TensorFlow model saving and extraction methods

# The procedures in this document are designed to be progressive and in line with the teaching materials and learning progress, please follow the notes in sections.
# When executing, pay attention to the IDE's current work over the path, it is best to restart the controller once per segment, the output results are more accurate
 
 
# Part1: Saving and Loading Neural Network Models via Class Implementation
 
# Take note of the current working path when executing this program.
import tensorflow as tf 
 
v1 = ((1.0, shape=[1]), name="v1") 
v2 = ((2.0, shape=[1]), name="v2") 
result = v1 + v2 
 
saver = () 
 
with () as sess: 
 (tf.global_variables_initializer()) 
 (sess, "Model/") 
 
 
# Part2: Methods for loading TensorFlow models
 
import tensorflow as tf 
 
v1 = ((1.0, shape=[1]), name="v1") 
v2 = ((2.0, shape=[1]), name="v2") 
result = v1 + v2 
 
saver = () 
 
with () as sess: 
 (sess, "./Model/") # Note that the path is preceded by ". /"
 print((result)) # [ 3.] 
 
 
# Part3: If you don't want to duplicate the definition of operations on the computational graph, you can load the already-persistent graph directly.
 
import tensorflow as tf 
 
saver = .import_meta_graph("Model/") 
 
with () as sess: 
 (sess, "./Model/") # Note the way the path is written
 print((tf.get_default_graph().get_tensor_by_name("add:0"))) # [ 3.] 
 
 
# Part4: Classes also support renaming variables at save and load time
 
import tensorflow as tf 
 
# The name of the declared variable name does not match the name of the variable name in the saved model
u1 = ((1.0, shape=[1]), name="other-v1") 
u2 = ((2.0, shape=[1]), name="other-v2") 
result = u1 + u2 
 
# If the Saver class object is lifted directly, it will report that the variable was not found.
# Just rename the variable using a dictionary dict, {"name of saved variable name": rename variable name}
# The variable whose name was v1 is now loaded into the variable u1 (whose name is other-v1)
saver = ({"v1": u1, "v2": u2}) 
 
with () as sess: 
 (sess, "./Model/") 
 print((result)) # [ 3.] 
 
 
# Part5: Preservation of the sliding average model
 
import tensorflow as tf 
 
v = (0, dtype=tf.float32, name="v") 
for variables in tf.global_variables(): 
 print() # v:0 
 
ema = (0.99) 
maintain_averages_op = (tf.global_variables()) 
for variables in tf.global_variables(): 
 print() # v:0 
       # v/ExponentialMovingAverage:0 
 
saver = () 
 
with () as sess: 
 (tf.global_variables_initializer()) 
 ((v, 10)) 
 (maintain_averages_op) 
 (sess, "Model/model_ema.ckpt") 
 print(([v, (v)])) # [10.0, 0.099999905] 
 
 
# Part6: Read sliding averages of variables directly by renaming them
 
import tensorflow as tf 
 
v = (0, dtype=tf.float32, name="v") 
saver = ({"v/ExponentialMovingAverage": v}) 
 
with () as sess: 
 (sess, "./Model/model_ema.ckpt") 
 print((v)) # 0.0999999 
 
 
# Part7: Getting the variable renaming dictionary via the variables_to_restore() function
 
import tensorflow as tf 
 
v = (0, dtype=tf.float32, name="v") 
# Note that the variable name here must be the same as the name of the saved variable.
ema = (0.99) 
print(ema.variables_to_restore()) 
# {'v/ExponentialMovingAverage': < 'v:0' shape=() dtype=float32_ref>} 
# v here is taken from the name of the variable v above name="v"
 
saver = (ema.variables_to_restore()) 
 
with () as sess: 
 (sess, "./Model/model_ema.ckpt") 
 print((v)) # 0.0999999 
 
 
# Part8: Save the variables and their values in the computed graph as constants in a file with the convert_variables_to_constants function.
 
import tensorflow as tf 
from  import graph_util 
 
v1 = ((1.0, shape=[1]), name="v1") 
v2 = ((2.0, shape=[1]), name="v2") 
result = v1 + v2 
 
with () as sess: 
 (tf.global_variables_initializer()) 
 # Export the GraphDef part of the current calculation diagram, i.e. the part of the calculation process from the input layer to the output layer
 graph_def = tf.get_default_graph().as_graph_def() 
 output_graph_def = graph_util.convert_variables_to_constants(sess, 
              graph_def, ['add']) 
 
 with ("Model/combined_model.pb", 'wb') as f: 
  (output_graph_def.SerializeToString()) 
 
 
# Part9: Load a model with variables and their values.
 
import tensorflow as tf 
from  import gfile 
 
with () as sess: 
 model_filename = "Model/combined_model.pb" 
 with (model_filename, 'rb') as f: 
  graph_def = () 
  graph_def.ParseFromString(()) 
 
 result = tf.import_graph_def(graph_def, return_elements=["add:0"]) 
 print((result)) # [array([ 3.], dtype=float32)]