时间戳

首页 > 解决方案 > 如何使用图中的旧值更新具有新值的张量？

python - 如何使用图中的旧值更新具有新值的张量？

问题描述

我对张量流很陌生。我想使用张量的旧值来计算新值。tf.assign仅适用于tf.Variables. 我不确定如何为张量操作实现这一点。

以下代码不是实际的代码段，但想法是相同的。

data.csv
inp1            inp2
288.15          288.15
289.87912       303.10137
291.60825       318.05275
292.90509       329.26628
294.20194       340.47981
295.75815       353.93605
297.31436       367.39229
298.87057       380.84852
300.42679       394.30476
301.983         407.761


import tensorflow as tf
import pandas as pd

inp1 = tf.placeholder("float", [None, 1],name="inp1")
inp2 = tf.placeholder("float", [None, 1],name="inp2")


# PREVop means previous value of op i.e. op [i-1]

dummy_op = tf.add(inp1, 10)

op = tf.Variable(dummy_op,validate_shape=False,dtype=tf.float32)


# op[i] = (op[i-1]*inp2[i]) + inp1[i]
op = tf.add(tf.multiply(PREVop, inp2), inp1) 

label = tf.placeholder("float", [None,1],name="label")


learning_rate = 1e-2
loss_op = tf.losses.absolute_difference(label, op)
optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate, epsilon=0.1)
train_op = optimizer.minimize(loss_op)

inp = pd.read_csv("data.csv")
batch_size = 20
training_steps = 100

with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())
    print ("Training starts.......")
    for step in range(training_steps):
        avg_cost = 0.
        total_batch = len(inp)//batch_size
        for i in range(total_batch):
            idx = np.arange(len(inp))
            np.random.shuffle(idx)
            idx = idx[i*batch_size:(i+1)*batch_size]
            _, c = sess.run([train_op, loss_op], feed_dict={
            inp1: input1.values[idx] #pandas
            inp2: input2.values[idx] #pandas
            label: target.values[idx], #pandas
            })

在训练期间，我期望对每个样本op = tf.add(tf.multiply(PREVop, inp2), inp1)使用之前的值。op

任何建议，将不胜感激。

标签： pythonpython-3.xtensorflowtensor

解决方案

由于 的值op始终在变化，因此您可以tf.Variable()在每次迭代后使用它来存储其值。在这里，tf.Variable()在开始时用零张量初始化。

import tensorflow as tf
import numpy as np

inp1 = tf.placeholder(tf.float32, [None, 1], name="inp1")
inp2 = tf.placeholder(tf.float32, [None, 1], name="inp2")

PREVop = tf.Variable(tf.zeros([2, 1]), tf.float32)

out = tf.add(tf.multiply(PREVop, inp2), inp1) 
PREVop = tf.assign(PREVop, out)  

init = tf.global_variables_initializer()

with tf.Session() as sess:
    sess.run(init)
    for i in range(3):
        res, var = sess.run([out, PREVop], feed_dict={inp1:np.random.rand(2, 1), inp2:np.random.rand(2, 1)})
        print('out operation result: \n{}'.format(res))
        print('PREVop value after assigning: \n{}'.format(var))
        print(20*'-')

输出：

out operation result: 
[[0.86163723]
 [0.7938016 ]]
PREVop value after assigning: 
[[0.86163723]
 [0.7938016 ]]
--------------------
out operation result: 
[[0.5666107]
 [0.9492748]]
PREVop value after assigning: 
[[0.5666107]
 [0.9492748]]
--------------------
out operation result: 
[[0.89638215]
 [0.93310213]]
PREVop value after assigning: 
[[0.89638215]
 [0.93310213]]
--------------------

更新：因此，您希望使用 进行初始化PREVop，tf.add(inp1, 10)然后使用 的值对其进行更新op，即tf.add(tf.multiply(PREVop, inp2), inp1). 我正在添加一种方法来做到这一点，但老实说，我不喜欢它。

代码：

batch_size = 2

inp1 = tf.placeholder(tf.float32, [None, 1],name="inp1")
inp2 = tf.placeholder(tf.float32, [None, 1],name="inp2")

inp3 = tf.placeholder(tf.float32, [None, 1], name="inp3")
PREVop = tf.Variable(tf.zeros([batch_size, 1]), dtype=tf.float32)
PREVop = tf.assign(PREVop, inp3)

out = tf.add(tf.multiply(PREVop, inp2), inp1)  

inp = pd.read_csv("data.csv", sep=' ')
x_train = inp.iloc[:,:-1]
training_steps = 100

with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())
    out_res = 0
    for step in range(training_steps):
        total_batch = len(inp)//batch_size
        for i in range(total_batch):
            batch_x = x_train[i*batch_size:min((i+1)*batch_size, len(inp))]

            if step==0 and i==0:
                _, res = sess.run([PREVop, out], feed_dict={inp1: batch_x['inp1'].values.reshape(2, 1), 
                                                            inp2: batch_x['inp2'].values.reshape(2, 1), 
                                                            inp3: batch_x['inp1'].values.reshape(2, 1)+10})   
            else:
                _, res = sess.run([PREVop, out], feed_dict={inp1: batch_x['inp1'].values.reshape(2, 1), 
                                                            inp2: batch_x['inp2'].values.reshape(2, 1), 
                                                            inp3: out_res})
            out_res = res

在上面的代码中，batch_size=2我使用了一个占位符inp3并将 value 提供给它inp1+10，稍后将其分配给 variable PREVop。这在开始时只发生一次，后来PREVop被赋值out（在下面的代码中）。

推荐阅读