Cannot restore model in tensorflow [duplicate]

Question

After you train a model in Tensorflow:

1. How do you save the trained model?
2. How do you later restore this saved model?

Answer 1

I am improving my answer to add more details for saving and restoring models.

In(and after) Tensorflow version 0.11:

Save the model:

import tensorflow as tf

#Prepare to feed input, i.e. feed_dict and placeholders
w1 = tf.placeholder("float", name="w1")
w2 = tf.placeholder("float", name="w2")
b1= tf.Variable(2.0,name="bias")
feed_dict ={w1:4,w2:8}

#Define a test operation that we will restore
w3 = tf.add(w1,w2)
w4 = tf.multiply(w3,b1,name="op_to_restore")
sess = tf.Session()
sess.run(tf.global_variables_initializer())

#Create a saver object which will save all the variables
saver = tf.train.Saver()

#Run the operation by feeding input
print sess.run(w4,feed_dict)
#Prints 24 which is sum of (w1+w2)*b1 

#Now, save the graph
saver.save(sess, 'my_test_model',global_step=1000)

Restore the model:

import tensorflow as tf

sess=tf.Session()    
#First let's load meta graph and restore weights
saver = tf.train.import_meta_graph('my_test_model-1000.meta')
saver.restore(sess,tf.train.latest_checkpoint('./'))


# Access saved Variables directly
print(sess.run('bias:0'))
# This will print 2, which is the value of bias that we saved


# Now, let's access and create placeholders variables and
# create feed-dict to feed new data

graph = tf.get_default_graph()
w1 = graph.get_tensor_by_name("w1:0")
w2 = graph.get_tensor_by_name("w2:0")
feed_dict ={w1:13.0,w2:17.0}

#Now, access the op that you want to run. 
op_to_restore = graph.get_tensor_by_name("op_to_restore:0")

print sess.run(op_to_restore,feed_dict)
#This will print 60 which is calculated 

This and some more advanced use-cases have been explained very well here.

A quick complete tutorial to save and restore Tensorflow models

Answer 2

In (and after) TensorFlow version 0.11.0RC1, you can save and restore your model directly by calling tf.train.export_meta_graph and tf.train.import_meta_graph according to https://www.tensorflow.org/programmers_guide/meta_graph.

Save the model

w1 = tf.Variable(tf.truncated_normal(shape=[10]), name='w1')
w2 = tf.Variable(tf.truncated_normal(shape=[20]), name='w2')
tf.add_to_collection('vars', w1)
tf.add_to_collection('vars', w2)
saver = tf.train.Saver()
sess = tf.Session()
sess.run(tf.global_variables_initializer())
saver.save(sess, 'my-model')
# `save` method will call `export_meta_graph` implicitly.
# you will get saved graph files:my-model.meta

Restore the model

sess = tf.Session()
new_saver = tf.train.import_meta_graph('my-model.meta')
new_saver.restore(sess, tf.train.latest_checkpoint('./'))
all_vars = tf.get_collection('vars')
for v in all_vars:
    v_ = sess.run(v)
    print(v_)

Answer 3

Tensorflow 2 Docs

Saving Checkpoints

Adapted from the docs

# -------------------------
# -----  Toy Context  -----
# -------------------------
import tensorflow as tf


class Net(tf.keras.Model):
    """A simple linear model."""

    def __init__(self):
        super(Net, self).__init__()
        self.l1 = tf.keras.layers.Dense(5)

    def call(self, x):
        return self.l1(x)


def toy_dataset():
    inputs = tf.range(10.0)[:, None]
    labels = inputs * 5.0 + tf.range(5.0)[None, :]
    return (
        tf.data.Dataset.from_tensor_slices(dict(x=inputs, y=labels)).repeat().batch(2)
    )


def train_step(net, example, optimizer):
    """Trains `net` on `example` using `optimizer`."""
    with tf.GradientTape() as tape:
        output = net(example["x"])
        loss = tf.reduce_mean(tf.abs(output - example["y"]))
    variables = net.trainable_variables
    gradients = tape.gradient(loss, variables)
    optimizer.apply_gradients(zip(gradients, variables))
    return loss


# ----------------------------
# -----  Create Objects  -----
# ----------------------------

net = Net()
opt = tf.keras.optimizers.Adam(0.1)
dataset = toy_dataset()
iterator = iter(dataset)
ckpt = tf.train.Checkpoint(
    step=tf.Variable(1), optimizer=opt, net=net, iterator=iterator
)
manager = tf.train.CheckpointManager(ckpt, "./tf_ckpts", max_to_keep=3)

# ----------------------------
# -----  Train and Save  -----
# ----------------------------

ckpt.restore(manager.latest_checkpoint)
if manager.latest_checkpoint:
    print("Restored from {}".format(manager.latest_checkpoint))
else:
    print("Initializing from scratch.")

for _ in range(50):
    example = next(iterator)
    loss = train_step(net, example, opt)
    ckpt.step.assign_add(1)
    if int(ckpt.step) % 10 == 0:
        save_path = manager.save()
        print("Saved checkpoint for step {}: {}".format(int(ckpt.step), save_path))
        print("loss {:1.2f}".format(loss.numpy()))


# ---------------------
# -----  Restore  -----
# ---------------------

# In another script, re-initialize objects
opt = tf.keras.optimizers.Adam(0.1)
net = Net()
dataset = toy_dataset()
iterator = iter(dataset)
ckpt = tf.train.Checkpoint(
    step=tf.Variable(1), optimizer=opt, net=net, iterator=iterator
)
manager = tf.train.CheckpointManager(ckpt, "./tf_ckpts", max_to_keep=3)

# Re-use the manager code above ^

ckpt.restore(manager.latest_checkpoint)
if manager.latest_checkpoint:
    print("Restored from {}".format(manager.latest_checkpoint))
else:
    print("Initializing from scratch.")

for _ in range(50):
    example = next(iterator)
    # Continue training or evaluate etc.

