Keras wrapper that autosaves and auto-recovers not just the model weights but also the last epoch number and training history metrics.
pip install keras-buoy
>>> resumableModel = ResumableModel(model, save_every_epochs=4, to_path='kerascheckpoint.h5') >>> history = resumableModel.fit(x = x_train, y = y_train, validation_split=0.1, batch_size = 256, verbose=2, epochs=15) 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 ...
When training is interrupted due to a crash/accidental Ctrl+C 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, save_every_epochs = 4, to_path='/path/to/save/model_weights.h5') history = resumable_model.fit(x = x_train, y = y_train, validation_split = 0.1, batch_size = 256, verbose = 2, epochs = 12)
save_every_epochs (int) will save the model, history, and epoch counter every so often. In case of a crash, recovery will happen from the last saved epoch multiple.
to_path (str) is where the model weights will be saved, and must have the .h5 extension.
resumable_model.fit(...) is the same as Keras' model.fit(...).
It returns history which is the history dict of the Keras History object. Note that it does not return the Keras.History object itself, just the dict.
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.
This project has been set up using PyScaffold 3.2.3. For details and usage information on PyScaffold see https://pyscaffold.org/.