README.md

To use these checkpoints, you need to use the following model structure for Transformer

Import used packages

import math

import torch
from torch import nn

PositionalEncoding

class PositionalEncoding(nn.Module):
    def __init__(self, d_model: int, dropout: float = 0.1, max_len: int = 5000) -> None:
        super().__init__()
        self.dropout = nn.Dropout(p=dropout)

        pe = torch.zeros(max_len, d_model)
        position = torch.arange(0, max_len, dtype=torch.float).unsqueeze(
            1
        )  # (max_len, 1)
        div_term = torch.exp(
            torch.arange(0, d_model, 2).float() * (-math.log(10000.0) / d_model)
        )
        pe[:, 0::2] = torch.sin(position * div_term)  # (max_len, d_model // 2)
        truncated_div_term = div_term[: d_model // 2]
        pe[:, 1::2] = torch.cos(position * truncated_div_term)  #
        pe = pe.unsqueeze(0).transpose(0, 1)  # (max_len, 1, d_model)
        self.register_buffer("pe", pe)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        x = x + self.pe[: x.size(0), :, :]
        return self.dropout(x)

RPBClass

class RelativePositionBiasV2(nn.Module):
    def __init__(self, n_heads, num_buckets=32, max_distance=128, bidirectional=True):
        super().__init__()
        assert num_buckets % 2 == 0, "num_buckets should be even for bidirectional"
        self.n_heads = n_heads
        self.num_buckets = num_buckets
        self.max_distance = max_distance
        self.bidirectional = bidirectional
        self.emb = nn.Embedding(num_buckets, n_heads)

    def _relative_position_bucket(self, relative_position):
        """
        relative_position: [Tq, Tk] = k - q
        returns bucket ids in [0, num_buckets-1]
        """
        num_buckets = self.num_buckets
        max_distance = self.max_distance

        ret = torch.zeros_like(relative_position, dtype=torch.long)
        n = -relative_position  # want smaller buckets for n > 0 (keys before queries)

        if self.bidirectional:
            half = num_buckets // 2
            ret += (n < 0).long() * half
            n = n.abs()
            num_buckets = half  # remaining buckets for non-negative distances
        else:
            n = torch.clamp(n, min=0)

        # Now n >= 0
        max_exact = num_buckets // 2
        is_small = n < max_exact
        # Avoid log(0) and division by zero; also ensure max_distance > max_exact
        denom = max(1.0, math.log(max(max_distance, max_exact + 1) / max(1, max_exact)))
        val_if_large = (
            max_exact
            + (
                (torch.log(n.float() / max(1, max_exact) + 1e-6) / denom)
                * (num_buckets - max_exact)
            ).long()
        )
        val_if_large = torch.clamp(val_if_large, max=num_buckets - 1)

        ret += torch.where(is_small, n.long(), val_if_large)
        # Final clamp for absolute safety when bidirectional half-split was applied
        return torch.clamp(ret, min=0, max=self.num_buckets - 1)

    def forward(self, Tq, Tk, device=None):
        device = device or torch.device("cpu")
        qpos = torch.arange(Tq, device=device)[:, None]
        kpos = torch.arange(Tk, device=device)[None, :]
        buckets = self._relative_position_bucket(kpos - qpos)  # [Tq, Tk]
        bias = self.emb(buckets)  # [Tq, Tk, H]
        return bias.permute(2, 0, 1)  # [H, Tq, Tk]

Transformer Base Class

class BaseTransformerComp(nn.Module):
    """Base class for transformer-based intra-stock components."""