More links

• exhaustive and useful tutorial on saved_model -> https://www.tensorflow.org/guide/saved_model

• keras detailed guide to save models -> https://www.tensorflow.org/guide/keras/save_and_serialize

Checkpoints capture the exact value of all parameters (tf.Variable objects) used by a model. Checkpoints do not contain any description of the computation defined by the model and thus are typically only useful when source code that will use the saved parameter values is available.

The SavedModel format on the other hand includes a serialized description of the computation defined by the model in addition to the parameter values (checkpoint). Models in this format are independent of the source code that created the model. They are thus suitable for deployment via TensorFlow Serving, TensorFlow Lite, TensorFlow.js, or programs in other programming languages (the C, C++, Java, Go, Rust, C# etc. TensorFlow APIs).

(Highlights are my own)

---

Tensorflow < 2

---

From the docs:

Save

# Create some variables.
v1 = tf.get_variable("v1", shape=[3], initializer = tf.zeros_initializer)
v2 = tf.get_variable("v2", shape=[5], initializer = tf.zeros_initializer)

inc_v1 = v1.assign(v1+1)
dec_v2 = v2.assign(v2-1)

# 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, and save the
# variables to disk.
with tf.Session() as sess:
  sess.run(init_op)
  # Do some work with the model.
  inc_v1.op.run()
  dec_v2.op.run()
  # Save the variables to disk.
  save_path = saver.save(sess, "/tmp/model.ckpt")
  print("Model saved in path: %s" % save_path)

Restore

tf.reset_default_graph()

# Create some variables.
v1 = tf.get_variable("v1", shape=[3])
v2 = tf.get_variable("v2", shape=[5])

# Add ops to save and restore all the variables.
saver = tf.train.Saver()

# Later, launch the model, use the saver to restore variables from disk, and
# do some work with the model.
with tf.Session() as sess:
  # Restore variables from disk.
  saver.restore(sess, "/tmp/model.ckpt")
  print("Model restored.")
  # Check the values of the variables
  print("v1 : %s" % v1.eval())
  print("v2 : %s" % v2.eval())

simple_save

Many good answer, for completeness I'll add my 2 cents: simple_save. Also a standalone code example using the tf.data.Dataset API.

Python 3 ; Tensorflow 1.14

import tensorflow as tf
from tensorflow.saved_model import tag_constants

with tf.Graph().as_default():
    with tf.Session() as sess:
        ...

        # Saving
        inputs = {
            "batch_size_placeholder": batch_size_placeholder,
            "features_placeholder": features_placeholder,
            "labels_placeholder": labels_placeholder,
        }
        outputs = {"prediction": model_output}
        tf.saved_model.simple_save(
            sess, 'path/to/your/location/', inputs, outputs
        )

Restoring:

graph = tf.Graph()
with restored_graph.as_default():
    with tf.Session() as sess:
        tf.saved_model.loader.load(
            sess,
            [tag_constants.SERVING],
            'path/to/your/location/',
        )
        batch_size_placeholder = graph.get_tensor_by_name('batch_size_placeholder:0')
        features_placeholder = graph.get_tensor_by_name('features_placeholder:0')
        labels_placeholder = graph.get_tensor_by_name('labels_placeholder:0')
        prediction = restored_graph.get_tensor_by_name('dense/BiasAdd:0')

        sess.run(prediction, feed_dict={
            batch_size_placeholder: some_value,
            features_placeholder: some_other_value,
            labels_placeholder: another_value
        })

Standalone example

Original blog post

The following code generates random data for the sake of the demonstration.

1. We start by creating the placeholders. They will hold the data at runtime. From them, we create the Dataset and then its Iterator. We get the iterator's generated tensor, called input_tensor which will serve as input to our model.
2. The model itself is built from input_tensor: a GRU-based bidirectional RNN followed by a dense classifier. Because why not.
3. The loss is a softmax_cross_entropy_with_logits, optimized with Adam. After 2 epochs (of 2 batches each), we save the "trained" model with tf.saved_model.simple_save. If you run the code as is, then the model will be saved in a folder called simple/ in your current working directory.
4. In a new graph, we then restore the saved model with tf.saved_model.loader.load. We grab the placeholders and logits with graph.get_tensor_by_name and the Iterator initializing operation with graph.get_operation_by_name.
5. Lastly we run an inference for both batches in the dataset, and check that the saved and restored model both yield the same values. They do!

