Related
So I'm training a model on a machine with GPU. Of course I save it in the end of the training:
a = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES)
saver = tf.train.Saver(a)
saver.save(sess, save_path)
Now I have one file, but every time I restore the model from the same file I get different numbers in the matrices, and different predictions for the same examples.
I restore the model like this:
saver = tf.train.import_meta_graph('{}.meta'.format(save_path))
sess.run(tf.global_variables_initializer())
saver.restore(sess, save_path)
What is happening here?
When you call sess.run(tf.global_variables_initializer()) after importing the frozen graph, you probably reinitialise some variables that you should not.
Instead, you should initialise only the uninitialised variables. One way to do it would be (credit to this answer)
uninitialized_vars = []
for var in tf.all_variables():
try:
sess.run(var)
except tf.errors.FailedPreconditionError:
uninitialized_vars.append(var)
init_new_vars_op = tf.initialize_variables(uninitialized_vars)
I am trying to save the latest LSTM State from training to be reused during the prediction stage later. The problem I am encountering is that in the TF LSTM model the State is passed around from one training iteration to next via a combination of a placeholder and a numpy array -- neither of which seems to be included in the Graph by default when the session is saved.
To work around this, I am creating a dedicated TF variable to hold the latest version of the state so as to add it to the Session graph, like so:
# latest State from last training iteration:
_, y, ostate, smm = sess.run([train_step, Y, H, summaries], feed_dict=feed_dict)
# now add to TF variable:
savedState = tf.Variable(ostate, dtype=tf.float32, name='savedState')
tf.variables_initializer([savedState]).run()
save_path = saver.save(sess, pathModel + '/my_model.ckpt')
This seems to add the savedState variable to the saved session graph well, and is easily recoverable later with the rest of the Session.
The problem though, is that the only way I have managed to actually use that variable later in the restored Session, is that if I initialize all variables in the session AFTER I recover it (which seems to reset all trained variables, including the weights/biases/etc.!). If I initialize variables first and THEN recover the session (which works fine in terms of preserving the trained varialbes), then I am getting an error that I'm trying to access an uninitialized variable.
I know there is a way to initialize a specific individual varialbe (which i am using while saving it originally) but the problem is that when we recover them, we refer to them by name as strings, we don't just pass the variable itself?!
# This produces an error 'trying to use an uninitialized varialbe
gInit = tf.global_variables_initializer().run()
new_saver = tf.train.import_meta_graph(pathModel + 'my_model.ckpt.meta')
new_saver.restore(sess, pathModel + 'my_model.ckpt')
fullState = sess.run('savedState:0')
What is the right way to get this done? As a workaround, I am currently saving the State to CSV just as a numpy array and then recover it the same way. It works OK, but clearly not the cleanest solution given that every other aspect of saving/restoring the TF session works perfectly.
Any suggestions appreciated!
**EDIT:
Here's the code that works well, as described in the accepted answer below:
# make sure to define the State variable before the Saver variable:
savedState = tf.get_variable('savedState', shape=[BATCHSIZE, CELL_SIZE * LAYERS])
saver = tf.train.Saver(max_to_keep=1)
# last training iteration:
_, y, ostate, smm = sess.run([train_step, Y, H, summaries], feed_dict=feed_dict)
# now save the State and the whole model:
assignOp = tf.assign(savedState, ostate)
sess.run(assignOp)
save_path = saver.save(sess, pathModel + '/my_model.ckpt')
# later on, in some other program, recover the model and the State:
# make sure to initialize all variables BEFORE recovering the model!
gInit = tf.global_variables_initializer().run()
local_saver = tf.train.import_meta_graph(pathModel + 'my_model.ckpt.meta')
local_saver.restore(sess, pathModel + 'my_model.ckpt')
# recover the state from training and get its last dimension
fullState = sess.run('savedState:0')
h = fullState[-1]
h = np.reshape(h, [1, -1])
I haven't tested yet whether this approach unintentionally initializes any other variables in the saved Session, but don't see why it should, since we only run the specific one.
