It will be good to have an option to early stop or easily add other callbacks from ker

Add callback for earlystopping about nuts-ml HOT 1 OPEN

maet3608 commented on June 7, 2024

Add callback for earlystopping

from nuts-ml.

Comments (1)

maet3608 commented on June 7, 2024

The KerasNetwork wrapper does not support callbacks because the underlying Keras API (train_on_batch()) does not support it. However, you can use model.fit_generator
(and not use the KerasNetwork wrapper):

model = Sequential()
...

batches = train_samples >> Cycle() >> read_big_data >> ... >> build_batch
model.fit_generator(batches, callbacks=...)

More info can be found here: https://maet3608.github.io/nuts-ml/tutorial/network.html

You have to make sure that train_samples can be loaded into memory,
and all big data, e.g. images are loaded after Cycle() !

Alternatively, you can run the epoch loop explicitly and do early stopping there, e.g.

network = KerasNetwork(model)
for epoch in range(EPOCHS):
    t_loss, t_acc = (train_samples >> read_image >> transform >> augment >>
                     Shuffle(100) >> build_batch >> network.train() >> Unzip())

    v_loss, v_acc = (val_samples >> read_image >> transform >>
                     build_batch >> network.validate() >> Unzip())

    network.save_best(v_loss >> Mean(), isloss=True)  # early stopping

from nuts-ml.

Recommend Projects