def evaluate(self, theta: Theta, xy_test: Partition) -> Tuple[float, float]:
model = self.model_provider.init_model()
model.set_weights(theta)
ds_val = prep.init_ds_val(xy_test)
# Assume the validation `tf.data.Dataset` to yield exactly one batch containing
# all examples in the validation set
loss, accuracy = model.evaluate(ds_val, steps=1, verbose=0)
return loss, accuracy
def fit(self, model: tf.keras.Model, epochs: int) -> KerasHistory:
ds_train = prep.init_ds_train(self.xy_train, self.num_classes,
self.batch_size)
ds_val = prep.init_ds_val(self.xy_val, self.num_classes)
hist = model.fit(
ds_train,
epochs=epochs,
validation_data=ds_val,
callbacks=[LoggingCallback(str(self.cid), logging.info)],
shuffle=False, # Shuffling is handled via tf.data.Dataset
steps_per_epoch=self.steps_train,
validation_steps=self.steps_val,
verbose=0,
)
return cast_to_float(hist.history)
def create_evalueate_fn(
orig_model: tf.keras.Model,
xy_val: Partition) -> Callable[[Theta], Tuple[float, float]]:
ds_val = prep.init_ds_val(xy_val)
model = tf.keras.models.clone_model(orig_model)
# FIXME refactor model compilation
model.compile(
loss=tf.keras.losses.categorical_crossentropy,
optimizer=tf.keras.optimizers.Adam(),
metrics=["accuracy"],
)
def fn(theta: Theta) -> Tuple[float, float]:
model.set_weights(theta)
# Assume the validation `tf.data.Dataset` to yield exactly one batch containing
# all examples in the validation set
return model.evaluate(ds_val, steps=1)
return fn
def evaluate(self, xy_val: Partition) -> Tuple[float, float]:
ds_val = prep.init_ds_val(xy_val)
# Assume the validation `tf.data.Dataset` to yield exactly one batch containing
# all examples in the validation set
loss, accuracy = self.model.evaluate(ds_val, steps=1)
return float(loss), float(accuracy)
def __init__(
self, model: tf.keras.Model, xy_val: Tuple[np.ndarray, np.ndarray]
) -> None:
self.model = model
self.ds_val = prep.init_ds_val(xy_val)
def __init__(self, model: tf.keras.Model, xy_val: Partition) -> None:
self.model = model
self.ds_val = prep.init_ds_val(xy_val)