The issue is that creating a new tf.Variable after the Saver was constructed means that the Saver has no knowledge of the new variable. It still gets saved in the metagraph, but not saved in the checkpoint:
import tensorflow as tf
with tf.Graph().as_default():
var_a = tf.get_variable("a", shape=[])
saver = tf.train.Saver()
var_b = tf.get_variable("b", shape=[])
print(saver._var_list) # [<tf.Variable 'a:0' shape=() dtype=float32_ref>]
initializer = tf.global_variables_initializer()
with tf.Session() as session:
session.run([initializer])
saver.save(session, "/tmp/model", global_step=0)
with tf.Graph().as_default():
new_saver = tf.train.import_meta_graph("/tmp/model-0.meta")
print(saver._var_list) # [<tf.Variable 'a:0' shape=() dtype=float32_ref>]
with tf.Session() as session:
new_saver.restore(session, "/tmp/model-0") # Only var_a gets restored!
I've annotated the quick reproduction of your issue above with the variables that the Saver knows about.
Now, the solution is relatively easy. I would suggest creating the Variable before the Saver, then using tf.assign to update its value (make sure you run the op returned by tf.assign). The assigned value will be saved in checkpoints and restored just like other variables.
This could be handled better by the Saver as a special case when None is passed to its var_list constructor argument (i.e. it could pick up new variables automatically). Feel free to open a feature request on Github for this.
I am using tensorflow checkpointing after every 10 epochs using the following code :
checkpoint_dir = os.path.abspath(os.path.join(out_dir, "checkpoints"))
checkpoint_prefix = os.path.join(checkpoint_dir, "model")
...
if current_step % checkpoint_every == 0:
path = saver.save(sess, checkpoint_prefix, global_step=current_step)
print("Saved model checkpoint to {}\n".format(path))
The problem is that, as the new files are getting generated, previous 5 model files are getting deleted automatically.
This is the expected behavior, the docs for tf.train.Saver say that by default the 5 most recent checkpoint files are kept. To adjust that, set max_to_keep the the desired value.
I am a bit of a beginner with tensorflow so please excuse if this is a stupid question and the answer is obvious.
I have created a Tensorflow graph where starting with placeholders for X and y I have optimized some tensors which represent my model. Part of the graph is something where a vector of predictions can be calculated, e.g. for linear regression something like
y_model = tf.add(tf.mul(X,w),d)
y_vals = sess.run(y_model,feed_dict={....})
After training has been completed I have acceptable values for w and d and now I want to save my model for later. Then, in a different python session I want to restore the model so that I can again run
## Starting brand new python session
import tensorflow as tf
## somehow restor the graph and the values here: how????
## so that I can run this:
y_vals = sess.run(y_model,feed_dict={....})
for some different data and get back the y-values.
I want this to work in a way where the graph for calculating the y-values from the placeholders is also stored and restored - as long as the placeholders get fed the correct data, this should work transparently without the user (the one who applies the model) needing to know what the graph looks like).
As far as I understand tf.train.Saver().save(..) only saves the variables but I also want to save the graph. I think that tf.train.export_meta_graph could be relevant here but I do not understand how to use it correctly, the documentation is a bit cryptic to me and the examples do not even use export_meta_graph anywhere.
From the docs, try this:
# Create some variables.
v1 = tf.Variable(..., name="v1")
v2 = tf.Variable(..., name="v2")
...
# Add an op to initialize the variables.
init_op = tf.global_variables_initializer()
# Add ops to save and restore all the variables.
saver = tf.train.Saver()
# Later, launch the model, initialize the variables, do some work, save the
# variables to disk.
with tf.Session() as sess:
sess.run(init_op)
# Do some work with the model.
..
# Save the variables to disk.
save_path = saver.save(sess, "/tmp/model.ckpt")
print("Model saved in file: %s" % save_path)
You can specify the path.