Code:

import os
import shutil
import numpy as np
import tensorflow as tf
from tensorflow.python.saved_model import tag_constants


def model(graph, input_tensor):
    """Create the model which consists of
    a bidirectional rnn (GRU(10)) followed by a dense classifier

    Args:
        graph (tf.Graph): Tensors' graph
        input_tensor (tf.Tensor): Tensor fed as input to the model

    Returns:
        tf.Tensor: the model's output layer Tensor
    """
    cell = tf.nn.rnn_cell.GRUCell(10)
    with graph.as_default():
        ((fw_outputs, bw_outputs), (fw_state, bw_state)) = tf.nn.bidirectional_dynamic_rnn(
            cell_fw=cell,
            cell_bw=cell,
            inputs=input_tensor,
            sequence_length=[10] * 32,
            dtype=tf.float32,
            swap_memory=True,
            scope=None)
        outputs = tf.concat((fw_outputs, bw_outputs), 2)
        mean = tf.reduce_mean(outputs, axis=1)
        dense = tf.layers.dense(mean, 5, activation=None)

        return dense


def get_opt_op(graph, logits, labels_tensor):
    """Create optimization operation from model's logits and labels

    Args:
        graph (tf.Graph): Tensors' graph
        logits (tf.Tensor): The model's output without activation
        labels_tensor (tf.Tensor): Target labels

    Returns:
        tf.Operation: the operation performing a stem of Adam optimizer
    """
    with graph.as_default():
        with tf.variable_scope('loss'):
            loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(
                    logits=logits, labels=labels_tensor, name='xent'),
                    name="mean-xent"
                    )
        with tf.variable_scope('optimizer'):
            opt_op = tf.train.AdamOptimizer(1e-2).minimize(loss)
        return opt_op


if __name__ == '__main__':
    # Set random seed for reproducibility
    # and create synthetic data
    np.random.seed(0)
    features = np.random.randn(64, 10, 30)
    labels = np.eye(5)[np.random.randint(0, 5, (64,))]

    graph1 = tf.Graph()
    with graph1.as_default():
        # Random seed for reproducibility
        tf.set_random_seed(0)
        # Placeholders
        batch_size_ph = tf.placeholder(tf.int64, name='batch_size_ph')
        features_data_ph = tf.placeholder(tf.float32, [None, None, 30], 'features_data_ph')
        labels_data_ph = tf.placeholder(tf.int32, [None, 5], 'labels_data_ph')
        # Dataset
        dataset = tf.data.Dataset.from_tensor_slices((features_data_ph, labels_data_ph))
        dataset = dataset.batch(batch_size_ph)
        iterator = tf.data.Iterator.from_structure(dataset.output_types, dataset.output_shapes)
        dataset_init_op = iterator.make_initializer(dataset, name='dataset_init')
        input_tensor, labels_tensor = iterator.get_next()

        # Model
        logits = model(graph1, input_tensor)
        # Optimization
        opt_op = get_opt_op(graph1, logits, labels_tensor)

        with tf.Session(graph=graph1) as sess:
            # Initialize variables
            tf.global_variables_initializer().run(session=sess)
            for epoch in range(3):
                batch = 0
                # Initialize dataset (could feed epochs in Dataset.repeat(epochs))
                sess.run(
                    dataset_init_op,
                    feed_dict={
                        features_data_ph: features,
                        labels_data_ph: labels,
                        batch_size_ph: 32
                    })
                values = []
                while True:
                    try:
                        if epoch < 2:
                            # Training
                            _, value = sess.run([opt_op, logits])
                            print('Epoch {}, batch {} | Sample value: {}'.format(epoch, batch, value[0]))
                            batch += 1
                        else:
                            # Final inference
                            values.append(sess.run(logits))
                            print('Epoch {}, batch {} | Final inference | Sample value: {}'.format(epoch, batch, values[-1][0]))
                            batch += 1
                    except tf.errors.OutOfRangeError:
                        break
            # Save model state
            print('\nSaving...')
            cwd = os.getcwd()
            path = os.path.join(cwd, 'simple')
            shutil.rmtree(path, ignore_errors=True)
            inputs_dict = {
                "batch_size_ph": batch_size_ph,
                "features_data_ph": features_data_ph,
                "labels_data_ph": labels_data_ph
            }
            outputs_dict = {
                "logits": logits
            }
            tf.saved_model.simple_save(
                sess, path, inputs_dict, outputs_dict
            )
            print('Ok')
    # Restoring
    graph2 = tf.Graph()
    with graph2.as_default():
        with tf.Session(graph=graph2) as sess:
            # Restore saved values
            print('\nRestoring...')
            tf.saved_model.loader.load(
                sess,
                [tag_constants.SERVING],
                path
            )
            print('Ok')
            # Get restored placeholders
            labels_data_ph = graph2.get_tensor_by_name('labels_data_ph:0')
            features_data_ph = graph2.get_tensor_by_name('features_data_ph:0')
            batch_size_ph = graph2.get_tensor_by_name('batch_size_ph:0')
            # Get restored model output
            restored_logits = graph2.get_tensor_by_name('dense/BiasAdd:0')
            # Get dataset initializing operation
            dataset_init_op = graph2.get_operation_by_name('dataset_init')

            # Initialize restored dataset
            sess.run(
                dataset_init_op,
                feed_dict={
                    features_data_ph: features,
                    labels_data_ph: labels,
                    batch_size_ph: 32
                }

