def _build_model(self) -> Estimator:
"""
Initializes the estimator architecture.
"""
super()._build_model()
if self.hparams.encoder_model != "LASER":
self.layer = (
int(self.hparams.layer)
if self.hparams.layer != "mix"
else self.hparams.layer
)
self.scalar_mix = (
ScalarMixWithDropout(
mixture_size=self.encoder.num_layers,
dropout=self.hparams.scalar_mix_dropout,
do_layer_norm=True,
)
if self.layer == "mix" and self.hparams.pool != "default"
else None
)
self.ff = FeedForward(
in_dim=self.encoder.output_units * 4,
hidden_sizes=self.hparams.hidden_sizes,
activations=self.hparams.activations,
dropout=self.hparams.dropout,
)
def _build_model(self) -> Estimator:
"""
Initializes the estimator architecture.
"""
super()._build_model()
if self.hparams.encoder_model != "LASER":
self.layer = (int(self.hparams.layer) if
self.hparams.layer != "mix" else self.hparams.layer)
self.scalar_mix = (ScalarMixWithDropout(
mixture_size=self.encoder.num_layers,
dropout=self.hparams.scalar_mix_dropout,
do_layer_norm=True,
) if self.layer == "mix" and self.hparams.pool != "default" else
None)
input_emb_sz = (self.encoder.output_units *
6 if self.hparams.pool != "cls+avg" else
self.encoder.output_units * 2 * 6)
self.ff = FeedForward(
in_dim=input_emb_sz,
hidden_sizes=self.hparams.hidden_sizes,
activations=self.hparams.activations,
dropout=self.hparams.dropout,
final_activation=(
self.hparams.final_activation if hasattr(
self.hparams, "final_activation"
) # compatability with older checkpoints!
else "Sigmoid"),
)
def _build_model(self) -> ModelBase:
"""
Initializes the ranking model architecture.
"""
super()._build_model()
self.metrics = WMTKendall()
if self.hparams.encoder_model != "LASER":
self.layer = (int(self.hparams.layer) if
self.hparams.layer != "mix" else self.hparams.layer)
self.scalar_mix = (ScalarMixWithDropout(
mixture_size=self.encoder.num_layers,
dropout=self.hparams.scalar_mix_dropout,
do_layer_norm=True,
) if self.layer == "mix" and self.hparams.pool != "default" else
None)