And if you want to restore the model, try:
with tf.Session() as sess:
saver = tf.train.import_meta_graph('/tmp/model.ckpt.meta')
saver.restore(sess, "/tmp/model.ckpt")
Saving Graph in Tensorflow:
import tensorflow as tf
# Create some placeholder variables
x_pl = tf.placeholder(..., name="x")
y_pl = tf.placeholder(..., name="y")
# Add some operation to the Graph
add_op = tf.add(x, y)
with tf.Session() as sess:
# Add variable initializer
init = tf.global_variables_initializer()
# Add ops to save variables to checkpoints
# Unless var_list is specified Saver will save ALL named variables
# in Graph
# Optionally set maximum of 3 latest models to be saved
saver = tf.train.Saver(max_to_keep=3)
# Run variable initializer
sess.run(init)
for i in range(no_steps):
# Feed placeholders with some data and run operation
sess.run(add_op, feed_dict={x_pl: i+1, y_pl: i+5})
saver.save(sess, "path/to/checkpoint/model.ckpt", global_step=i)
This will save the following files:
1) Meta Graph
.meta file:
MetaGraphDef protocol buffer representation of MetaGraph which saves the complete Tf Graph structure i.e. the GraphDef that describes the dataflow and all metadata associated with it e.g. all variables, operations, collections, etc.
importing the graph structure will recreate the Graph and all its variables, then the corresponding values for these variables can be restored from the checkpoint file
if you don't want to restore the Graph however you can reconstruct all of the information in the MetaGraphDef by re-executing the Python code that builds the model n.b. you must recreate the EXACT SAME variables first before restoring their values from the checkpoint
since Meta Graph file is not always needed, you can switch off writing the file in saver.save using write_meta_graph=False
2) Checkpoint files
.data file:
binary file containing VALUES of all saved variables outlined in tf.train.Saver() (default is all variables)
.index file:
immutable table describing all tensors and their metadata checkpoint file:
keeps a record of latest checkpoint files saved
Restoring Graph in Tensorflow:
import tensorflow as tf
latest_checkpoint = tf.train.latest_checkpoint("path/to/checkpoint")
# Load latest checkpoint Graph via import_meta_graph:
# - construct protocol buffer from file content
# - add all nodes to current graph and recreate collections
# - return Saver
saver = tf.train.import_meta_graph(latest_checkpoint + '.meta')
# Start session
with tf.Session() as sess:
# Restore previously trained variables from disk
print("Restoring Model: {}".format("path/to/checkpoint"))
saver.restore(sess, latest_checkpoint)
# Retrieve protobuf graph definition
graph = tf.get_default_graph()
print("Restored Operations from MetaGraph:")
for op in graph.get_operations():
print(op.name)
# Access restored placeholder variables
x_pl = graph.get_tensor_by_name("x_pl:0")
y_pl = graph.get_tensor_by_name("y_pl:0")
# Access restored operation to re run
accuracy_op = graph.get_tensor_by_name("accuracy_op:0")
This is just a quick example with the basics, for a working implementation see here.
In order to save the graph, you need to freeze the graph.
Here is the python script for freezing the graph : https://github.com/tensorflow/tensorflow/blob/master/tensorflow/python/tools/freeze_graph.py
Here is a code snippet for freezing graph:
from tensorflow.python.tools import freeze_graph
freeze_graph.freeze_graph(input_graph_path, input_saver_def_path,
input_binary, checkpoint_path, output_node
restore_op_name, filename_tensor_name,
output_frozen_graph_name, True, "")
where output node corresponds to output tensor variable.
output = tf.nn.softmax(outer_layer_name,name="output")
I'm saving my session state like so:
self._saver = tf.saver()
self._saver.save(self._session, '/network', global_step=self._time)
When I later restore I want to get the value of the global_step for the checkpoint I restore from. This is in order to set some hyper parameters from it.
The hacky way to do this would be to run through and parse the file names in the checkpoint directory. But surly there has to be a better, built in way to do this?
General pattern is to have a global_step variable to keep track of steps
global_step = tf.Variable(0, name='global_step', trainable=False)
train_op = optimizer.minimize(loss, global_step=global_step)
Then you can save with
saver.save(sess, save_path, global_step=global_step)
When you restore, the value of global_step is restored as well
This is a bit of a hack, but the other answers did not work for me at all
ckpt = tf.train.get_checkpoint_state(checkpoint_dir)
#Extract from checkpoint filename
step = int(os.path.basename(ckpt.model_checkpoint_path).split('-')[1])
Update 9/2017
I'm not sure if this started working due to updates, but the following method seems to be effective in getting global_step to update and load properly:
Create two ops. One to hold global_step and another to increment it:
global_step = tf.Variable(0, trainable=False, name='global_step')
increment_global_step = tf.assign_add(global_step,1,
name = 'increment_global_step')
Now in your training loop run the increment op every time you run your training op.
