Return R @ Rᵀ, which should be close to the identity matrix.
(self)
| 716 | Sampling using low-dimensional random projections (Johnson-Lindenstrauss lemma). |
| 717 | """ |
| 718 | if not self.prepared: |
| 719 | raise ValueError("Sampler not prepared. Call .prepare() first.") |
| 720 | logits_p = self.compute_logits(hidden_states) |
| 721 | probs = (logits_p / temperature).softmax(dim=1) |
| 722 | samples = _fast_multinomial(probs, num_samples) |
| 723 | return samples |
| 724 |
nothing calls this directly
no outgoing calls
no test coverage detected