb
graph = tf.get_default_graph()
graph_def = graph.as_graph_def()

dump_graph_name = "dump_graph.pb"
with tf.gfile.GFile(dump_graph_name, "wb") as f:
    f.write(graph_def.SerializeToString())