When converting tensor to ndarray in tensorflow, the run or Eval function is constantly called in the loop, and the code runs more and more slowly!

problem

   I have such a need ： I currently have a trained encoder Model , Its output is Tensor type , I want to turn it into ndarray type . By searching for information , I found it possible to take advantage of sess.run() hold Tensor Convert to ndarray, So I call... In my code sess.run() Successfully converted data type .
   however , My data transformation will call in every loop , That is, the loop always calls sess.run(), So the question came , Every cycle ,sess.run It takes longer than the last time , This leads to slower and slower training . It takes time from the first call 0.17s To the next 100 When called 0.27s, And this is 100 Time , If you train 10000 Time , I don't know how long it will take , So this problem has to be solved ！

Question why

   If you keep building in a cycle tensorflow Figure if the node is running again , It can lead to tensorflow It's getting slower and slower . concrete problems See the code comments , Uncommented lines of code can be ignored , The problem code is as follows ：

import gym
from gym.spaces import Box
import numpy as np
from tensorflow import keras
import tensorflow as tf
import time

class MyWrapper(gym.ObservationWrapper):
    def __init__(self, env, encoder, latent_dim = 2):
        super().__init__(env)
        self._observation_space = Box(-np.inf, np.inf, shape=(7 + latent_dim,), dtype=np.float32)
        self.observation_space = self._observation_space
        self.encoder = encoder #  This is the model I have trained in advance 
        tf.InteractiveSession()
        self.sess = tf.get_default_session()
		self.sess.run(tf.global_variables_initializer())
	
    def observation(self, obs):
        obs = np.reshape(obs, (1, -1))
        latent_z_tensor = self.encoder(obs)[2] #  That's the problem , This line of code is calling run when , Will continue to create graph nodes , So it's getting slower and slower 
        
        t=time.time() #  Test run time 
        latent_z_arr = sels.sess.run(latent_z_tensor) #  Every time run when , The above diagram will be rebuilt 
        print(time.time()-t) #  Test run time 

        obs = np.reshape(obs, (-1,))
        latent_z_arr = np.reshape(latent_z_arr, (-1,))

        obs = obs.tolist()
        obs.extend(latent_z_arr.tolist())
        obs = np.array(obs)
        return obs

Solutions

At initialization , Just build the graph structure , Use tf.placeholder Placeholders represent obs This variable , Examples of specific schemes are as follows （ You can focus only on the annotated lines ）：

import gym
from gym.spaces import Box
import numpy as np
from tensorflow import keras
import tensorflow as tf
import time

class MyWrapper(gym.ObservationWrapper):
    def __init__(self, env, encoder, latent_dim = 2):
        super().__init__(env)
        self._observation_space = Box(-np.inf, np.inf, shape=(7 + latent_dim,), dtype=np.float32)
        self.observation_space = self._observation_space
        self.encoder = encoder
        tf.InteractiveSession()
        self.sess = tf.get_default_session()
        self.obs=tf.placeholder(dtype=tf.float32,shape=(1,7)) #  The point is that these two lines of code , When initializing, build the diagram first , Use placeholders to represent obs, In actual operation, you only need to feed data obs Just fine 
        self.latent_z_tensor = self.encoder(self.obs)[2] #  Build the diagram at initialization 
        self.sess.run(tf.global_variables_initializer())

    def observation(self, obs):
        obs = np.reshape(obs, (1, -1))
        t=time.time() #  Test run time 
        latent_z_arr = self.sess.run(self.latent_z_tensor, feed_dict={
    self.obs:obs}) #  Just feed the data here , The diagram will not be rebuilt .
        print(time.time()-t) #  Test run time 

        obs = np.reshape(obs, (-1,))
        latent_z_arr = np.reshape(latent_z_arr, (-1,))

        obs = obs.tolist()
        obs.extend(latent_z_arr.tolist())
        obs = np.array(obs)
        return obs

Now? , Data type conversion completed , The slow running of the code also solves the problem ！

This problem has been solved , After four or five days of searching, I finally got it , At this moment , The happiness brought by solving the problem has dispelled the haze of lovelorn in the previous two days , I'm so happy .