DeepHealth/backbones.py

import math

import torch
import torch.nn as nn
import torch.nn.functional as F


class TimeRoPE(nn.Module):
    def __init__(self, dim: int, base: float = 10000.0):
        super().__init__()
        assert dim % 2 == 0, "RoPE dim must be even"
        self.dim = dim
        # inv_freq: (dim // 2,) — not trainable, but should move with device
        inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2).float() / dim))
        self.register_buffer("inv_freq", inv_freq, persistent=False)

    def precompute_cache(self, tau: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]:
        t = tau.unsqueeze(-1)  # (B, L, 1)
        angles = t * self.inv_freq  # (B, L, dim//2)
        # Pre-expand for heads and interleave once (avoids N_layers repeats)
        cos = angles.cos().unsqueeze(1).repeat_interleave(2, dim=-1)
        sin = angles.sin().unsqueeze(1).repeat_interleave(2, dim=-1)
        return cos, sin  # (B, 1, L, dim)

    @staticmethod
    def _rotate_half(x: torch.Tensor) -> torch.Tensor:
        """Rotate pairs: ``[-x2, x1, -x4, x3, ...]``."""
        x1 = x[..., 0::2]
        x2 = x[..., 1::2]
        return torch.stack((-x2, x1), dim=-1).flatten(-2)

    @staticmethod
    def apply_from_cache(
        q: torch.Tensor,
        k: torch.Tensor,
        rope_cache: tuple[torch.Tensor, torch.Tensor],
    ) -> tuple[torch.Tensor, torch.Tensor]:
        cos, sin = rope_cache  # each (B, 1, L, dim)
        q_rot = q * cos + TimeRoPE._rotate_half(q) * sin
        k_rot = k * cos + TimeRoPE._rotate_half(k) * sin
        return q_rot, k_rot

    def forward(
        self,
        tau: torch.Tensor,
        q: torch.Tensor,
        k: torch.Tensor,
    ) -> tuple[torch.Tensor, torch.Tensor]:
        cache = self.precompute_cache(tau)
        return self.apply_from_cache(q, k, cache)


class GaussianRBFTimeBasis(nn.Module):
    def __init__(
        self,
        n_bases: int = 16,
        max_time_diff: float = 40.0,
    ):
        super().__init__()
        self.n_bases = n_bases

        # Evenly spaced RBF centres for non-negative linear time differences.
        # Causal masking enforces query_time >= key_time, so diff is >= 0.
        centers = torch.linspace(0.0, max_time_diff, n_bases)
        self.register_buffer("centers", centers,
                             persistent=False)  # (n_bases,)

        # Learnable log-widths (initialized to center spacing on linear scale).
        init_width = max(max_time_diff / max(n_bases - 1, 1), 1e-3)
        init_log_width = math.log(init_width)
        self.log_widths = nn.Parameter(torch.full((n_bases,), init_log_width))

    def precompute_cache(self, tau: torch.Tensor) -> torch.Tensor:

        time_coord = tau.float()  # (B, L)
        # Pairwise signed difference: query_i − key_j
        diff = time_coord.unsqueeze(
            2) - time_coord.unsqueeze(1)  # (B, L_q, L_k)
        # Gaussian RBF: exp(-0.5 * ((diff - c) / w)^2)
        diff = diff.unsqueeze(-1)  # (B, L, L, 1)
        widths = self.log_widths.exp()  # (n_bases,)
        rbf_acts = torch.exp(
            -0.5 * ((diff - self.centers) / widths).square()
            # (B, L, L, n_bases)
        )
        return rbf_acts


class TemporalAttention(nn.Module):
    def __init__(
        self,
        n_embd: int,
        n_head: int,
        n_rbf_bases: int = 16,
        dropout: float = 0.0,
        use_time_rope: bool = True,
        use_rbf_bias: bool = True,
    ):
        super().__init__()
        assert n_embd % n_head == 0, "n_embd must be divisible by n_head"
        self.n_head = n_head
        self.d_head = n_embd // n_head
        self.scale = 1.0 / math.sqrt(self.d_head)
        self.use_time_rope = use_time_rope
        self.use_rbf_bias = use_rbf_bias

        # QKV projection (fused for efficiency)
        self.qkv = nn.Linear(n_embd, 3 * n_embd, bias=False)
        # Output projection
        self.out_proj = nn.Linear(n_embd, n_embd, bias=False)

