def weight_inputs(self, inputs: BatchEncoding) -> List[float]:
device = self.transformer.model.device
all_attentions = self.transformer.model(**inputs.to(device),
output_attentions=True)[-1]
weights = self._aggregate_attentions(all_attentions, self.heads,
self.agg_strategy).detach().cpu()
if self.normalize_weights:
norm = torch.linalg.norm(weights, ord=2)
weights = torch.tensor(weights) / torch.max(
norm, 1e-10 * torch.ones_like(norm))
return weights.detach().cpu().numpy().tolist()
def embed_inputs(self, inputs: BatchEncoding) -> List[List[List[float]]]:
device = self.transformer.model.device
outputs = self.transformer.model(**inputs.to(device),
output_hidden_states=True)[-1]
embeddings_t = self._embedings_from_outputs(outputs)
return embeddings_t.detach().cpu().tolist()