sess.run([train_op,increment_global_step],feed_dict=feed_dict)
If you ever want to retrieve you global step value as an integer at any point, just use the following command after loading the model:
sess.run(global_step)
This can be useful for creating filenames or calculating what your current epoch is without having a second tensorflow Variable for holding that value. For instance, calculating the current epoch on loading would be something like:
loaded_epoch = sess.run(global_step)//(batch_size*num_train_records)
I had the same issue as Lawrence Du, I could not find a way to get the global_step by restoring the model. So I applied his hack to the inception v3 training code in the Tensorflow/models github repo I'm using. The code below also contains a fix related to the pretrained_model_checkpoint_path.
If you have a better solution, or know what I'm missing please leave a comment!
In any case, this code works for me:
...
# When not restoring start at 0
last_step = 0
if FLAGS.pretrained_model_checkpoint_path:
# A model consists of three files, use the base name of the model in
# the checkpoint path. E.g. my-model-path/model.ckpt-291500
#
# Because we need to give the base name you can't assert (will always fail)
# assert tf.gfile.Exists(FLAGS.pretrained_model_checkpoint_path)
variables_to_restore = tf.get_collection(
slim.variables.VARIABLES_TO_RESTORE)
restorer = tf.train.Saver(variables_to_restore)
restorer.restore(sess, FLAGS.pretrained_model_checkpoint_path)
print('%s: Pre-trained model restored from %s' %
(datetime.now(), FLAGS.pretrained_model_checkpoint_path))
# HACK : global step is not restored for some unknown reason
last_step = int(os.path.basename(FLAGS.pretrained_model_checkpoint_path).split('-')[1])
# assign to global step
sess.run(global_step.assign(last_step))
...
for step in range(last_step + 1, FLAGS.max_steps):
...
You can use the global_step variable to keep track of steps, but if in your code, you are initializing or assigning this value to another step variable, it may not be consistent.
For instance, you define your global_step using:
global_step = tf.Variable(0, name='global_step', trainable=False)
Assign to your training operation:
train_op = optimizer.minimize(loss, global_step=global_step)
Save in your checkpoint:
saver.save(sess, checkpoint_path, global_step=global_step)
And restore from your checkpoint:
saver.restore(sess, checkpoint_path)
the value of global_step is restored as well but if you are assigning it to another variable, say step, then you must do something like:
step = global_step.eval(session=sess)
The variable step, contains the last saved global_step in the checkpoint.
It will be nice to also define the global_step from graph than as zero variable (as earlier defined):
global_step = tf.train.get_or_create_global_step()
This will get your last global_step if exist or create one if not.
TL;DR
As tensorflow variable (will be evaluated in the session)
global_step = tf.train.get_or_create_global_step()
# use global_step variable to calculate your hyperparameter
# this variable will be evaluated later in the session
saver = tf.train.Saver()
with tf.Session() as sess:
# restore all variables from checkpoint
saver.restore(sess, checkpoint_path)
# than init table and local variables and start training/evaluation ...
Or: As numpy integer (without any session):
reader = tf.train.NewCheckpointReader(absolute_checkpoint_path)
global_step = reader.get_tensor('global_step')
Long Answer
There are at least two ways retrieving the global from a checkpoint. As tensorflow variable or as numpy integer. Parsing the filename will not work, if the global_step was not provided as a parameter in the save method of the Saver. For pretrained models see the remark at the end of the answer.
As Tensorflow variable
If you need the global_step variable to calculate some hyperparameters you can just use tf.train.get_or_create_global_step(). This will return a tensorflow variable. Because the variable will be evaluated later in the session you can only use tensorflow operations to calculate your hyperparameters. So e.g.: max(global_step, 100) will not work. You have to use tensorflow equivalent tf.maximum(global_step, 100) that can be evaluated later in the session.
