update新结构

example/README.md

@@ -72,3 +72,5 @@ python example/run_pipeline.py --device auto
 - The script only samples the first 5000 rows to stay fast.
 - `prepare_data.py` runs without PyTorch, but `train.py` and `sample.py` require it.
 - `train.py` and `sample.py` auto-select GPU if available; otherwise they fall back to CPU.
+- Optional feature-graph mixer (`model_use_feature_graph`) adds a learnable relation prior across feature channels.
+- Optional two-stage temporal backbone (`use_temporal_stage1`) learns a GRU-based sequence trend; diffusion models the residual.

example/config.json

@@ -32,11 +32,24 @@
   "model_ff_mult": 2,
   "model_pos_dim": 64,
   "model_use_pos_embed": true,
+  "model_use_feature_graph": true,
+  "feature_graph_scale": 0.1,
+  "feature_graph_dropout": 0.0,
+  "use_temporal_stage1": true,
+  "temporal_hidden_dim": 256,
+  "temporal_num_layers": 1,
+  "temporal_dropout": 0.0,
+  "temporal_loss_weight": 1.0,
   "disc_mask_scale": 0.9,
   "cont_loss_weighting": "inv_std",
   "cont_loss_eps": 1e-6,
-  "cont_target": "x0",
+  "cont_target": "v",
   "cont_clamp_x0": 5.0,
+  "quantile_loss_weight": 0.1,
+  "quantile_points": [0.05, 0.25, 0.5, 0.75, 0.95],
+  "quantile_loss_warmup_steps": 200,
+  "quantile_loss_clip": 6.0,
+  "quantile_loss_huber_delta": 1.0,
   "shuffle_buffer": 256,
   "sample_batch_size": 8,
   "sample_seq_len": 128

example/export_samples.py

@@ -13,7 +13,7 @@ import torch
 import torch.nn.functional as F
 
 from data_utils import load_split
-from hybrid_diffusion import HybridDiffusionModel, cosine_beta_schedule
+from hybrid_diffusion import HybridDiffusionModel, TemporalGRUGenerator, cosine_beta_schedule
 from platform_utils import resolve_device, safe_path, ensure_dir, resolve_path
 
 
@@ -140,6 +140,10 @@ def main():
                 raise SystemExit("use_condition enabled but no files matched data_glob: %s" % cfg_glob)
     cont_target = str(cfg.get("cont_target", "eps"))
     cont_clamp_x0 = float(cfg.get("cont_clamp_x0", 0.0))
+    use_temporal_stage1 = bool(cfg.get("use_temporal_stage1", False))
+    temporal_hidden_dim = int(cfg.get("temporal_hidden_dim", 256))
+    temporal_num_layers = int(cfg.get("temporal_num_layers", 1))
+    temporal_dropout = float(cfg.get("temporal_dropout", 0.0))
 
     model = HybridDiffusionModel(
         cont_dim=len(cont_cols),
@@ -151,6 +155,9 @@ def main():
         ff_mult=int(cfg.get("model_ff_mult", 2)),
         pos_dim=int(cfg.get("model_pos_dim", 64)),
         use_pos_embed=bool(cfg.get("model_use_pos_embed", True)),
+        use_feature_graph=bool(cfg.get("model_use_feature_graph", False)),
+        feature_graph_scale=float(cfg.get("feature_graph_scale", 0.1)),
+        feature_graph_dropout=float(cfg.get("feature_graph_dropout", 0.0)),
         cond_vocab_size=cond_vocab_size if use_condition else 0,
         cond_dim=int(cfg.get("cond_dim", 32)),
         use_tanh_eps=bool(cfg.get("use_tanh_eps", False)),
@@ -163,6 +170,20 @@ def main():
         model.load_state_dict(torch.load(args.model_path, map_location=device, weights_only=True))
     model.eval()
 
+    temporal_model = None
+    if use_temporal_stage1:
+        temporal_model = TemporalGRUGenerator(
+            input_dim=len(cont_cols),
+            hidden_dim=temporal_hidden_dim,
+            num_layers=temporal_num_layers,
+            dropout=temporal_dropout,
+        ).to(device)
+        temporal_path = Path(args.model_path).with_name("temporal.pt")
+        if not temporal_path.exists():
+            raise SystemExit(f"missing temporal model file: {temporal_path}")
+        temporal_model.load_state_dict(torch.load(temporal_path, map_location=device, weights_only=True))
+        temporal_model.eval()
+
     betas = cosine_beta_schedule(args.timesteps).to(device)
     alphas = 1.0 - betas
     alphas_cumprod = torch.cumprod(alphas, dim=0)
@@ -189,6 +210,10 @@ def main():
             cond_id = torch.full((args.batch_size,), int(args.condition_id), device=device, dtype=torch.long)
         cond = cond_id
 
