TensorFlow 2.0 Beta is available Learn more

Save and restore models

View on TensorFlow.org Run in Google Colab View source on GitHub Download notebook

Model progress can be saved during—and after—training. This means a model can resume where it left off and avoid long training times. Saving also means you can share your model and others can recreate your work. When publishing research models and techniques, most machine learning practitioners share:

  • code to create the model, and
  • the trained weights, or parameters, for the model

Sharing this data helps others understand how the model works and try it themselves with new data.

Options

There are different ways to save TensorFlow models—depending on the API you're using. This guide uses tf.keras, a high-level API to build and train models in TensorFlow. For other approaches, see the TensorFlow Save and Restore guide or Saving in eager.

Setup

Installs and imports

Install and import TensorFlow and dependencies:

!pip install -q h5py pyyaml

Get an example dataset

We'll use the MNIST dataset to train our model to demonstrate saving weights. To speed up these demonstration runs, only use the first 1000 examples:

from __future__ import absolute_import, division, print_function, unicode_literals

import os

!pip install -q tensorflow==2.0.0-beta1
import tensorflow as tf
from tensorflow import keras

tf.__version__
'2.0.0-beta1'
(train_images, train_labels), (test_images, test_labels) = tf.keras.datasets.mnist.load_data()

train_labels = train_labels[:1000]
test_labels = test_labels[:1000]

train_images = train_images[:1000].reshape(-1, 28 * 28) / 255.0
test_images = test_images[:1000].reshape(-1, 28 * 28) / 255.0

Define a model

Let's build a simple model we'll use to demonstrate saving and loading weights.

# Returns a short sequential model
def create_model():
  model = tf.keras.models.Sequential([
    keras.layers.Dense(512, activation='relu', input_shape=(784,)),
    keras.layers.Dropout(0.2),
    keras.layers.Dense(10, activation='softmax')
  ])

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

  return model


# Create a basic model instance
model = create_model()
model.summary()
Model: "sequential"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
dense (Dense)                (None, 512)               401920    
_________________________________________________________________
dropout (Dropout)            (None, 512)               0         
_________________________________________________________________
dense_1 (Dense)              (None, 10)                5130      
=================================================================
Total params: 407,050
Trainable params: 407,050
Non-trainable params: 0
_________________________________________________________________

Save checkpoints during training

The primary use case is to automatically save checkpoints during and at the end of training. This way you can use a trained model without having to retrain it, or pick-up training where you left of—in case the training process was interrupted.

tf.keras.callbacks.ModelCheckpoint is a callback that performs this task. The callback takes a couple of arguments to configure checkpointing.

Checkpoint callback usage

Train the model and pass it the ModelCheckpoint callback:

checkpoint_path = "training_1/cp.ckpt"
checkpoint_dir = os.path.dirname(checkpoint_path)

# Create checkpoint callback
cp_callback = tf.keras.callbacks.ModelCheckpoint(checkpoint_path,
                                                 save_weights_only=True,
                                                 verbose=1)

model = create_model()

model.fit(train_images, train_labels,  epochs = 10,
          validation_data = (test_images,test_labels),
          callbacks = [cp_callback])  # pass callback to training

# This may generate warnings related to saving the state of the optimizer.
# These warnings (and similar warnings throughout this notebook)
# are in place to discourage outdated usage, and can be ignored.
WARNING: Logging before flag parsing goes to stderr.
W0628 04:40:04.167828 140342847055616 deprecation.py:323] From /tmpfs/src/tf_docs_env/lib/python3.5/site-packages/tensorflow/python/ops/math_grad.py:1250: add_dispatch_support.<locals>.wrapper (from tensorflow.python.ops.array_ops) is deprecated and will be removed in a future version.
Instructions for updating:
Use tf.where in 2.0, which has the same broadcast rule as np.where