        # Layer-specific projection from shared RBF basis activations to per-head attention bias.
        self.rbf_proj = nn.Linear(n_rbf_bases, n_head, bias=False)
        self.time_bias_scale = nn.Parameter(torch.tensor(0.0))

        self.resid_drop = nn.Dropout(dropout)
        self.reset_parameters()

    def reset_parameters(self) -> None:
        """Match the previous version's GPT-style weight initialization."""
        nn.init.normal_(self.qkv.weight, mean=0.0, std=0.02)
        nn.init.normal_(self.out_proj.weight, mean=0.0, std=0.02)
        nn.init.zeros_(self.rbf_proj.weight)

    def forward(
        self,
        x: torch.Tensor,
        rope_cache: tuple[torch.Tensor, torch.Tensor] | None = None,
        rbf_cache: torch.Tensor | None = None,
        attn_mask: torch.Tensor | None = None,
    ) -> torch.Tensor:
        if self.use_time_rope:
            assert rope_cache is not None, "rope_cache must be provided when use_time_rope is True"
        if self.use_rbf_bias:
            assert rbf_cache is not None, "rbf_cache must be provided when use_rbf_bias is True"

        B, L, _ = x.shape
        H, D = self.n_head, self.d_head

        # --- QKV ----------------------------------------------------------
        qkv = self.qkv(x).reshape(B, L, 3, H, D).permute(2, 0, 3, 1, 4)
        q, k, v = qkv.unbind(0)  # each (B, H, L, D)

        # --- Apply RoPE (from shared cache) --------------------------------
        if self.use_time_rope:
            q, k = TimeRoPE.apply_from_cache(q, k, rope_cache)

        # Build additive attention bias mask: time bias + causal/padding mask.
        time_bias = None
        if self.use_rbf_bias:
            time_bias = self.rbf_proj(rbf_cache).permute(
                0, 3, 1, 2)  # (B, H, L, L)
            time_bias = self.time_bias_scale.tanh() * time_bias

        if time_bias is not None and attn_mask is not None:
            attn_bias = time_bias + attn_mask.to(time_bias.dtype)
        elif time_bias is not None:
            attn_bias = time_bias
        elif attn_mask is not None:
            attn_bias = attn_mask
        else:
            attn_bias = None

        out = F.scaled_dot_product_attention(
            q,
            k,
            v,
            attn_mask=attn_bias,
            dropout_p=0.0,
            is_causal=False,
            scale=self.scale,
        )

        # --- Aggregate & project out --------------------------------------
        out = out.transpose(1, 2).reshape(B, L, H * D)
        return self.resid_drop(self.out_proj(out))


class SwiGLU(nn.Module):
    def __init__(
        self,
        n_embd: int,
        hidden_dim: int | None = None,
        dropout: float = 0.0,
        bias: bool = True,
    ):
        super().__init__()
        hidden_dim = hidden_dim if hidden_dim is not None else int(
            n_embd * 2.5)

        self.w1 = nn.Linear(n_embd, hidden_dim, bias=bias)  # gate path
        self.w2 = nn.Linear(n_embd, hidden_dim, bias=bias)  # value path
        # output projection
        self.w3 = nn.Linear(hidden_dim, n_embd, bias=bias)
        self.drop = nn.Dropout(dropout)
        self.reset_parameters()

    def reset_parameters(self) -> None:
        """GPT-style parameter initialization for MLP paths."""
        nn.init.normal_(self.w1.weight, mean=0.0, std=0.02)
        nn.init.normal_(self.w2.weight, mean=0.0, std=0.02)
        nn.init.normal_(self.w3.weight, mean=0.0, std=0.02)
        if self.w1.bias is not None:
            nn.init.zeros_(self.w1.bias)
            nn.init.zeros_(self.w2.bias)
            nn.init.zeros_(self.w3.bias)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        """``(B, L, n_embd) -> (B, L, n_embd)``."""
        return self.drop(self.w3(F.silu(self.w1(x)) * self.w2(x)))


class GPTBlock(nn.Module):
    def __init__(
        self,
        n_embd: int,
        n_head: int,