Within the session you can initialize the global step variable with a checkpoint using saver.restore(sess, checkpoint_path)
global_step = tf.train.get_or_create_global_step()
# use global_step variable to calculate your hyperparameter
# this variable will be evaluated later in the session
hyper_parameter = tf.maximum(global_step, 100)
saver = tf.train.Saver()
with tf.Session() as sess:
# restore all variables from checkpoint
saver.restore(sess, checkpoint_path)
# than init table and local variables and start training/evaluation ...
# for verification you can print the global step and your hyper parameter
print(sess.run([global_step, hyper_parameter]))
Or: As numpy integer (without session)
If you need the global step variable as scalar without starting a session you can also read this variable directly from your checkpoint file(s). You just need a NewCheckpointReader. Because of a bug in older tensorflow versions you should convert the path of the checkpoint file to an absolute path. With the reader you can get all the tensors of the model as numpy variables.
The name of the global step variable is a constant string tf.GraphKeys.GLOBAL_STEP defined as 'global_step'.
absolute_checkpoint_path = os.path.abspath(checkpoint_path)
reader = tf.train.NewCheckpointReader(absolute_checkpoint_path)
global_step = reader.get_tensor(tf.GraphKeys.GLOBAL_STEP)
Remark to pretrained models: In most pretrained models that are available online the global step is reset to zero. So, these models can be used to initialize the model parameters for finetuning without overwrite the global step.
The current 0.10rc0 version seems to be different, there's no tf.saver() any more. Now it's tf.train.Saver(). Also, the save command adds info onto save_path filename for the global_step, so we can't just call restore on the same save_path since that not the actual save file.
The easiest way I see right now is to use the SessionManager along with a saver like this:
my_checkpoint_dir = "/tmp/checkpoint_dir"
# make a saver to use with SessionManager for restoring
saver = tf.train.Saver()
# Build an initialization operation to run below.
init = tf.initialize_all_variables()
# use a SessionManager to help with automatic variable restoration
sm = tf.train.SessionManager()
# try to find the latest checkpoint in my_checkpoint_dir, then create a session with that restored
# if no such checkpoint, then call the init_op after creating a new session
sess = sm.prepare_session("", init_op=init, saver=saver, checkpoint_dir=my_checkpoint_dir))
That's it. Now you have a session that's either restored from the my_checkpoint_dir (make sure that directory exists before calling this), or if there's no checkpoint there then it creates a new session and calls the init_op to initialize your variables.
When you want to save, you just save to any name you want in that directory and pass the global_step in. Here's an example where I save the step variable in a loop as the global_step, so it comes back to that point if you kill the program and restart it so it restores the checkpoint:
checkpoint_path = os.path.join(my_checkpoint_dir, 'model.ckpt')
saver.save(sess, checkpoint_path, global_step=step)
This creates files in my_checkpoint_dir like "model.ckpt-1000" where 1000 is the global_step passed in. If it keeps running, then you get more like "model.ckpt-2000". The SessionManager above picks up the latest one of these when the program is restarted. The checkpoint_path can be whatever file name you want, as long as it's in the checkpoint_dir. The save() will create that file with the global_step appended (as shown above). It also creates a "checkpoint" index file, which is how the SessionManager then finds the latest save checkpoint.
just note my solution on global step saving and restore.
Save:
global_step = tf.Variable(0, trainable=False, name='global_step')
saver.save(sess, model_path + model_name, global_step=_global_step)
Restore:
if os.path.exists(model_path):
saver.restore(sess, tf.train.latest_checkpoint(model_path))
print("Model restore finished, current globle step: %d" % global_step.eval())
The reason that a variable is not restored as expected is most likely due to the fact that it was created after your tf.Saver() object was created.
The place where you create the tf.Saver() object matters when you don't explicitly specify a var_list, or specify None for var_list. The expected behavior for many programmers is that all variables in the graph are saved when the save() method is called, but this is not the case, and it should perhaps be documented as such. A snapshot of all variables in the graph is saved at the time of object creation.
Unless you're having any performance issues, it's safest to create the saver object right when you decide to save your progress. Otherwise, make sure to create the saver object after you create all your variables.
Also, the global_step that is passed to saver.save(sess, save_path, global_step=global_step) is merely a counter used for creating the filename and has nothing to do with whether it will be restored as a global_step variable. This is a parameter misnomer IMO since if you're saving your progress at the end of each epoch, it's probably best to pass your epoch number for this parameter.