+    trend = None
+    if temporal_model is not None:
+        trend = temporal_model.generate(args.batch_size, args.seq_len, device)
+
     for t in reversed(range(args.timesteps)):
         t_batch = torch.full((args.batch_size,), t, device=device, dtype=torch.long)
         eps_pred, logits = model(x_cont, x_disc, t_batch, cond)
@@ -201,6 +226,10 @@ def main():
             if cont_clamp_x0 > 0:
                 x0_pred = torch.clamp(x0_pred, -cont_clamp_x0, cont_clamp_x0)
             eps_pred = (x_cont - torch.sqrt(a_bar_t) * x0_pred) / torch.sqrt(1.0 - a_bar_t)
+        elif cont_target == "v":
+            v_pred = eps_pred
+            x0_pred = torch.sqrt(a_bar_t) * x_cont - torch.sqrt(1.0 - a_bar_t) * v_pred
+            eps_pred = torch.sqrt(1.0 - a_bar_t) * x_cont + torch.sqrt(a_bar_t) * v_pred
         coef1 = 1.0 / torch.sqrt(a_t)
         coef2 = (1 - a_t) / torch.sqrt(1 - a_bar_t)
         mean_x = coef1 * (x_cont - coef2 * eps_pred)
@@ -225,6 +254,8 @@ def main():
                     )
                     x_disc[:, :, i][mask] = sampled[mask]
 
+    if trend is not None:
+        x_cont = x_cont + trend
     # move to CPU for export
     x_cont = x_cont.cpu()
     x_disc = x_disc.cpu()

example/hybrid_diffusion.py

@@ -66,6 +66,64 @@ class SinusoidalTimeEmbedding(nn.Module):
         return emb
 
 
+class FeatureGraphMixer(nn.Module):
+    """Learnable feature relation mixer (dataset-agnostic)."""
+
+    def __init__(self, dim: int, scale: float = 0.1, dropout: float = 0.0):
+        super().__init__()
+        self.scale = scale
+        self.dropout = nn.Dropout(dropout) if dropout > 0 else nn.Identity()
+        self.A = nn.Parameter(torch.zeros(dim, dim))
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        # x: (B, T, D)
+        # Symmetric relation to stabilize
+        A = (self.A + self.A.t()) * 0.5
+        mixed = torch.matmul(x, A) * self.scale
+        mixed = self.dropout(mixed)
+        return x + mixed
+
+
+class TemporalGRUGenerator(nn.Module):
+    """Stage-1 temporal generator (autoregressive GRU) for sequence backbone."""
+
+    def __init__(self, input_dim: int, hidden_dim: int = 256, num_layers: int = 1, dropout: float = 0.0):
+        super().__init__()
+        self.start_token = nn.Parameter(torch.zeros(input_dim))
+        self.gru = nn.GRU(
+            input_dim,
+            hidden_dim,
+            num_layers=num_layers,
+            dropout=dropout if num_layers > 1 else 0.0,
+            batch_first=True,
+        )
+        self.out = nn.Linear(hidden_dim, input_dim)
+
+    def forward_teacher(self, x: torch.Tensor) -> torch.Tensor:
+        """Teacher-forced next-step prediction. Returns trend sequence and preds."""
+        if x.size(1) < 2:
+            raise ValueError("sequence length must be >= 2 for teacher forcing")
+        inp = x[:, :-1, :]
+        out, _ = self.gru(inp)
+        pred_next = self.out(out)
+        trend = torch.zeros_like(x)
+        trend[:, 0, :] = x[:, 0, :]
+        trend[:, 1:, :] = pred_next
+        return trend, pred_next
+
+    def generate(self, batch_size: int, seq_len: int, device: torch.device) -> torch.Tensor:
+        """Autoregressively generate a backbone sequence."""
+        h = None
+        prev = self.start_token.unsqueeze(0).expand(batch_size, -1).to(device)
+        outputs = []
+        for _ in range(seq_len):
+            out, h = self.gru(prev.unsqueeze(1), h)
+            nxt = self.out(out.squeeze(1))
+            outputs.append(nxt.unsqueeze(1))
+            prev = nxt
+        return torch.cat(outputs, dim=1)
+
+
 class HybridDiffusionModel(nn.Module):
     def __init__(
         self,
@@ -78,6 +136,9 @@ class HybridDiffusionModel(nn.Module):
         ff_mult: int = 2,
         pos_dim: int = 64,
         use_pos_embed: bool = True,
+        use_feature_graph: bool = False,
+        feature_graph_scale: float = 0.1,
+        feature_graph_dropout: float = 0.0,
         cond_vocab_size: int = 0,
         cond_dim: int = 32,
         use_tanh_eps: bool = False,
@@ -92,6 +153,7 @@ class HybridDiffusionModel(nn.Module):
         self.eps_scale = eps_scale
         self.pos_dim = pos_dim
         self.use_pos_embed = use_pos_embed
+        self.use_feature_graph = use_feature_graph
 