            )
            # Compute inference for both batches in dataset
            restored_values = []
            for i in range(2):
                restored_values.append(sess.run(restored_logits))
                print('Restored values: ', restored_values[i][0])

    # Check if original inference and restored inference are equal
    valid = all((v == rv).all() for v, rv in zip(values, restored_values))
    print('\nInferences match: ', valid)

This will print:

$ python3 save_and_restore.py

Epoch 0, batch 0 | Sample value: [-0.13851789 -0.3087595   0.12804556  0.20013677 -0.08229901]
Epoch 0, batch 1 | Sample value: [-0.00555491 -0.04339041 -0.05111827 -0.2480045  -0.00107776]
Epoch 1, batch 0 | Sample value: [-0.19321944 -0.2104792  -0.00602257  0.07465433  0.11674127]
Epoch 1, batch 1 | Sample value: [-0.05275984  0.05981954 -0.15913513 -0.3244143   0.10673307]
Epoch 2, batch 0 | Final inference | Sample value: [-0.26331693 -0.13013336 -0.12553    -0.04276478  0.2933622 ]
Epoch 2, batch 1 | Final inference | Sample value: [-0.07730117  0.11119192 -0.20817074 -0.35660955  0.16990358]

Saving...
INFO:tensorflow:Assets added to graph.
INFO:tensorflow:No assets to write.
INFO:tensorflow:SavedModel written to: b'/some/path/simple/saved_model.pb'
Ok

Restoring...
INFO:tensorflow:Restoring parameters from b'/some/path/simple/variables/variables'
Ok
Restored values:  [-0.26331693 -0.13013336 -0.12553    -0.04276478  0.2933622 ]
Restored values:  [-0.07730117  0.11119192 -0.20817074 -0.35660955  0.16990358]

Inferences match:  True

Answer 4

For TensorFlow version < 0.11.0RC1:

The checkpoints that are saved contain values for the Variables in your model, not the model/graph itself, which means that the graph should be the same when you restore the checkpoint.

Here's an example for a linear regression where there's a training loop that saves variable checkpoints and an evaluation section that will restore variables saved in a prior run and compute predictions. Of course, you can also restore variables and continue training if you'd like.

x = tf.placeholder(tf.float32)
y = tf.placeholder(tf.float32)

w = tf.Variable(tf.zeros([1, 1], dtype=tf.float32))
b = tf.Variable(tf.ones([1, 1], dtype=tf.float32))
y_hat = tf.add(b, tf.matmul(x, w))

...more setup for optimization and what not...

saver = tf.train.Saver()  # defaults to saving all variables - in this case w and b

with tf.Session() as sess:
    sess.run(tf.initialize_all_variables())
    if FLAGS.train:
        for i in xrange(FLAGS.training_steps):
            ...training loop...
            if (i + 1) % FLAGS.checkpoint_steps == 0:
                saver.save(sess, FLAGS.checkpoint_dir + 'model.ckpt',
                           global_step=i+1)
    else:
        # Here's where you're restoring the variables w and b.
        # Note that the graph is exactly as it was when the variables were
        # saved in a prior training run.
        ckpt = tf.train.get_checkpoint_state(FLAGS.checkpoint_dir)
        if ckpt and ckpt.model_checkpoint_path:
            saver.restore(sess, ckpt.model_checkpoint_path)
        else:
            ...no checkpoint found...

        # Now you can run the model to get predictions
        batch_x = ...load some data...
        predictions = sess.run(y_hat, feed_dict={x: batch_x})

Here are the docs for Variables, which cover saving and restoring. And here are the docs for the Saver.

Answer 5

My environment: Python 3.6, Tensorflow 1.3.0

Though there have been many solutions, most of them is based on tf.train.Saver. When we load a .ckpt saved by Saver, we have to either redefine the tensorflow network or use some weird and hard-remembered name, e.g. 'placehold_0:0','dense/Adam/Weight:0'. Here I recommend to use tf.saved_model, one simplest example given below, your can learn more from Serving a TensorFlow Model:

Save the model:

import tensorflow as tf

# define the tensorflow network and do some trains
x = tf.placeholder("float", name="x")
w = tf.Variable(2.0, name="w")
b = tf.Variable(0.0, name="bias")

h = tf.multiply(x, w)
y = tf.add(h, b, name="y")
sess = tf.Session()
sess.run(tf.global_variables_initializer())

# save the model
export_path =  './savedmodel'
builder = tf.saved_model.builder.SavedModelBuilder(export_path)

tensor_info_x = tf.saved_model.utils.build_tensor_info(x)
tensor_info_y = tf.saved_model.utils.build_tensor_info(y)

prediction_signature = (
  tf.saved_model.signature_def_utils.build_signature_def(
      inputs={'x_input': tensor_info_x},
      outputs={'y_output': tensor_info_y},
      method_name=tf.saved_model.signature_constants.PREDICT_METHOD_NAME))

builder.add_meta_graph_and_variables(
  sess, [tf.saved_model.tag_constants.SERVING],
  signature_def_map={
      tf.saved_model.signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY:
          prediction_signature 
  },
  )
builder.save()

Load the model:

import tensorflow as tf
sess=tf.Session() 
signature_key = tf.saved_model.signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY
input_key = 'x_input'
output_key = 'y_output'

export_path =  './savedmodel'
meta_graph_def = tf.saved_model.loader.load(
           sess,
          [tf.saved_model.tag_constants.SERVING],
          export_path)
signature = meta_graph_def.signature_def

x_tensor_name = signature[signature_key].inputs[input_key].name
y_tensor_name = signature[signature_key].outputs[output_key].name

x = sess.graph.get_tensor_by_name(x_tensor_name)
y = sess.graph.get_tensor_by_name(y_tensor_name)

y_out = sess.run(y, {x: 3.0})

Answer 6

There are two parts to the model, the model definition, saved by Supervisor as graph.pbtxt in the model directory and the numerical values of tensors, saved into checkpoint files like model.ckpt-1003418.