    def __init__(
        self,
        input_dim: int,
        hidden_dim: int,
        num_layers: int,
        num_heads: int,
        dropout: float = 0.1,
        mask_type: str = "none",
    ) -> None:
        super().__init__()
        self.input_dim = input_dim
        self.hidden_dim = hidden_dim
        self.num_layers = num_layers
        self.num_heads = num_heads
        self.dropout_rate = dropout
        self.mask_type = mask_type

    def _reshape_input(self, x: torch.Tensor) -> tuple[torch.Tensor, int, int]:
        """
        Reshape input from [batch, seq_len, n_stocks, n_feats] to [seq_len, batch*n_stocks, n_feats].
        Returns reshaped tensor and original batch/n_stocks sizes for later reconstruction.
        """
        batch, seq_len, n_stocks, n_feats = x.shape

        if batch == 0 or seq_len == 0 or n_stocks == 0:
            raise ValueError(
                f"Invalid input dimensions: batch={batch}, seq_len={seq_len}, "
                f"n_stocks={n_stocks}, n_feats={n_feats}"
            )

        x = x.permute(0, 2, 1, 3).contiguous()
        x = x.reshape(batch * n_stocks, seq_len, n_feats)  # [b * s, t, f]
        x = x.permute(1, 0, 2).contiguous()  # [t, b * s, f]

        return x, batch, n_stocks

    def _reshape_output(
        self, x: torch.Tensor, batch: int, n_stocks: int
    ) -> torch.Tensor:
        """Reshape output from [seq_len, batch*n_stocks, hidden_dim] to [batch, n_stocks, hidden_dim]."""
        output = x[-1]  # Take last time step: [b * s, hidden_dim]
        output = output.reshape(batch, n_stocks, -1)  # [b, s, hidden_dim]
        return output

    def _generate_causal_mask(self, seq_len: int, device: torch.device) -> torch.Tensor:
        """Generate causal attention mask."""
        mask = torch.triu(
            torch.ones(seq_len, seq_len, device=device) * float("-inf"), diagonal=1
        )
        return mask

Transformer Encoder Layer with RPB

class TransformerEncoderLayerWithRPB(nn.Module):
    def __init__(
        self,
        d_model: int,
        nhead: int,
        dim_feedforward: int,
        dropout: float,
        rbp,
    ):
        super().__init__()
        self.d_model = d_model
        self.nhead = nhead
        self.rbp = rbp

        # QKV projections
        self.qkv_proj = nn.Linear(d_model, 3 * d_model)
        self.out_proj = nn.Linear(d_model, d_model)

        # FFN layers
        self.linear1 = nn.Linear(d_model, dim_feedforward)
        self.dropout = nn.Dropout(dropout)
        self.linear2 = nn.Linear(dim_feedforward, d_model)

        # Normalization and dropout
        self.norm1 = nn.LayerNorm(d_model)
        self.norm2 = nn.LayerNorm(d_model)
        self.dropout1 = nn.Dropout(dropout)
        self.dropout2 = nn.Dropout(dropout)
        self.activation = F.relu

    def forward(
        self,
        src: torch.Tensor,
        src_mask: Optional[torch.Tensor] = None,
        src_key_padding_mask: Optional[torch.Tensor] = None,
        is_causal: bool = False,
    ) -> torch.Tensor:
        seq_len, batch_size, d_model = src.shape
        head_dim = d_model // self.nhead
        qkv = self.qkv_proj(src)
        q, k, v = qkv.chunk(3, dim=-1)
        q = q.reshape(seq_len, batch_size, self.nhead, head_dim).permute(1, 2, 0, 3)
        k = k.reshape(seq_len, batch_size, self.nhead, head_dim).permute(1, 2, 0, 3)
        v = v.reshape(seq_len, batch_size, self.nhead, head_dim).permute(1, 2, 0, 3)
        attn_weights = torch.matmul(q, k.transpose(-2, -1)) / math.sqrt(head_dim)