Train on 1000 samples, validate on 1000 samples
Epoch 1/10
 992/1000 [============================>.] - ETA: 0s - loss: 1.1842 - accuracy: 0.6482
Epoch 00001: saving model to training_1/cp.ckpt
1000/1000 [==============================] - 1s 661us/sample - loss: 1.1784 - accuracy: 0.6500 - val_loss: 0.7726 - val_accuracy: 0.7600
Epoch 2/10
 992/1000 [============================>.] - ETA: 0s - loss: 0.4351 - accuracy: 0.8810
Epoch 00002: saving model to training_1/cp.ckpt
1000/1000 [==============================] - 0s 190us/sample - loss: 0.4320 - accuracy: 0.8820 - val_loss: 0.5391 - val_accuracy: 0.8320
Epoch 3/10
 544/1000 [===============>..............] - ETA: 0s - loss: 0.2756 - accuracy: 0.9393
Epoch 00003: saving model to training_1/cp.ckpt
1000/1000 [==============================] - 0s 171us/sample - loss: 0.2998 - accuracy: 0.9230 - val_loss: 0.4611 - val_accuracy: 0.8550
Epoch 4/10
 608/1000 [=================>............] - ETA: 0s - loss: 0.1979 - accuracy: 0.9556
Epoch 00004: saving model to training_1/cp.ckpt
1000/1000 [==============================] - 0s 157us/sample - loss: 0.2056 - accuracy: 0.9530 - val_loss: 0.4361 - val_accuracy: 0.8620
Epoch 5/10
 576/1000 [================>.............] - ETA: 0s - loss: 0.1565 - accuracy: 0.9653
Epoch 00005: saving model to training_1/cp.ckpt
1000/1000 [==============================] - 0s 157us/sample - loss: 0.1513 - accuracy: 0.9670 - val_loss: 0.4293 - val_accuracy: 0.8630
Epoch 6/10
 576/1000 [================>.............] - ETA: 0s - loss: 0.1097 - accuracy: 0.9826
Epoch 00006: saving model to training_1/cp.ckpt
1000/1000 [==============================] - 0s 155us/sample - loss: 0.1137 - accuracy: 0.9810 - val_loss: 0.4072 - val_accuracy: 0.8690
Epoch 7/10
 608/1000 [=================>............] - ETA: 0s - loss: 0.0892 - accuracy: 0.9885
Epoch 00007: saving model to training_1/cp.ckpt
1000/1000 [==============================] - 0s 158us/sample - loss: 0.0859 - accuracy: 0.9900 - val_loss: 0.4117 - val_accuracy: 0.8670
Epoch 8/10
 544/1000 [===============>..............] - ETA: 0s - loss: 0.0591 - accuracy: 0.9945
Epoch 00008: saving model to training_1/cp.ckpt
1000/1000 [==============================] - 0s 165us/sample - loss: 0.0678 - accuracy: 0.9910 - val_loss: 0.3934 - val_accuracy: 0.8710
Epoch 9/10
 608/1000 [=================>............] - ETA: 0s - loss: 0.0570 - accuracy: 0.9934
Epoch 00009: saving model to training_1/cp.ckpt
1000/1000 [==============================] - 0s 153us/sample - loss: 0.0510 - accuracy: 0.9940 - val_loss: 0.4002 - val_accuracy: 0.8710
Epoch 10/10
 576/1000 [================>.............] - ETA: 0s - loss: 0.0374 - accuracy: 0.9983
Epoch 00010: saving model to training_1/cp.ckpt
1000/1000 [==============================] - 0s 161us/sample - loss: 0.0389 - accuracy: 0.9980 - val_loss: 0.3910 - val_accuracy: 0.8730

<tensorflow.python.keras.callbacks.History at 0x7fa3c23d87f0>

This creates a single collection of TensorFlow checkpoint files that are updated at the end of each epoch:

!ls {checkpoint_dir}
checkpoint  cp.ckpt.data-00000-of-00001  cp.ckpt.index

Create a new, untrained model. When restoring a model from only weights, you must have a model with the same architecture as the original model. Since it's the same model architecture, we can share weights despite that it's a different instance of the model.

Now rebuild a fresh, untrained model, and evaluate it on the test set. An untrained model will perform at chance levels (~10% accuracy):

model = create_model()

loss, acc = model.evaluate(test_images, test_labels)
print("Untrained model, accuracy: {:5.2f}%".format(100*acc))
1000/1000 [==============================] - 0s 108us/sample - loss: 2.4145 - accuracy: 0.0780
Untrained model, accuracy:  7.80%

Then load the weights from the checkpoint, and re-evaluate:

model.load_weights(checkpoint_path)
loss,acc = model.evaluate(test_images, test_labels)
print("Restored model, accuracy: {:5.2f}%".format(100*acc))
1000/1000 [==============================] - 0s 44us/sample - loss: 0.3910 - accuracy: 0.8730
Restored model, accuracy: 87.30%

Checkpoint callback options

The callback provides several options to give the resulting checkpoints unique names, and adjust the checkpointing frequency.

Train a new model, and save uniquely named checkpoints once every 5-epochs:

# 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)
W0628 04:40:07.198855 140342847055616 callbacks.py:859] `period` argument is deprecated. Please use `save_freq` to specify the frequency in number of samples seen.