        attn_dropout: float = 0.0,
        mlp_dropout: float = 0.0,
        use_time_rope: bool = False,
        use_rbf_bias: bool = False,
        use_cross_attention: bool = False,
        n_rbf_bases: int = 16,
    ):
        super().__init__()
        self.attn = TemporalAttention(
            n_embd=n_embd,
            n_head=n_head,
            n_rbf_bases=n_rbf_bases,
            dropout=attn_dropout,
            use_time_rope=use_time_rope,
            use_rbf_bias=use_rbf_bias,
        )
        self.use_cross_attention = bool(use_cross_attention)
        self.cross_attn = (
            CrossAttention(
                n_embd=n_embd,
                n_head=n_head,
                dropout=attn_dropout,
                n_rbf_bases=n_rbf_bases,
            )
            if self.use_cross_attention
            else None
        )
        self.mlp = SwiGLU(n_embd=n_embd, dropout=mlp_dropout)
        self.ln1 = nn.LayerNorm(n_embd)
        self.ln2 = nn.LayerNorm(n_embd)

    def forward(
        self,
        x: torch.Tensor,
        rope_cache: tuple[torch.Tensor, torch.Tensor] | None = None,
        rbf_cache: torch.Tensor | None = None,
        attn_mask: torch.Tensor | None = None,
        t_disease: torch.Tensor | None = None,
        h_token: torch.Tensor | None = None,
        t_token: torch.Tensor | None = None,
        token_mask: torch.Tensor | None = None,
    ) -> torch.Tensor:
        x = x + self.attn(self.ln1(x), rope_cache, rbf_cache, attn_mask)
        if self.use_cross_attention:
            if (
                self.cross_attn is None
                or t_disease is None
                or h_token is None
                or t_token is None
                or token_mask is None
            ):
                raise ValueError(
                    "GPTBlock cross-attention requires t_disease, h_token, "
                    "t_token, and token_mask."
                )
            x = self.cross_attn(
                h_disease=x,
                t_disease=t_disease,
                h_token=h_token,
                t_token=t_token,
                token_mask=token_mask,
            )
        x = x + self.mlp(self.ln2(x))
        return x


class TokenAutoDiscretization(nn.Module):
    def __init__(
        self,
        n_cont_types: int,
        n_bins: int,
        n_embd: int,
    ):
        super().__init__()
        if n_cont_types <= 0:
            raise ValueError(f"n_cont_types must be > 0, got {n_cont_types}")
        if n_bins <= 1:
            raise ValueError(f"n_bins must be > 1, got {n_bins}")
        if n_embd <= 0:
            raise ValueError(f"n_embd must be > 0, got {n_embd}")

        self.n_cont_types = n_cont_types
        self.n_bins = n_bins
        self.n_embd = n_embd
        self.weight = nn.Parameter(torch.empty(n_cont_types, n_bins))
        self.bias = nn.Parameter(torch.empty(n_cont_types, n_bins))
        self.bin_emb = nn.Parameter(torch.empty(n_cont_types, n_bins, n_embd))
        self.reset_parameters()

    def reset_parameters(self) -> None:
        nn.init.normal_(self.weight, mean=0.0, std=0.02)
        nn.init.zeros_(self.bias)
        nn.init.normal_(self.bin_emb, mean=0.0, std=0.02)

    def forward(
        self,
        cont_type_idx: torch.LongTensor,  # (N,)
        value: torch.Tensor,              # (N,)
    ) -> torch.Tensor:
        if cont_type_idx.dim() != 1:
            raise ValueError(
                f"cont_type_idx must be 1D, got {tuple(cont_type_idx.shape)}"
            )
        if value.dim() != 1:
            raise ValueError(f"value must be 1D, got {tuple(value.shape)}")
        if cont_type_idx.numel() != value.numel():
            raise ValueError("cont_type_idx and value must have the same length")

        w = self.weight[cont_type_idx]      # (N, n_bins)
        b = self.bias[cont_type_idx]        # (N, n_bins)
        e = self.bin_emb[cont_type_idx]     # (N, n_bins, D)
        logits = value.unsqueeze(-1) * w + b
        probs = torch.softmax(logits, dim=-1)
        return torch.einsum("nb,nbd->nd", probs, e)


class BaselineEncoder(nn.Module):
    PAD_KIND = 0
    CONT_KIND = 1
    CATE_KIND = 2

