TensorFlow 模型保存/載入
我們在上線使用一個算法模型的時候,首先必須將已經訓練好的模型保存下來。tensorflow保存模型的方式與sklearn不太一樣,sklearn很直接,一個sklearn.externals.joblib的dump與load方法就可以保存與載入使用。而tensorflow由於有graph, operation 這些概念,保存與載入模型稍顯麻煩。由於TensorFlow 的版本一直在更新, 保存模型的方法也發生了改變,在python 環境,和在C++ 環境(移動端) 等不同的平台需要的模型文件也是不也一樣的。
(https://stackoverflow.com/questions/44516609/tensorflow-what-is-the-relationship-between-ckpt-file-and-ckpt-meta-and-ckp有一個解釋如下)
-
the .ckpt file is the old version output of
saver.save(sess)
, which is the equivalent of your.ckpt-data
(see below) -
the "checkpoint" file is only here to tell some TF functions which is the latest checkpoint file.
-
.ckpt-meta
contains the metagraph, i.e. the structure of your computation graph, without the values of the variables (basically what you can see in tensorboard/graph). -
.ckpt-data
contains the values for all the variables, without the structure. To restore a model in python, you'll usually use the meta and data files with (but you can also use the.pb
file):
saver = tf.train.import_meta_graph(path_to_ckpt_meta)
saver.restore(sess, path_to_ckpt_data)
-
I don't know exactly for
.ckpt-index
, I guess it's some kind of index needed internally to map the two previous files correctly. Anyway it's not really necessary usually, you can restore a model with only.ckpt-meta
and.ckpt-data
. -
the
.pb
file can save your whole graph (meta + data). To load and use (but not train) a graph in c++ you'll usually use it, created withfreeze_graph
, which creates the.pb
file from the meta and data. Be careful, (at least in previous TF versions and for some people) the py function provided byfreeze_graph
did not work properly, so you'd have to use the script version. Tensorflow also provides atf.train.Saver.to_proto()
method, but I don't know what it does exactly.
一、基本方法
網上搜索tensorflow模型保存,搜到的大多是基本的方法。即
保存
- 定義變量
- 使用saver.save()方法保存
載入
- 定義變量
- 使用saver.restore()方法載入
如 保存 代碼如下
import tensorflow as tf import numpy as np W = tf.Variable([[1,1,1],[2,2,2]],dtype = tf.float32,name='w') b = tf.Variable([[0,1,2]],dtype = tf.float32,name='b') init = tf.initialize_all_variables() saver = tf.train.Saver() with tf.Session() as sess: sess.run(init) save_path = saver.save(sess,"save/model.ckpt")
載入代碼如下
import tensorflow as tf import numpy as np W = tf.Variable(tf.truncated_normal(shape=(2,3)),dtype = tf.float32,name='w') b = tf.Variable(tf.truncated_normal(shape=(1,3)),dtype = tf.float32,name='b') saver = tf.train.Saver() with tf.Session() as sess: saver.restore(sess,"save/model.ckpt")
二、不需重新定義網絡結構的方法
tf.train.import_meta_graph( meta_graph_or_file, clear_devices=False, import_scope=None, **kwargs )
這個方法可以從文件中將保存的graph的所有節點加載到當前的default graph中,並返回一個saver。也就是說,我們在保存的時候,除了將變量的值保存下來,其實還有將對應graph中的各種節點保存下來,所以模型的結構也同樣被保存下來了。
比如我們想要保存計算最后預測結果的y
,則應該在訓練階段將它添加到collection中。具體代碼如下
保存
### 定義模型 input_x = tf.placeholder(tf.float32, shape=(None, in_dim), name='input_x') input_y = tf.placeholder(tf.float32, shape=(None, out_dim), name='input_y') w1 = tf.Variable(tf.truncated_normal([in_dim, h1_dim], stddev=0.1), name='w1') b1 = tf.Variable(tf.zeros([h1_dim]), name='b1') w2 = tf.Variable(tf.zeros([h1_dim, out_dim]), name='w2') b2 = tf.Variable(tf.zeros([out_dim]), name='b2') keep_prob = tf.placeholder(tf.float32, name='keep_prob') hidden1 = tf.nn.relu(tf.matmul(self.input_x, w1) + b1) hidden1_drop = tf.nn.dropout(hidden1, self.keep_prob) ### 定義預測目標 y = tf.nn.softmax(tf.matmul(hidden1_drop, w2) + b2) # 創建saver saver = tf.train.Saver(...variables...) # 假如需要保存y,以便在預測時使用 tf.add_to_collection('pred_network', y) sess = tf.Session() for step in xrange(1000000): sess.run(train_op) if step % 1000 == 0: # 保存checkpoint, 同時也默認導出一個meta_graph # graph名為'my-model-{global_step}.meta'. saver.save(sess, 'my-model', global_step=step)
載入
with tf.Session() as sess: new_saver = tf.train.import_meta_graph('my-save-dir/my-model-10000.meta') new_saver.restore(sess, 'my-save-dir/my-model-10000') # tf.get_collection() 返回一個list. 但是這里只要第一個參數即可 y = tf.get_collection('pred_network')[0] graph = tf.get_default_graph() # 因為y中有placeholder,所以sess.run(y)的時候還需要用實際待預測的樣本以及相應的參數來填充這些placeholder,而這些需要通過graph的get_operation_by_name方法來獲取。 input_x = graph.get_operation_by_name('input_x').outputs[0] keep_prob = graph.get_operation_by_name('keep_prob').outputs[0] # 使用y進行預測 sess.run(y, feed_dict={input_x:...., keep_prob:1.0})
這里有兩點需要注意的:
一、 saver.restore()時填的文件名,因為在saver.save的時候,每個checkpoint會保存三個文件,如 my-model-10000.meta
, my-model-10000.index
, my-model-10000.data-00000-of-00001
在import_meta_graph
時填的就是meta文件名,我們知道權值都保存在my-model-10000.data-00000-of-00001
這個文件中,但是如果在restore方法中填這個文件名,就會報錯,應該填的是前綴,這個前綴可以使用tf.train.latest_checkpoint(checkpoint_dir)
這個方法獲取。
二、模型的y中有用到placeholder,在sess.run()的時候肯定要feed對應的數據,因此還要根據具體placeholder的名字,從graph中使用get_operation_by_name
方法獲取。