The model definition can be restored using tf.import_graph_def, and the weights are restored using Saver.

However, Saver uses special collection holding list of variables that's attached to the model Graph, and this collection is not initialized using import_graph_def, so you can't use the two together at the moment (it's on our roadmap to fix). For now, you have to use approach of Ryan Sepassi -- manually construct a graph with identical node names, and use Saver to load the weights into it.

(Alternatively you could hack it by using by using import_graph_def, creating variables manually, and using tf.add_to_collection(tf.GraphKeys.VARIABLES, variable) for each variable, then using Saver)

Answer 7

You can also take this easier way.

Step 1: initialize all your variables

W1 = tf.Variable(tf.truncated_normal([6, 6, 1, K], stddev=0.1), name="W1")
B1 = tf.Variable(tf.constant(0.1, tf.float32, [K]), name="B1")

Similarly, W2, B2, W3, .....

Step 2: save the session inside model Saver and save it

model_saver = tf.train.Saver()

# Train the model and save it in the end
model_saver.save(session, "saved_models/CNN_New.ckpt")

Step 3: restore the model

with tf.Session(graph=graph_cnn) as session:
    model_saver.restore(session, "saved_models/CNN_New.ckpt")
    print("Model restored.") 
    print('Initialized')

Step 4: check your variable

W1 = session.run(W1)
print(W1)

---

While running in different python instance, use

with tf.Session() as sess:
    # Restore latest checkpoint
    saver.restore(sess, tf.train.latest_checkpoint('saved_model/.'))

    # Initalize the variables
    sess.run(tf.global_variables_initializer())

    # Get default graph (supply your custom graph if you have one)
    graph = tf.get_default_graph()

    # It will give tensor object
    W1 = graph.get_tensor_by_name('W1:0')

    # To get the value (numpy array)
    W1_value = session.run(W1)

Answer 8

In most cases, saving and restoring from disk using a tf.train.Saver is your best option:

... # build your model
saver = tf.train.Saver()

with tf.Session() as sess:
    ... # train the model
    saver.save(sess, "/tmp/my_great_model")

with tf.Session() as sess:
    saver.restore(sess, "/tmp/my_great_model")
    ... # use the model

You can also save/restore the graph structure itself (see the MetaGraph documentation for details). By default, the Saver saves the graph structure into a .meta file. You can call import_meta_graph() to restore it. It restores the graph structure and returns a Saver that you can use to restore the model's state:

saver = tf.train.import_meta_graph("/tmp/my_great_model.meta")

with tf.Session() as sess:
    saver.restore(sess, "/tmp/my_great_model")
    ... # use the model

However, there are cases where you need something much faster. For example, if you implement early stopping, you want to save checkpoints every time the model improves during training (as measured on the validation set), then if there is no progress for some time, you want to roll back to the best model. If you save the model to disk every time it improves, it will tremendously slow down training. The trick is to save the variable states to memory, then just restore them later:

... # build your model

# get a handle on the graph nodes we need to save/restore the model
graph = tf.get_default_graph()
gvars = graph.get_collection(tf.GraphKeys.GLOBAL_VARIABLES)
assign_ops = [graph.get_operation_by_name(v.op.name + "/Assign") for v in gvars]
init_values = [assign_op.inputs[1] for assign_op in assign_ops]

with tf.Session() as sess:
    ... # train the model

    # when needed, save the model state to memory
    gvars_state = sess.run(gvars)

    # when needed, restore the model state
    feed_dict = {init_value: val
                 for init_value, val in zip(init_values, gvars_state)}
    sess.run(assign_ops, feed_dict=feed_dict)

A quick explanation: when you create a variable X, TensorFlow automatically creates an assignment operation X/Assign to set the variable's initial value. Instead of creating placeholders and extra assignment ops (which would just make the graph messy), we just use these existing assignment ops. The first input of each assignment op is a reference to the variable it is supposed to initialize, and the second input (assign_op.inputs[1]) is the initial value. So in order to set any value we want (instead of the initial value), we need to use a feed_dict and replace the initial value. Yes, TensorFlow lets you feed a value for any op, not just for placeholders, so this works fine.

Answer 9

As Yaroslav said, you can hack restoring from a graph_def and checkpoint by importing the graph, manually creating variables, and then using a Saver.

I implemented this for my personal use, so I though I'd share the code here.

Link: https://gist.github.com/nikitakit/6ef3b72be67b86cb7868

(This is, of course, a hack, and there is no guarantee that models saved this way will remain readable in future versions of TensorFlow.)

