models/QFormerAdapter.py

import torch
from torch import nn
from utils.position_coding import (
    LearnablePositionalEmbedding,
    RotaryPositionalEncoding,
    SinusoidalPositionalEncoding,
)


class QFormerBlock(nn.Module):
    def __init__(self, dim, num_heads, dropout=0.1, mlp_ratio=4.0):
        super().__init__()
        self.query_norm = nn.LayerNorm(dim)
        self.self_attn = nn.MultiheadAttention(
            embed_dim=dim,
            num_heads=num_heads,
            dropout=dropout,
            batch_first=True,
        )

        self.cross_query_norm = nn.LayerNorm(dim)
        self.memory_norm = nn.LayerNorm(dim)
        self.cross_attn = nn.MultiheadAttention(
            embed_dim=dim,
            num_heads=num_heads,
            dropout=dropout,
            batch_first=True,
        )

        hidden_dim = int(dim * mlp_ratio)
        self.ffn = nn.Sequential(
            nn.LayerNorm(dim),
            nn.Linear(dim, hidden_dim),
            nn.GELU(),
            nn.Dropout(dropout),
            nn.Linear(hidden_dim, dim),
            nn.Dropout(dropout),
        )

    def forward(self, queries, memory):
        query_residual = queries
        norm_queries = self.query_norm(queries)
        queries = query_residual + self.self_attn(
            norm_queries,
            norm_queries,
            norm_queries,
            need_weights=False,
        )[0]

        query_residual = queries
        queries = query_residual + self.cross_attn(
            self.cross_query_norm(queries),
            self.memory_norm(memory),
            self.memory_norm(memory),
            need_weights=False,
        )[0]

        queries = queries + self.ffn(queries)
        return queries


class QFormerAdapter(nn.Module):
    """Stage-aware Q-Former adapter baseline.

    It keeps the Q-Former learnable query tokens, but preserves the time-series
    structure before flattening memory for cross-attention. Stage 3 and 4 use
    cycle positional encoding so cross-cycle degradation trends remain visible.
    """

    def __init__(self, args):
        super().__init__()
        self.prefix_num = args.prefix_num
        self.prefix_token = nn.Parameter(torch.randn(1, args.prefix_num, args.it_d_model))
        self.instruction_norm = nn.LayerNorm(args.it_d_model)
        self.stage_embedding = nn.Embedding(5, args.it_d_model)
        self.memory_norm = nn.LayerNorm(args.it_d_model)
        self.time_pos = SinusoidalPositionalEncoding(args.it_d_model)
        self.var_pos = LearnablePositionalEmbedding(args.it_d_model)
        self.cycle_pos = RotaryPositionalEncoding(args.it_d_model)
        self.layers = nn.ModuleList(
            [
                QFormerBlock(
                    dim=args.it_d_model,
                    num_heads=args.it_n_heads,
                    dropout=args.it_dropout,
                    mlp_ratio=4.0,
                )
                for _ in range(args.it_layers)
            ]
        )
        self.norm = nn.LayerNorm(args.it_d_model)

    def _stage_list(self, stage, batch_size, device):
        if stage is None:
            return [0] * batch_size
        if torch.is_tensor(stage):
            values = stage.detach().to(device=device, dtype=torch.long).view(-1).tolist()
        elif isinstance(stage, (int, float)):
            values = [int(stage)]
        else:
            values = [int(s) for s in stage]

        if len(values) == 1 and batch_size > 1:
            return values * batch_size
        if len(values) != batch_size:
            raise ValueError(f"Expected {batch_size} stage values, got {len(values)}.")
        return values

    def _encode_memory(self, memory, stage):
        batch_size = memory.shape[0]
        stage_list = self._stage_list(stage, batch_size, memory.device)

        cycle_index = [i for i, s in enumerate(stage_list) if s not in (3, 4)]
        cross_cycle_index = [i for i, s in enumerate(stage_list) if s in (3, 4)]
        original_indices = cycle_index + cross_cycle_index
        reorder_map = {idx: i for i, idx in enumerate(original_indices)}
        reverse_indices = [reorder_map[i] for i in range(batch_size)]

        processed_memories = []

        if cycle_index:
            sub_memory = memory[cycle_index]
            b, l, v, d = sub_memory.shape
            sub_memory = sub_memory.view(b * l, v, d)
            sub_memory = sub_memory + self.time_pos(sub_memory)
            sub_memory = sub_memory.view(b, l, v, d)
            processed_memories.append(sub_memory)

        if cross_cycle_index:
            sub_memory = memory[cross_cycle_index]
            b, l, v, d = sub_memory.shape
            sub_memory = sub_memory.view(b * v, l, d)
            sub_memory = sub_memory + self.cycle_pos(sub_memory)
            sub_memory = sub_memory.view(b, l, v, d)
            processed_memories.append(sub_memory)

        memory = torch.cat(processed_memories, dim=0)[reverse_indices]

        b, l, v, d = memory.shape
        memory = memory.view(b * l, v, d)
        memory = memory + self.var_pos(memory)
        memory = memory.view(b, l * v, d)
        return self.memory_norm(memory)

    def forward(self, x, memory, stage=None, attn_mask=None):
        batch_size = x.shape[0]
        queries = self.prefix_token.repeat(batch_size, 1, 1)

        if x is not None and x.numel() > 0:
            instruction_bias = self.instruction_norm(x).mean(dim=1, keepdim=True)
            queries = queries + instruction_bias

        stage_ids = torch.tensor(
            self._stage_list(stage, batch_size, queries.device),
            device=queries.device,
            dtype=torch.long,
        ).clamp(min=0, max=self.stage_embedding.num_embeddings - 1)
        queries = queries + self.stage_embedding(stage_ids).unsqueeze(1)

        memory = self._encode_memory(memory, stage)

        for layer in self.layers:
            queries = layer(queries, memory)

        return self.norm(queries)