    def __init__(
        self,
        n_embd: int,
        n_head: int,
        n_types: int,
        n_cont_types: int,
        n_categories: int,
        cont_type_ids: list[int],
        n_value_kinds: int = 3,
        n_bins: int = 16,
        n_tab_layer: int = 2,
        dropout: float = 0.0,
    ):
        super().__init__()
        if len(cont_type_ids) != n_cont_types:
            raise ValueError(
                "cont_type_ids length must match n_cont_types, got "
                f"{len(cont_type_ids)} vs {n_cont_types}"
            )
        if n_types <= 0:
            raise ValueError(f"n_types must include PAD and be > 0, got {n_types}")
        if n_categories <= 0:
            raise ValueError(
                f"n_categories must include PAD and be > 0, got {n_categories}"
            )
        if n_value_kinds <= self.CATE_KIND:
            raise ValueError(
                f"n_value_kinds must be > {self.CATE_KIND}, got {n_value_kinds}"
            )

        self.n_embd = n_embd
        self.type_emb = nn.Embedding(n_types, n_embd, padding_idx=0)
        self.kind_emb = nn.Embedding(n_value_kinds, n_embd, padding_idx=0)
        self.cont_value_encoder = (
            TokenAutoDiscretization(
                n_cont_types=n_cont_types,
                n_bins=n_bins,
                n_embd=n_embd,
            )
            if n_cont_types > 0
            else None
        )
        self.cate_value_emb = nn.Embedding(
            n_categories,
            n_embd,
            padding_idx=0,
        )

        cont_type_index = torch.full((n_types,), -1, dtype=torch.long)
        for idx, type_id in enumerate(cont_type_ids):
            if type_id <= 0 or type_id >= n_types:
                raise ValueError(
                    f"continuous type id {type_id} must be in [1, {n_types})"
                )
            cont_type_index[type_id] = idx
        self.register_buffer(
            "cont_type_index",
            cont_type_index,
            persistent=False,
        )

        self.blocks = nn.ModuleList([
            GPTBlock(
                n_embd=n_embd,
                n_head=n_head,
                use_time_rope=False,
                use_rbf_bias=False,
                mlp_dropout=dropout,
            ) for _ in range(n_tab_layer)
        ])
        self.ln = nn.LayerNorm(n_embd)
        self.reset_parameters()

    def reset_parameters(self) -> None:
        nn.init.normal_(self.type_emb.weight, mean=0.0, std=0.02)
        nn.init.zeros_(self.type_emb.weight[0])
        nn.init.normal_(self.kind_emb.weight, mean=0.0, std=0.02)
        nn.init.zeros_(self.kind_emb.weight[0])
        nn.init.normal_(self.cate_value_emb.weight, mean=0.0, std=0.02)
        nn.init.zeros_(self.cate_value_emb.weight[0])

    def _make_attn_mask(self, mask: torch.Tensor, dtype: torch.dtype):
        return torch.zeros(
            mask.size(0),
            1,
            1,
            mask.size(1),
            device=mask.device,
            dtype=dtype,
        ).masked_fill(~mask[:, None, None, :], -1e4)

    def forward(
        self,
        other_type: torch.LongTensor,        # (B, K), 0 = padding
        other_value: torch.Tensor,           # (B, K), cate stores global id
        other_value_kind: torch.LongTensor,  # (B, K), 0=PAD, 1=CONT, 2=CATE
    ):
        if other_type.shape != other_value.shape:
            raise ValueError(
                "other_type and other_value must have the same shape, got "
                f"{tuple(other_type.shape)} vs {tuple(other_value.shape)}"
            )
        if other_type.shape != other_value_kind.shape:
            raise ValueError(
                "other_type and other_value_kind must have the same shape, got "
                f"{tuple(other_type.shape)} vs {tuple(other_value_kind.shape)}"
            )

        other_valid = other_type > 0
        type_emb = self.type_emb(other_type)
        kind_emb = self.kind_emb(other_value_kind)
        value_emb = torch.zeros_like(type_emb)

        cont_pos = other_valid & (other_value_kind == self.CONT_KIND)
        if cont_pos.any():
            if self.cont_value_encoder is None:
                raise ValueError("continuous tokens found but n_cont_types is 0")
            cont_idx = self.cont_type_index[other_type[cont_pos]]
            if (cont_idx < 0).any():
                bad_type = other_type[cont_pos][cont_idx < 0][0].item()
                raise ValueError(
                    f"type_id={bad_type} is marked continuous but is not in "
                    "cont_type_ids"
                )
            value_emb[cont_pos] = self.cont_value_encoder(
                cont_type_idx=cont_idx,
                value=other_value[cont_pos].to(type_emb.dtype),
            )