         self.cond_vocab_size = cond_vocab_size
         self.cond_dim = cond_dim
@@ -106,8 +168,17 @@ class HybridDiffusionModel(nn.Module):
         disc_embed_dim = sum(e.embedding_dim for e in self.disc_embeds)
 
         self.cont_proj = nn.Linear(cont_dim, cont_dim)
+        self.feature_dim = cont_dim + disc_embed_dim
+        if use_feature_graph:
+            self.feature_graph = FeatureGraphMixer(
+                self.feature_dim,
+                scale=feature_graph_scale,
+                dropout=feature_graph_dropout,
+            )
+        else:
+            self.feature_graph = None
         pos_dim = pos_dim if use_pos_embed else 0
-        in_dim = cont_dim + disc_embed_dim + time_dim + pos_dim + (cond_dim if self.cond_embed is not None else 0)
+        in_dim = self.feature_dim + time_dim + pos_dim + (cond_dim if self.cond_embed is not None else 0)
         self.in_proj = nn.Linear(in_dim, hidden_dim)
         self.backbone = nn.GRU(
             hidden_dim,
@@ -149,7 +220,10 @@ class HybridDiffusionModel(nn.Module):
             cond_feat = self.cond_embed(cond).unsqueeze(1).expand(-1, x_cont.size(1), -1)
 
         cont_feat = self.cont_proj(x_cont)
-        parts = [cont_feat, disc_feat, time_emb]
+        feat = torch.cat([cont_feat, disc_feat], dim=-1)
+        if self.feature_graph is not None:
+            feat = self.feature_graph(feat)
+        parts = [feat, time_emb]
         if pos_emb is not None:
             parts.append(pos_emb.unsqueeze(0).expand(x_cont.size(0), -1, -1))
         if cond_feat is not None:

example/plot_loss.py

@@ -34,6 +34,7 @@ def main():
     loss = []
     loss_cont = []
     loss_disc = []
+    loss_quant = []
 
     with log_path.open("r", encoding="utf-8", newline="") as f:
         reader = csv.DictReader(f)
@@ -42,6 +43,8 @@ def main():
             loss.append(float(row["loss"]))
             loss_cont.append(float(row["loss_cont"]))
             loss_disc.append(float(row["loss_disc"]))
+            if "loss_quantile" in row:
+                loss_quant.append(float(row["loss_quantile"]))
 
     if not steps:
         raise SystemExit("no rows in log file: %s" % log_path)
@@ -50,6 +53,8 @@ def main():
     plt.plot(steps, loss, label="total")
     plt.plot(steps, loss_cont, label="continuous")
     plt.plot(steps, loss_disc, label="discrete")
+    if loss_quant:
+        plt.plot(steps, loss_quant, label="quantile")
     plt.xlabel("step")
     plt.ylabel("loss")
     plt.title("Training Loss")