Answer 10

If it is an internally saved model, you just specify a restorer for all variables as

restorer = tf.train.Saver(tf.all_variables())

and use it to restore variables in a current session:

restorer.restore(self._sess, model_file)

For the external model you need to specify the mapping from the its variable names to your variable names. You can view the model variable names using the command

python /path/to/tensorflow/tensorflow/python/tools/inspect_checkpoint.py --file_name=/path/to/pretrained_model/model.ckpt

The inspect_checkpoint.py script can be found in './tensorflow/python/tools' folder of the Tensorflow source.

To specify the mapping, you can use my Tensorflow-Worklab, which contains a set of classes and scripts to train and retrain different models. It includes an example of retraining ResNet models, located here

Answer 11

Here's my simple solution for the two basic cases differing on whether you want to load the graph from file or build it during runtime.

This answer holds for Tensorflow 0.12+ (including 1.0).

Rebuilding the graph in code

Saving

graph = ... # build the graph
saver = tf.train.Saver()  # create the saver after the graph
with ... as sess:  # your session object
    saver.save(sess, 'my-model')

Loading

graph = ... # build the graph
saver = tf.train.Saver()  # create the saver after the graph
with ... as sess:  # your session object
    saver.restore(sess, tf.train.latest_checkpoint('./'))
    # now you can use the graph, continue training or whatever

Loading also the graph from a file

When using this technique, make sure all your layers/variables have explicitly set unique names. Otherwise Tensorflow will make the names unique itself and they'll be thus different from the names stored in the file. It's not a problem in the previous technique, because the names are "mangled" the same way in both loading and saving.

Saving

graph = ... # build the graph

for op in [ ... ]:  # operators you want to use after restoring the model
    tf.add_to_collection('ops_to_restore', op)

saver = tf.train.Saver()  # create the saver after the graph
with ... as sess:  # your session object
    saver.save(sess, 'my-model')

Loading

with ... as sess:  # your session object
    saver = tf.train.import_meta_graph('my-model.meta')
    saver.restore(sess, tf.train.latest_checkpoint('./'))
    ops = tf.get_collection('ops_to_restore')  # here are your operators in the same order in which you saved them to the collection

Answer 12

tf.keras Model saving with TF2.0

I see great answers for saving models using TF1.x. I want to provide couple of more pointers in saving tensorflow.keras models which is a little complicated as there are many ways to save a model.

Here I am providing an example of saving a tensorflow.keras model to model_path folder under current directory. This works well with most recent tensorflow (TF2.0). I will update this description if there is any change in near future.

Saving and loading entire model

import tensorflow as tf
from tensorflow import keras
mnist = tf.keras.datasets.mnist

#import data
(x_train, y_train),(x_test, y_test) = mnist.load_data()
x_train, x_test = x_train / 255.0, x_test / 255.0

# create a model
def create_model():
  model = tf.keras.models.Sequential([
    tf.keras.layers.Flatten(input_shape=(28, 28)),
    tf.keras.layers.Dense(512, activation=tf.nn.relu),
    tf.keras.layers.Dropout(0.2),
    tf.keras.layers.Dense(10, activation=tf.nn.softmax)
    ])
# compile the model
  model.compile(optimizer='adam',
              loss='sparse_categorical_crossentropy',
              metrics=['accuracy'])
  return model

# Create a basic model instance
model=create_model()

model.fit(x_train, y_train, epochs=1)
loss, acc = model.evaluate(x_test, y_test,verbose=1)
print("Original model, accuracy: {:5.2f}%".format(100*acc))

# Save entire model to a HDF5 file
model.save('./model_path/my_model.h5')

# Recreate the exact same model, including weights and optimizer.
new_model = keras.models.load_model('./model_path/my_model.h5')
loss, acc = new_model.evaluate(x_test, y_test)
print("Restored model, accuracy: {:5.2f}%".format(100*acc))

Saving and loading model Weights only

If you are interested in saving model weights only and then load weights to restore the model, then

model.fit(x_train, y_train, epochs=5)
loss, acc = model.evaluate(x_test, y_test,verbose=1)
print("Original model, accuracy: {:5.2f}%".format(100*acc))

# Save the weights
model.save_weights('./checkpoints/my_checkpoint')

# Restore the weights
model = create_model()
model.load_weights('./checkpoints/my_checkpoint')

loss,acc = model.evaluate(x_test, y_test)
print("Restored model, accuracy: {:5.2f}%".format(100*acc))

Saving and restoring using keras checkpoint callback

# include the epoch in the file name. (uses `str.format`)
checkpoint_path = "training_2/cp-{epoch:04d}.ckpt"
checkpoint_dir = os.path.dirname(checkpoint_path)

cp_callback = tf.keras.callbacks.ModelCheckpoint(
    checkpoint_path, verbose=1, save_weights_only=True,
    # Save weights, every 5-epochs.
    period=5)

model = create_model()
model.save_weights(checkpoint_path.format(epoch=0))
model.fit(train_images, train_labels,
          epochs = 50, callbacks = [cp_callback],
          validation_data = (test_images,test_labels),
          verbose=0)

latest = tf.train.latest_checkpoint(checkpoint_dir)

new_model = create_model()
new_model.load_weights(latest)
loss, acc = new_model.evaluate(test_images, test_labels)
print("Restored model, accuracy: {:5.2f}%".format(100*acc))

saving model with custom metrics

import tensorflow as tf
from tensorflow import keras
mnist = tf.keras.datasets.mnist

