Skip to main content

Keras wrapper that autosaves what ModelCheckpoint cannot.

Project description

Keras wrapper that autosaves and auto-recovers not just the model weights but also the last epoch number and training history metrics.

See it in action in this Colab notebook!

pip install keras-buoy


When training is interrupted and you rerun the whole code, it recovers the model weights and the epoch counter to the last saved values. Then it resumes training as if nothing happened. At the end, the Keras History.history dictionaries are combined so that the training history looks like one single training run.


>>> from tensorflow import keras
>>> from keras_buoy.models import ResumableModel

>>> model = keras.Sequential()
>>> resumable_model = ResumableModel(model,
>>> history =,

Recovered model from kerascheckpoint.h5 at epoch 8.

Epoch 9/15
1125/1125 - 5s - loss: 0.4790 - top_k_categorical_accuracy: 0.9698 - val_loss: 1.1075 - val_top_k_categorical_accuracy: 0.9206
Epoch 10/15
1125/1125 - 5s - loss: 0.4758 - top_k_categorical_accuracy: 0.9701 - val_loss: 1.1119 - val_top_k_categorical_accuracy: 0.9214
Epoch 11/15
1125/1125 - 5s - loss: 0.4753 - top_k_categorical_accuracy: 0.9702 - val_loss: 1.1000 - val_top_k_categorical_accuracy: 0.9215
Epoch 12/15

Try it out yourself in this Colab notebook.



Creates a resumable model.


Parameter name Description
model (tf.keras.Model) The instance of tf.keras.Model which you want to make resumable.
save_every_epochs (int) Specifies how often to save the model, history, and epoch counter. In case of a crash, recovery will happen from the last saved epoch multiple.
custom_objects (dict) At recovery time, this is passed into tf.keras.models.load_model(...) exactly as shown in Tensorflow docs so you can load your model with a custom loss for example.
to_path (str) Specifies the path where the model weights will be saved. If it ends with .h5, then it saves in the Keras H5 format instead of the default TensorFlow SavedModel format.

If to_path is mymodel.h5, then there will be mymodel_epoch_num.pkl and mymodel_history.pkl in the same directory as mymodel.h5, which hold backups for the epoch counter and the history dict, respectively.


A ResumableModel instance. You can call .fit(...) on it.

Fits a resumable model.


The accepted parameters are the same as except you cannot specify initial_epoch.


history (dict): The history dict of the Keras History object. Note that it does not return the Keras.History object itself, just the dict.


This project has been set up using PyScaffold 3.2.3. For details and usage information on PyScaffold see

Project details

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Files for keras-buoy, version
Filename, size File type Python version Upload date Hashes
Filename, size keras-buoy- (19.0 kB) File type Source Python version None Upload date Hashes View

Supported by

AWS AWS Cloud computing Datadog Datadog Monitoring Facebook / Instagram Facebook / Instagram PSF Sponsor Fastly Fastly CDN Google Google Object Storage and Download Analytics Huawei Huawei PSF Sponsor Microsoft Microsoft PSF Sponsor NVIDIA NVIDIA PSF Sponsor Pingdom Pingdom Monitoring Salesforce Salesforce PSF Sponsor Sentry Sentry Error logging StatusPage StatusPage Status page