        cate_pos = other_valid & (other_value_kind == self.CATE_KIND)
        if cate_pos.any():
            cate_id = other_value[cate_pos].long()
            value_emb[cate_pos] = self.cate_value_emb(cate_id)

        f = type_emb + kind_emb + value_emb
        f = f * other_valid.unsqueeze(-1).to(f.dtype)

        if f.size(1) == 0:
            return f, other_valid

        attn_mask = self._make_attn_mask(other_valid, f.dtype)
        for block in self.blocks:
            f = block(f, attn_mask=attn_mask)
            f = f * other_valid.unsqueeze(-1).to(f.dtype)

        h = self.ln(f)
        h = h * other_valid.unsqueeze(-1).to(h.dtype)
        return h, other_valid


class CrossAttention(nn.Module):
    def __init__(
        self,
        n_embd: int,
        n_head: int,
        dropout: float = 0.0,
        n_rbf_bases: int = 16,
        max_time_diff: float = 40.0,
    ):
        super().__init__()
        assert n_embd % n_head == 0, "n_embd must be divisible by n_head"
        self.n_head = n_head
        self.d_head = n_embd // n_head
        self.scale = 1.0 / math.sqrt(self.d_head)
        self.dropout = dropout
        self.mask_value = -1e4

        self.q_proj = nn.Linear(n_embd, n_embd, bias=False)
        self.k_proj = nn.Linear(n_embd, n_embd, bias=False)
        self.v_proj = nn.Linear(n_embd, n_embd, bias=False)
        self.out_proj = nn.Linear(n_embd, n_embd, bias=False)

        self.time_rope = TimeRoPE(self.d_head)
        self.rbf_time_basis = GaussianRBFTimeBasis(
            n_bases=n_rbf_bases,
            max_time_diff=max_time_diff,
        )
        self.rbf_proj = nn.Linear(n_rbf_bases, n_head, bias=False)
        self.time_bias_scale = nn.Parameter(torch.tensor(0.0))
        self.resid_drop = nn.Dropout(dropout)
        self.ln = nn.LayerNorm(n_embd)
        self.out_ln = nn.LayerNorm(n_embd)
        self.reset_parameters()

    def reset_parameters(self) -> None:
        nn.init.normal_(self.q_proj.weight, mean=0.0, std=0.02)
        nn.init.normal_(self.k_proj.weight, mean=0.0, std=0.02)
        nn.init.normal_(self.v_proj.weight, mean=0.0, std=0.02)
        nn.init.normal_(self.out_proj.weight, mean=0.0, std=0.02)
        nn.init.normal_(self.rbf_proj.weight, mean=0.0, std=0.02)

    def _make_attn_mask(
        self,
        token_mask: torch.Tensor,  # (B, K), True = valid
        t_disease: torch.Tensor,     # (B, L)
        t_token: torch.Tensor,       # (B, K)
        dtype: torch.dtype,
    ):
        valid_token = token_mask[:, None, :]  # (B, 1, K)
        visible_by_time = t_token[:, None, :] <= t_disease[:, :, None]
        valid = visible_by_time & valid_token  # (B, L, K)

        has_any_valid = valid.any(dim=-1, keepdim=True)
        safe_valid = valid.clone()
        safe_valid[..., 0:1] = torch.where(
            has_any_valid,
            safe_valid[..., 0:1],
            torch.ones_like(safe_valid[..., 0:1]),
        )

        attn_mask = torch.zeros(
            safe_valid.shape,
            device=safe_valid.device,
            dtype=dtype,
        ).masked_fill(~safe_valid, self.mask_value)
        return attn_mask[:, None, :, :], valid

    def _cross_rbf_cache(
        self,
        t_disease: torch.Tensor,
        t_token: torch.Tensor,
    ) -> torch.Tensor:
        tau = torch.cat([t_disease, t_token], dim=1)
        rbf_cache = self.rbf_time_basis.precompute_cache(tau)
        n_disease = t_disease.size(1)
        return rbf_cache[:, :n_disease, n_disease:, :]