(x_train, y_train),(x_test, y_test) = mnist.load_data()
x_train, x_test = x_train / 255.0, x_test / 255.0

# Custom Loss1 (for example) 
@tf.function() 
def customLoss1(yTrue,yPred):
  return tf.reduce_mean(yTrue-yPred) 

# Custom Loss2 (for example) 
@tf.function() 
def customLoss2(yTrue, yPred):
  return tf.reduce_mean(tf.square(tf.subtract(yTrue,yPred))) 

def create_model():
  model = tf.keras.models.Sequential([
    tf.keras.layers.Flatten(input_shape=(28, 28)),
    tf.keras.layers.Dense(512, activation=tf.nn.relu),  
    tf.keras.layers.Dropout(0.2),
    tf.keras.layers.Dense(10, activation=tf.nn.softmax)
    ])
  model.compile(optimizer='adam',
              loss='sparse_categorical_crossentropy',
              metrics=['accuracy', customLoss1, customLoss2])
  return model

# Create a basic model instance
model=create_model()

# Fit and evaluate model 
model.fit(x_train, y_train, epochs=1)
loss, acc,loss1, loss2 = model.evaluate(x_test, y_test,verbose=1)
print("Original model, accuracy: {:5.2f}%".format(100*acc))

model.save("./model.h5")

new_model=tf.keras.models.load_model("./model.h5",custom_objects={'customLoss1':customLoss1,'customLoss2':customLoss2})

Saving keras model with custom ops

When we have custom ops as in the following case (tf.tile), we need to create a function and wrap with a Lambda layer. Otherwise, model cannot be saved.

import numpy as np
import tensorflow as tf
from tensorflow.keras.layers import Input, Lambda
from tensorflow.keras import Model

def my_fun(a):
  out = tf.tile(a, (1, tf.shape(a)[0]))
  return out

a = Input(shape=(10,))
#out = tf.tile(a, (1, tf.shape(a)[0]))
out = Lambda(lambda x : my_fun(x))(a)
model = Model(a, out)

x = np.zeros((50,10), dtype=np.float32)
print(model(x).numpy())

model.save('my_model.h5')

#load the model
new_model=tf.keras.models.load_model("my_model.h5")

I think I have covered a few of the many ways of saving tf.keras model. However, there are many other ways. Please comment below if you see your use case is not covered above. Thanks!

Answer 13

You can also check out examples in TensorFlow/skflow, which offers save and restore methods that can help you easily manage your models. It has parameters that you can also control how frequently you want to back up your model.

Answer 14

If you use tf.train.MonitoredTrainingSession as the default session, you don't need to add extra code to do save/restore things. Just pass a checkpoint dir name to MonitoredTrainingSession's constructor, it will use session hooks to handle these.

Answer 15

All the answers here are great, but I want to add two things.

First, to elaborate on @user7505159's answer, the "./" can be important to add to the beginning of the file name that you are restoring.

For example, you can save a graph with no "./" in the file name like so:

# Some graph defined up here with specific names

saver = tf.train.Saver()
save_file = 'model.ckpt'

with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())
    saver.save(sess, save_file)

But in order to restore the graph, you may need to prepend a "./" to the file_name:

# Same graph defined up here

saver = tf.train.Saver()
save_file = './' + 'model.ckpt' # String addition used for emphasis

with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())
    saver.restore(sess, save_file)

You will not always need the "./", but it can cause problems depending on your environment and version of TensorFlow.

It also want to mention that the sess.run(tf.global_variables_initializer()) can be important before restoring the session.

If you are receiving an error regarding uninitialized variables when trying to restore a saved session, make sure you include sess.run(tf.global_variables_initializer()) before the saver.restore(sess, save_file) line. It can save you a headache.

Answer 16

As described in issue 6255:

use '**./**model_name.ckpt'
saver.restore(sess,'./my_model_final.ckpt')

instead of

saver.restore('my_model_final.ckpt')

Answer 17

According to the new Tensorflow version, tf.train.Checkpoint is the preferable way of saving and restoring a model:

Checkpoint.save and Checkpoint.restore write and read object-based checkpoints, in contrast to tf.train.Saver which writes and reads variable.name based checkpoints. Object-based checkpointing saves a graph of dependencies between Python objects (Layers, Optimizers, Variables, etc.) with named edges, and this graph is used to match variables when restoring a checkpoint. It can be more robust to changes in the Python program, and helps to support restore-on-create for variables when executing eagerly. Prefer tf.train.Checkpoint over tf.train.Saver for new code.

Here is an example:

import tensorflow as tf
import os

tf.enable_eager_execution()

checkpoint_directory = "/tmp/training_checkpoints"
checkpoint_prefix = os.path.join(checkpoint_directory, "ckpt")

checkpoint = tf.train.Checkpoint(optimizer=optimizer, model=model)
status = checkpoint.restore(tf.train.latest_checkpoint(checkpoint_directory))
for _ in range(num_training_steps):
  optimizer.minimize( ... )  # Variables will be restored on creation.
status.assert_consumed()  # Optional sanity checks.
checkpoint.save(file_prefix=checkpoint_prefix)

More information and example here.