        # Add RBP after QK^T
        rbp_bias = self.rbp(
            seq_len, seq_len, device=src.device
        )  # [nhead, seq_len, seq_len]
        attn_weights = attn_weights + rbp_bias.unsqueeze(
            0
        )  # [batch, nhead, seq_len, seq_len]

        if src_mask is not None:
            attn_weights = attn_weights + src_mask.unsqueeze(0).unsqueeze(0)

        if src_key_padding_mask is not None:
            attn_weights = attn_weights.masked_fill(
                src_key_padding_mask.unsqueeze(1).unsqueeze(2), float("-inf")
            )

        attn_weights = F.softmax(attn_weights, dim=-1)
        attn_weights = self.dropout1(attn_weights)
        attn_output = torch.matmul(attn_weights, v)  # [batch, nhead, seq_len, head_dim]
        attn_output = attn_output.permute(2, 0, 1, 3).reshape(
            seq_len, batch_size, d_model
        )
        attn_output = self.out_proj(attn_output)
        src2 = src + self.dropout1(attn_output)
        src2 = self.norm1(src2)
        ffn_output = self.linear2(self.dropout(self.activation(self.linear1(src2))))
        src3 = src2 + self.dropout2(ffn_output)
        src3 = self.norm2(src3)

        return src3

RPB Components

class TransformerRPBComp(BaseTransformerComp):
    """TransformerComp with Relative Bias Pooling."""

    def __init__(
        self,
        input_dim: int,
        hidden_dim: int,
        num_layers: int,
        num_heads: int,
        dropout: float = 0.1,
        mask_type: str = "none",
    ) -> None:
        super().__init__(input_dim, hidden_dim, num_layers, num_heads, dropout)
        self.feature_layer = nn.Linear(input_dim, hidden_dim)
        self.pe = PositionalEncoding(hidden_dim, dropout)
        self.encoder_norm = nn.LayerNorm(hidden_dim)
        self.mask_type = mask_type
        self.rbp = RelativePositionBiasV2(n_heads=num_heads)
        self.encoder_layers = nn.ModuleList(
            [
                TransformerEncoderLayerWithRPB(
                    d_model=hidden_dim,
                    nhead=num_heads,
                    dim_feedforward=hidden_dim * 4,
                    dropout=dropout,
                    rbp=self.rbp,
                )
                for _ in range(num_layers)
            ]
        )

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        """x.shape [batch, seq_len, n_stocks, n_feats]"""
        x, batch, n_stocks = self._reshape_input(x)
        seq_len = x.shape[0]

        x = self.encoder_norm(self.pe(self.feature_layer(x)))  # [t, b * s, d_model]

        if self.mask_type == "causal":
            mask = self._generate_causal_mask(seq_len, x.device).permute(1, 0)
        else:
            mask = None

        for layer in self.encoder_layers:
            x = layer(x, src_mask=mask)

        return self._reshape_output(x, batch, n_stocks)

Transformer Module

class Transformer(nn.Module):
    def __init__(
        self,
        input_dim: int,
        output_dim: int = 1,
        hidden_dim: int = 256,
        num_layers: int = 2,
        num_heads: int = 4,
        dropout: float = 0.1,
        tfm_type: str = "base",
        mask_type: str = "none",
    ) -> None:
        """
        tfm_type: "base", "rope", "rpb"
        mask_type: "none", "alibi", "causal"
        """
        super().__init__()
        self.tfm_type = tfm_type
        self.mask_type = mask_type

        tfm_type_mapper = {
            "base": TransformerComp,
            "alibi": TransformerComp,
            "rope": TransformerRoPEComp,
            "rpb": TransformerRPBComp,
        }
        self.transformer_encoder = tfm_type_mapper[self.tfm_type](
            input_dim=input_dim,
            hidden_dim=hidden_dim,
            num_layers=num_layers,
            num_heads=num_heads,
            dropout=dropout,
            mask_type=mask_type,
        )
        self.fc_out = nn.Sequential(
            nn.Linear(hidden_dim, hidden_dim, bias=True),
            nn.GELU(),
            nn.Linear(hidden_dim, output_dim, bias=True),
        )

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        tfm_out = self.transformer_encoder(x)  # [b, s, d_model]
        final_out = self.fc_out(tfm_out).squeeze(-1)  # [b, s]

        return final_out

Model Configuration

input_dim: 8,
output_dim: 1,
hidden_dim: 64,
num_layers: 2,
num_heads: 4,
dropout: 0.0,
tfm_type: "rpb",
mask_type: "causal",