    def forward(
        self,
        h_disease: torch.Tensor,          # (B, L, D)
        t_disease: torch.Tensor,          # (B, L)
        h_token: torch.Tensor,            # (B, K, D)
        t_token: torch.Tensor,            # (B, K)
        token_mask: torch.Tensor,         # (B, K), True = valid
        need_weights: bool = False,
    ):
        B, L, _ = h_disease.shape
        K = h_token.size(1)
        H, Dh = self.n_head, self.d_head
        if K == 0:
            empty_weights = h_disease.new_zeros(B, L, 0)
            if need_weights:
                return h_disease, empty_weights
            return h_disease

        attn_mask, valid = self._make_attn_mask(
            token_mask=token_mask.to(device=h_disease.device, dtype=torch.bool),
            t_disease=t_disease,
            t_token=t_token,
            dtype=h_disease.dtype,
        )

        q = self.q_proj(h_disease).reshape(B, L, H, Dh).transpose(1, 2)
        k = self.k_proj(h_token).reshape(B, K, H, Dh).transpose(1, 2)
        v = self.v_proj(h_token).reshape(B, K, H, Dh).transpose(1, 2)

        q_cos, q_sin = self.time_rope.precompute_cache(t_disease)
        k_cos, k_sin = self.time_rope.precompute_cache(t_token)
        q = q * q_cos + TimeRoPE._rotate_half(q) * q_sin
        k = k * k_cos + TimeRoPE._rotate_half(k) * k_sin

        rbf_cache = self._cross_rbf_cache(t_disease, t_token)  # (B, L, K, R)
        time_bias = self.rbf_proj(rbf_cache).permute(0, 3, 1, 2)
        time_bias = self.time_bias_scale.tanh() * time_bias

        attn_bias = attn_mask.to(time_bias.dtype) + time_bias
        attn_out = F.scaled_dot_product_attention(
            q,
            k,
            v,
            attn_mask=attn_bias,
            dropout_p=self.dropout if self.training else 0.0,
            is_causal=False,
            scale=self.scale,
        )
        attn_out = attn_out.transpose(1, 2).reshape(B, L, H * Dh)
        attn_out = self.resid_drop(self.out_proj(attn_out))

        has_any_valid = valid.any(dim=-1)  # (B, L)
        attn_out = attn_out * has_any_valid.unsqueeze(-1).to(attn_out.dtype)

        attn_scores = torch.matmul(q, k.transpose(-2, -1)) * self.scale
        attn_scores = attn_scores + attn_bias
        attn_weights = torch.softmax(attn_scores, dim=-1).mean(dim=1)
        attn_weights = attn_weights * valid.to(attn_weights.dtype)
        weight_sum = attn_weights.sum(dim=-1, keepdim=True).clamp_min(1e-12)
        norm_weights = attn_weights / weight_sum
        norm_weights = norm_weights * has_any_valid.unsqueeze(-1).to(
            norm_weights.dtype
        )
        out = self.out_ln(h_disease + attn_out)
        out = torch.where(has_any_valid.unsqueeze(-1), out, h_disease)

        if need_weights:
            return out, norm_weights
        return out


class AgeSinusoidalEncoding(nn.Module):

    def __init__(self, embedding_dim: int):

        super().__init__()
        if embedding_dim % 2 != 0:
            raise ValueError(
                f"Embedding dimension must be an even number, but got {embedding_dim}")

        self.embedding_dim = embedding_dim

        i = torch.arange(0, self.embedding_dim, 2, dtype=torch.float32)
        divisor = torch.pow(10000, i / self.embedding_dim)
        self.register_buffer('divisor', divisor)
        self.linear = nn.Linear(embedding_dim, embedding_dim, bias=False)

    def forward(self, t: torch.Tensor) -> torch.Tensor:

        t_years = t
        # Broadcast (B, L, 1) against (1, 1, D/2) to get (B, L, D/2)
        args = t_years.unsqueeze(-1) / self.divisor.view(1, 1, -1)
        # Interleave cos and sin along the last dimension
        output = torch.zeros(t.shape[0], t.shape[1],
                             self.embedding_dim, device=t.device)
        output[:, :, 0::2] = torch.cos(args)
        output[:, :, 1::2] = torch.sin(args)
        output = self.linear(output)
        return output