contrastive

Name	Description
`ContrastiveTrainer`	Trainer for contrastive learning

Name

Description

ContrastiveTrainer

Trainer for contrastive learning

ContrastiveTrainer(
    encoder: keras.Model,
    projector: keras.Model | tuple[keras.Model, keras.Model],
    augmenter: keras.Layer | tuple[keras.Layer, keras.Layer] | None = None,
    probe: keras.Layer | keras.Model | None = None,
)

Creates a self-supervised contrastive trainer for a model.

Parameters

Parameters of ContrastiveTrainer
Name	Type	Default	Description
`encoder`	`keras.Model`	Required	The encoder model to be trained.
`projector`	`keras.Model \| tuple[keras.Model, keras.Model]`	Required	The projector model to be trained.
`augmenter`	`keras.Layer \| tuple[keras.Layer, keras.Layer] \| None`	None	The augmenter to be used for data augmentation.
`probe`	`keras.Layer \| keras.Model \| None`	None	The probe model to be trained. If None, no probe is used.

SAMPLES = 'data'

LABELS = 'labels'

AUG_SAMPLES_0 = 'augmented_data_0'

AUG_SAMPLES_1 = 'augmented_data_1'

augmenters: tuple[keras.Layer, keras.Layer]

encoder: keras.Model = encoder

projectors = projector if  else (projector, projector)

probe = probe

loss_metric = keras.metrics.Mean(name='loss')

encoder_metrics = []

probe_loss_metric = keras.metrics.Mean(name='probe_loss')

probe_metrics = []

metrics

compile(
    encoder_optimizer: keras.Optimizer,
    encoder_loss: keras.Loss,
    encoder_metrics: list[keras.Metric] | None = None,
    probe_optimizer: keras.Optimizer | None = None,
    probe_loss: keras.Loss | None = None,
    probe_metrics: list[keras.Metric] | None = None,
    **kwargs={},
)

fit(x=None, y=None, sample_weight=None, batch_size=None, validation_data=None, **kwargs={})

run_augmenters(x, y=None)

train_step(data)

test_step(data)

call(inputs)

linear_probe(num_classes, **kwargs={})

staticmethod

save(filepath, overwrite=True, zipped=True, **kwargs={})

We only save the encoder model

contrastive

Contrastive Trainer API