Epoch 00005: saving model to training_2/cp-0005.ckpt

Epoch 00010: saving model to training_2/cp-0010.ckpt

Epoch 00015: saving model to training_2/cp-0015.ckpt

Epoch 00020: saving model to training_2/cp-0020.ckpt

Epoch 00025: saving model to training_2/cp-0025.ckpt

Epoch 00030: saving model to training_2/cp-0030.ckpt

Epoch 00035: saving model to training_2/cp-0035.ckpt

Epoch 00040: saving model to training_2/cp-0040.ckpt

Epoch 00045: saving model to training_2/cp-0045.ckpt

Epoch 00050: saving model to training_2/cp-0050.ckpt

<tensorflow.python.keras.callbacks.History at 0x7fa3f9b87d30>

Now, look at the resulting checkpoints and choose the latest one:

! ls {checkpoint_dir}
checkpoint            cp-0025.ckpt.index
cp-0000.ckpt.data-00000-of-00001  cp-0030.ckpt.data-00000-of-00001
cp-0000.ckpt.index        cp-0030.ckpt.index
cp-0005.ckpt.data-00000-of-00001  cp-0035.ckpt.data-00000-of-00001
cp-0005.ckpt.index        cp-0035.ckpt.index
cp-0010.ckpt.data-00000-of-00001  cp-0040.ckpt.data-00000-of-00001
cp-0010.ckpt.index        cp-0040.ckpt.index
cp-0015.ckpt.data-00000-of-00001  cp-0045.ckpt.data-00000-of-00001
cp-0015.ckpt.index        cp-0045.ckpt.index
cp-0020.ckpt.data-00000-of-00001  cp-0050.ckpt.data-00000-of-00001
cp-0020.ckpt.index        cp-0050.ckpt.index
cp-0025.ckpt.data-00000-of-00001
latest = tf.train.latest_checkpoint(checkpoint_dir)
latest
'training_2/cp-0050.ckpt'

To test, reset the model and load the latest checkpoint:

model = create_model()
model.load_weights(latest)
loss, acc = model.evaluate(test_images, test_labels)
print("Restored model, accuracy: {:5.2f}%".format(100*acc))
1000/1000 [==============================] - 0s 102us/sample - loss: 0.4884 - accuracy: 0.8750
Restored model, accuracy: 87.50%

What are these files?

The above code stores the weights to a collection of checkpoint-formatted files that contain only the trained weights in a binary format. Checkpoints contain: * One or more shards that contain your model's weights. * An index file that indicates which weights are stored in a which shard.

If you are only training a model on a single machine, you'll have one shard with the suffix: .data-00000-of-00001

Manually save weights

Above you saw how to load the weights into a model.

Manually saving the weights is just as simple, use the Model.save_weights method.

# 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(test_images, test_labels)
print("Restored model, accuracy: {:5.2f}%".format(100*acc))
1000/1000 [==============================] - 0s 100us/sample - loss: 0.4884 - accuracy: 0.8750
Restored model, accuracy: 87.50%

Save the entire model

The model and optimizer can be saved to a file that contains both their state (weights and variables), and the model configuration. This allows you to export a model so it can be used without access to the original python code. Since the optimizer-state is recovered you can even resume training from exactly where you left off.

Saving a fully-functional model is very useful—you can load them in TensorFlow.js (HDF5, Saved Model) and then train and run them in web browsers, or convert them to run on mobile devices using TensorFlow Lite (HDF5, Saved Model)

As an HDF5 file

Keras provides a basic save format using the HDF5 standard. For our purposes, the saved model can be treated as a single binary blob.

model = create_model()

model.fit(train_images, train_labels, epochs=5)