Answer 18

For tensorflow 2.0, it is as simple as

# Save the model
model.save('path_to_my_model.h5')

To restore:

new_model = tensorflow.keras.models.load_model('path_to_my_model.h5')

Answer 19

For tensorflow-2.0

it's very simple.

import tensorflow as tf

SAVE

model.save("model_name")

RESTORE

model = tf.keras.models.load_model('model_name')

Answer 20

Here is a simple example using Tensorflow 2.0 SavedModel format (which is the recommended format, according to the docs) for a simple MNIST dataset classifier, using Keras functional API without too much fancy going on:

# Imports
import tensorflow as tf
from tensorflow.keras.layers import Input, Dense, Flatten
from tensorflow.keras.models import Model
import matplotlib.pyplot as plt

# Load data
mnist = tf.keras.datasets.mnist # 28 x 28
(x_train,y_train), (x_test, y_test) = mnist.load_data()

# Normalize pixels [0,255] -> [0,1]
x_train = tf.keras.utils.normalize(x_train,axis=1)
x_test = tf.keras.utils.normalize(x_test,axis=1)

# Create model
input = Input(shape=(28,28), dtype='float64', name='graph_input')
x = Flatten()(input)
x = Dense(128, activation='relu')(x)
x = Dense(128, activation='relu')(x)
output = Dense(10, activation='softmax', name='graph_output', dtype='float64')(x)
model = Model(inputs=input, outputs=output)

model.compile(optimizer='adam',
             loss='sparse_categorical_crossentropy',
             metrics=['accuracy'])

# Train
model.fit(x_train, y_train, epochs=3)

# Save model in SavedModel format (Tensorflow 2.0)
export_path = 'model'
tf.saved_model.save(model, export_path)

# ... possibly another python program 

# Reload model
loaded_model = tf.keras.models.load_model(export_path) 

# Get image sample for testing
index = 0
img = x_test[index] # I normalized the image on a previous step

# Predict using the signature definition (Tensorflow 2.0)
predict = loaded_model.signatures["serving_default"]
prediction = predict(tf.constant(img))

# Show results
print(np.argmax(prediction['graph_output']))  # prints the class number
plt.imshow(x_test[index], cmap=plt.cm.binary)  # prints the image

What is serving_default?

It's the name of the signature def of the tag you selected (in this case, the default serve tag was selected). Also, here explains how to find the tag's and signatures of a model using saved_model_cli.

Disclaimers

This is just a basic example if you just want to get it up and running, but is by no means a complete answer - maybe I can update it in the future. I just wanted to give a simple example using the SavedModel in TF 2.0 because I haven't seen one, even this simple, anywhere.

@Tom's answer is a SavedModel example, but it will not work on Tensorflow 2.0, because unfortunately there are some breaking changes.

@Vishnuvardhan Janapati's answer says TF 2.0, but it's not for SavedModel format.

Answer 21

You can save the variables in the network using

saver = tf.train.Saver() 
saver.save(sess, 'path of save/fileName.ckpt')

To restore the network for reuse later or in another script, use:

saver = tf.train.Saver()
saver.restore(sess, tf.train.latest_checkpoint('path of save/')
sess.run(....) 

Important points:

1. sess must be same between first and later runs (coherent structure).
2. saver.restore needs the path of the folder of the saved files, not an individual file path.

Answer 22

Following @Vishnuvardhan Janapati 's answer, here is another way to save and reload model with custom layer/metric/loss under TensorFlow 2.0.0

import tensorflow as tf
from tensorflow.keras.layers import Layer
from tensorflow.keras.utils.generic_utils import get_custom_objects

# custom loss (for example)  
def custom_loss(y_true,y_pred):
  return tf.reduce_mean(y_true - y_pred)
get_custom_objects().update({'custom_loss': custom_loss}) 

# custom loss (for example) 
class CustomLayer(Layer):
  def __init__(self, ...):
      ...
  # define custom layer and all necessary custom operations inside custom layer

get_custom_objects().update({'CustomLayer': CustomLayer})  

In this way, once you have executed such codes, and saved your model with tf.keras.models.save_model or model.save or ModelCheckpoint callback, you can re-load your model without the need of precise custom objects, as simple as

new_model = tf.keras.models.load_model("./model.h5"})

Answer 23

Tensorflow 2.6 : It has become way more simpler now, you can save model in 2 formats

1. Saved_model ( tf-serving compatible)
2. H5 or HDF5

Saving model in both formats :

 from tensorflow.keras import Model
 inputs = tf.keras.Input(shape=(224,224,3))
 y = tf.keras.layers.Conv2D(24, 3, activation='relu', input_shape=input_shape[1:])(inputs)
 outputs = tf.keras.layers.Dense(5, activation=tf.nn.softmax)(y)
 model = tf.keras.Model(inputs=inputs, outputs=outputs)
 model.save("saved_model/my_model") #To Save in Saved_model format
 model.save("my_model.h5") #To save model in H5 or HDF5 format

To load Model in both format

import tensorflow as tf
h5_model = tf.keras.models.load_model("my_model.h5") # loading model in h5 format
h5_model.summary()
saved_m = tf.keras.models.load_model("saved_model/my_model") #loading model in saved_model format
saved_m.summary()