tensorflow的保存模型大小理解（生成模型和打印模型参数）-20201110

tensorflow的保存模型大小理解过程

如何计算tensorflow计算model的大小size（1）
如何计算tensorflow计算model的大小size（2）
https://github.com/Prinsphield/Wechat_AutoJump/issues/56
https://stackoverflow.com/questions/39757985/why-are-there-two-additional-variables-in-the-checkpoint-for-each-layer（这篇说明Adam3倍原因，说实话，还不是很懂）
1.生成模型

import os
import tensorflow as tf
import tensorflow.contrib.slim as slim
 
CURRENT_DIR = os.getcwd()
train_dir  = CURRENT_DIR + '/logs/'
 
a = tf.placeholder(tf.float32, shape=[None, 6, 512], name='a')
b = slim.fully_connected(a, 1 * 512, activation_fn=tf.nn.relu, weights_initializer=slim.variance_scaling_initializer(), scope='fc1')

 
for tv in tf.trainable_variables():
    print (tv.name)
 
w = tf.get_default_graph().get_tensor_by_name("fc1/weights:0")   
b = tf.get_default_graph().get_tensor_by_name("fc1/biases:0")
 
     
with tf.Session() as sess:
    tf.global_variables_initializer().run()
    print(sess.run(b))
    print(sess.run(w))
    print(sess.run(b).shape)
    print(sess.run(w).shape)    
    print(type(sess.run(b)))
    print(type(sess.run(w)[0,0]))   
    checkpoint_path = os.path.join(train_dir,'abc.ckpt')
    saver = tf.train.Saver()
    saver.save(sess, checkpoint_path)
    print（'*********    model saved    *********' )

2.结合adam

import os
import tensorflow as tf
import tensorflow.contrib.slim as slim
from tensorflow.python import pywrap_tensorflow
 
 
 
CURRENT_DIR = os.getcwd()
train_dir  = CURRENT_DIR + '/logs/'
 
lr =0.01
 
a = tf.placeholder(tf.float32, shape=[None, 6, 512], name='a')
print a.shape
b     = slim.fully_connected(a, 1 * 512, activation_fn=tf.nn.relu, weights_initializer=slim.variance_scaling_initializer(), scope='fc1')
print b.shape
  
y = tf.placeholder(tf.float32, [None, 10])
cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=b, labels=y))
    
optimizer = tf.train.AdamOptimizer(lr).minimize(cost)
# optimizer = tf.train.GradientDescentOptimizer(lr).minimize(cost)   # 对比tf.train.GradientDescentOptimizer
 
for tv in tf.trainable_variables():
    print (tv.name)
  
w = tf.get_default_graph().get_tensor_by_name("fc1/weights:0")   
b = tf.get_default_graph().get_tensor_by_name("fc1/biases:0") 
 
  
      
with tf.Session() as sess:
    tf.global_variables_initializer().run()
    print(sess.run(b))
    print(sess.run(w))
    print(sess.run(b).shape)
    print(sess.run(w).shape)    
    print(type(sess.run(b)))
    print(type(sess.run(w)[0,0]))   
    checkpoint_path = os.path.join(train_dir,'abc.ckpt')
    saver = tf.train.Saver()
    saver.save(sess, checkpoint_path)
    print '*********    model saved    *********'  

3.打印模型参数

model_dir = train_dir
 
ckpt = tf.train.get_checkpoint_state(model_dir)
ckpt_path = ckpt.model_checkpoint_path
 
reader = pywrap_tensorflow.NewCheckpointReader(ckpt_path)
param_dict = reader.get_variable_to_shape_map()
 
for key, val in param_dict.items():
    try:
        print key, val
    except:
        pass

4.inspect_checkpoint.py有个报错，
from tensorflow.python.training import py_checkpoint_reader改为from tensorflow.python import pywrap_tensorflow
estimate.py中需要注意空格，params.txt的最后一行等信息，不然会报错

import ast
from six.moves import reduce

def main(filename):
    f = open(filename, 'r')
    storage = 0
    for i, line in enumerate(f):
        if len(line) <= 1: continue
        l = ast.literal_eval(line.split()[-1])
        if len(l) == 0:
            storage += 1 * 4
        else:
            storage += reduce(lambda x,y: x*y, l) * 4
    print(storage)

if __name__ == "__main__":
    filename = 'params.txt'
    main(filename)