# Save entire model to a HDF5 file
model.save('my_model.h5')
W0628 04:40:16.245705 140342847055616 util.py:244] Unresolved object in checkpoint: (root).optimizer.iter
W0628 04:40:16.247200 140342847055616 util.py:244] Unresolved object in checkpoint: (root).optimizer.beta_1
W0628 04:40:16.247903 140342847055616 util.py:244] Unresolved object in checkpoint: (root).optimizer.beta_2
W0628 04:40:16.248561 140342847055616 util.py:244] Unresolved object in checkpoint: (root).optimizer.decay
W0628 04:40:16.249153 140342847055616 util.py:244] Unresolved object in checkpoint: (root).optimizer.learning_rate
W0628 04:40:16.249761 140342847055616 util.py:252] A checkpoint was restored (e.g. tf.train.Checkpoint.restore or tf.keras.Model.load_weights) but not all checkpointed values were used. See above for specific issues. Use expect_partial() on the load status object, e.g. tf.train.Checkpoint.restore(...).expect_partial(), to silence these warnings, or use assert_consumed() to make the check explicit. See https://www.tensorflow.org/alpha/guide/checkpoints#loading_mechanics for details.
W0628 04:40:16.252574 140342847055616 util.py:244] Unresolved object in checkpoint: (root).optimizer.iter
W0628 04:40:16.253534 140342847055616 util.py:244] Unresolved object in checkpoint: (root).optimizer.beta_1
W0628 04:40:16.254374 140342847055616 util.py:244] Unresolved object in checkpoint: (root).optimizer.beta_2
W0628 04:40:16.255190 140342847055616 util.py:244] Unresolved object in checkpoint: (root).optimizer.decay
W0628 04:40:16.255778 140342847055616 util.py:244] Unresolved object in checkpoint: (root).optimizer.learning_rate
W0628 04:40:16.256367 140342847055616 util.py:252] A checkpoint was restored (e.g. tf.train.Checkpoint.restore or tf.keras.Model.load_weights) but not all checkpointed values were used. See above for specific issues. Use expect_partial() on the load status object, e.g. tf.train.Checkpoint.restore(...).expect_partial(), to silence these warnings, or use assert_consumed() to make the check explicit. See https://www.tensorflow.org/alpha/guide/checkpoints#loading_mechanics for details.

Train on 1000 samples
Epoch 1/5
1000/1000 [==============================] - 0s 261us/sample - loss: 1.1682 - accuracy: 0.6560
Epoch 2/5
1000/1000 [==============================] - 0s 105us/sample - loss: 0.4095 - accuracy: 0.8830
Epoch 3/5
1000/1000 [==============================] - 0s 101us/sample - loss: 0.2954 - accuracy: 0.9190
Epoch 4/5
1000/1000 [==============================] - 0s 102us/sample - loss: 0.2100 - accuracy: 0.9460
Epoch 5/5
1000/1000 [==============================] - 0s 99us/sample - loss: 0.1565 - accuracy: 0.9630

Now recreate the model from that file:

# Recreate the exact same model, including weights and optimizer.
new_model = keras.models.load_model('my_model.h5')
new_model.summary()
Model: "sequential_6"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
dense_12 (Dense)             (None, 512)               401920    
_________________________________________________________________
dropout_6 (Dropout)          (None, 512)               0         
_________________________________________________________________
dense_13 (Dense)             (None, 10)                5130      
=================================================================
Total params: 407,050
Trainable params: 407,050
Non-trainable params: 0
_________________________________________________________________

Check its accuracy:

loss, acc = new_model.evaluate(test_images, test_labels)
print("Restored model, accuracy: {:5.2f}%".format(100*acc))
1000/1000 [==============================] - 0s 96us/sample - loss: 0.4376 - accuracy: 0.8560
Restored model, accuracy: 85.60%

This technique saves everything:

  • The weight values
  • The model's configuration(architecture)
  • The optimizer configuration

Keras saves models by inspecting the architecture. Currently, it is not able to save TensorFlow optimizers (from tf.train). When using those you will need to re-compile the model after loading, and you will lose the state of the optimizer.

As a saved_model

Build a fresh model:

model = create_model()

model.fit(train_images, train_labels, epochs=5)
Train on 1000 samples
Epoch 1/5
1000/1000 [==============================] - 0s 254us/sample - loss: 1.1280 - accuracy: 0.6910
Epoch 2/5
1000/1000 [==============================] - 0s 107us/sample - loss: 0.4015 - accuracy: 0.8950
Epoch 3/5
1000/1000 [==============================] - 0s 107us/sample - loss: 0.2782 - accuracy: 0.9260
Epoch 4/5
1000/1000 [==============================] - 0s 96us/sample - loss: 0.2066 - accuracy: 0.9530
Epoch 5/5
1000/1000 [==============================] - 0s 96us/sample - loss: 0.1556 - accuracy: 0.9640

<tensorflow.python.keras.callbacks.History at 0x7fa3f8e2a048>

Create a saved_model, and place it in a time-stamped directory:

import time
saved_model_path = "./saved_models/{}".format(int(time.time()))

tf.keras.experimental.export_saved_model(model, saved_model_path)
saved_model_path
W0628 04:40:19.867537 140342847055616 deprecation.py:323] From /tmpfs/src/tf_docs_env/lib/python3.5/site-packages/tensorflow/python/saved_model/signature_def_utils_impl.py:253: build_tensor_info (from tensorflow.python.saved_model.utils_impl) is deprecated and will be removed in a future version.
Instructions for updating:
This function will only be available through the v1 compatibility library as tf.compat.v1.saved_model.utils.build_tensor_info or tf.compat.v1.saved_model.build_tensor_info.
W0628 04:40:19.869408 140342847055616 export_utils.py:182] Export includes no default signature!
W0628 04:40:20.103580 140342847055616 export_utils.py:182] Export includes no default signature!

