src/speech_decoding/models/v14_converged_v3/objective.py

Selected training excerpt · lines 189–191 · source SHA-256 5718458b537d

189        d = x_full.shape[-1]190        x_vis = x_full.gather(1, pack.idx[:, :, None].expand(-1, -1, d))  # (B, M_vis, d)191        return self.encoder.forward_flat_pack(x_vis, pack, tap_blocks=tap_blocks)