'./saved_models/1561696819'

List your saved models:

!ls saved_models/
1561696819

Reload a fresh keras model from the saved model.

new_model = tf.keras.experimental.load_from_saved_model(saved_model_path)
new_model.summary()
W0628 04:40:20.426350 140342847055616 util.py:244] Unresolved object in checkpoint: (root).optimizer
W0628 04:40:20.427524 140342847055616 util.py:244] Unresolved object in checkpoint: (root).optimizer.iter
W0628 04:40:20.428700 140342847055616 util.py:244] Unresolved object in checkpoint: (root).optimizer.beta_1
W0628 04:40:20.429565 140342847055616 util.py:244] Unresolved object in checkpoint: (root).optimizer.beta_2
W0628 04:40:20.430170 140342847055616 util.py:244] Unresolved object in checkpoint: (root).optimizer.decay
W0628 04:40:20.430845 140342847055616 util.py:244] Unresolved object in checkpoint: (root).optimizer.learning_rate
W0628 04:40:20.431444 140342847055616 util.py:244] Unresolved object in checkpoint: (root).optimizer's state 'm' for (root).layer_with_weights-0.kernel
W0628 04:40:20.431999 140342847055616 util.py:244] Unresolved object in checkpoint: (root).optimizer's state 'm' for (root).layer_with_weights-0.bias
W0628 04:40:20.433732 140342847055616 util.py:244] Unresolved object in checkpoint: (root).optimizer's state 'm' for (root).layer_with_weights-1.kernel
W0628 04:40:20.434602 140342847055616 util.py:244] Unresolved object in checkpoint: (root).optimizer's state 'm' for (root).layer_with_weights-1.bias
W0628 04:40:20.435332 140342847055616 util.py:244] Unresolved object in checkpoint: (root).optimizer's state 'v' for (root).layer_with_weights-0.kernel
W0628 04:40:20.436072 140342847055616 util.py:244] Unresolved object in checkpoint: (root).optimizer's state 'v' for (root).layer_with_weights-0.bias
W0628 04:40:20.436897 140342847055616 util.py:244] Unresolved object in checkpoint: (root).optimizer's state 'v' for (root).layer_with_weights-1.kernel
W0628 04:40:20.437806 140342847055616 util.py:244] Unresolved object in checkpoint: (root).optimizer's state 'v' for (root).layer_with_weights-1.bias
W0628 04:40:20.438539 140342847055616 util.py:252] A checkpoint was restored (e.g. tf.train.Checkpoint.restore or tf.keras.Model.load_weights) but not all checkpointed values were used. See above for specific issues. Use expect_partial() on the load status object, e.g. tf.train.Checkpoint.restore(...).expect_partial(), to silence these warnings, or use assert_consumed() to make the check explicit. See https://www.tensorflow.org/alpha/guide/checkpoints#loading_mechanics for details.

Model: "sequential_7"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
dense_14 (Dense)             (None, 512)               401920    
_________________________________________________________________
dropout_7 (Dropout)          (None, 512)               0         
_________________________________________________________________
dense_15 (Dense)             (None, 10)                5130      
=================================================================
Total params: 407,050
Trainable params: 407,050
Non-trainable params: 0
_________________________________________________________________

Run the restored model.

model.predict(test_images).shape
(1000, 10)
# The model has to be compiled before evaluating.
# This step is not required if the saved model is only being deployed.

new_model.compile(optimizer=model.optimizer,  # keep the optimizer that was loaded
              loss='sparse_categorical_crossentropy',
              metrics=['accuracy'])

# Evaluate the restored model.
loss, acc = new_model.evaluate(test_images, test_labels)
print("Restored model, accuracy: {:5.2f}%".format(100*acc))
1000/1000 [==============================] - 0s 100us/sample - loss: 0.4579 - accuracy: 0.8520
Restored model, accuracy: 85.20%

What's Next

That was a quick guide to saving and loading in with tf.keras.

#@title MIT License
#
# Copyright (c) 2017 François Chollet
#
# Permission is hereby granted, free of charge, to any person obtaining a
# copy of this software and associated documentation files (the "Software"),
# to deal in the Software without restriction, including without limitation
# the rights to use, copy, modify, merge, publish, distribute, sublicense,
# and/or sell copies of the Software, and to permit persons to whom the
# Software is furnished to do so, subject to the following conditions:
#
# The above copyright notice and this permission notice shall be included in
# all copies or substantial portions of the Software.
#
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
# THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
# FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
# DEALINGS IN THE SOFTWARE.