example/results/config_used.json

@@ -32,12 +32,26 @@
   "model_ff_mult": 2,
   "model_pos_dim": 64,
   "model_use_pos_embed": true,
+  "model_use_feature_graph": true,
+  "feature_graph_scale": 0.1,
+  "feature_graph_dropout": 0.0,
   "disc_mask_scale": 0.9,
   "shuffle_buffer": 256,
   "cont_loss_weighting": "inv_std",
   "cont_loss_eps": 1e-06,
-  "cont_target": "x0",
+  "cont_target": "v",
   "cont_clamp_x0": 5.0,
+  "quantile_loss_weight": 0.1,
+  "quantile_points": [
+    0.05,
+    0.25,
+    0.5,
+    0.75,
+    0.95
+  ],
+  "quantile_loss_warmup_steps": 200,
+  "quantile_loss_clip": 6.0,
+  "quantile_loss_huber_delta": 1.0,
   "sample_batch_size": 8,
   "sample_seq_len": 128
 }

example/results/train_log.csv

@@ -1,61 +1,61 @@
-epoch,step,loss,loss_cont,loss_disc
+epoch,step,loss,loss_cont,loss_disc,loss_quantile
-0,0,9801.450195,14001.333008,1.724242
+0,0,11527.750000,16467.474609,1.724242,0.113109
-0,10,8648.388672,12354.390625,1.050415
+0,10,8955.048828,12792.469727,1.065955,0.142707
-0,20,6285.666992,8979.191406,0.775979
+0,20,123049.507812,175784.562500,1.048191,0.181546
-0,30,6296.939941,8995.372070,0.598658
+0,30,30947.671875,44210.511719,1.034107,0.224363
-0,40,7126.128906,10180.011719,0.402291
+0,40,11166.339844,15951.691406,0.512211,0.163978
-0,50,5381.071289,7687.099121,0.340463
+0,50,14919.276367,21313.089844,0.369016,0.101041
-1,0,17876.904297,25538.212891,0.522857
+1,0,1113425.000000,1590606.875000,0.667305,0.236102
-1,10,9174.909180,13106.868164,0.338315
+1,10,38804.527344,55433.980469,2.453400,0.238323
-1,20,7635.462891,10907.713867,0.211883
+1,20,138075.984375,197251.281250,0.288013,0.157941
-1,30,5425.212402,7750.165039,0.323606
+1,30,56904.078125,81291.343750,0.429261,0.174654
-1,40,4372.716309,6246.610840,0.296102
+1,40,10662.019531,15231.303711,0.343701,0.079876
-1,50,3846.437988,5494.793945,0.273941
+1,50,9890.013672,14128.458008,0.292848,0.088391
-2,0,17057.958984,24368.269531,0.564671
+2,0,912398.500000,1303426.125000,0.795286,0.256503
-2,10,7089.009766,10127.021484,0.315390
+2,10,12546.185547,17922.542969,1.328752,0.110444
-2,20,4230.856445,6043.994141,0.201429
+2,20,109676.710938,156680.906250,0.245523,0.132549
-2,30,3744.107910,5348.593262,0.309118
+2,30,49427.507812,70610.578125,0.327188,0.088591
-2,40,3531.041992,5044.219238,0.295744
+2,40,27778.673828,39683.660156,0.345683,0.102361
-2,50,3570.459229,5100.528320,0.297530
+2,50,10311.509766,14730.566406,0.350199,0.091674
-3,0,14367.601562,20524.917969,0.529623
+3,0,1040308.062500,1486154.000000,0.807157,0.279805
-3,10,6734.334473,9620.347656,0.304721
+3,10,64799.246094,92570.117188,0.485949,0.198213
-3,20,6140.179688,8771.602539,0.193836
+3,20,336018.000000,480025.531250,0.410466,0.118048
-3,30,4089.454102,5841.940918,0.317864
+3,30,94216.312500,134594.562500,0.355044,0.114209
-3,40,3553.830811,5076.785645,0.269531
+3,40,19988.919922,28555.457031,0.298066,0.094291
-3,50,3590.448242,5129.088867,0.287063
+3,50,9181.969727,13116.940430,0.326489,0.137392
-4,0,14410.816406,20586.648438,0.543822
+4,0,741176.187500,1058822.750000,0.850633,0.233118
-4,10,6411.443359,9159.058594,0.341742
+4,10,39252.617188,56074.410156,1.690607,0.227370
-4,20,3816.795166,5452.479492,0.198213
+4,20,108992.304688,155703.140625,0.258920,0.285944
-4,30,4069.170898,5812.959961,0.329676
+4,30,37115.253906,53021.632812,0.337674,0.131537
-4,40,3484.921631,4978.332520,0.296284
+4,40,19358.708984,27655.123047,0.349937,0.169446
-4,50,2802.801514,4003.873779,0.299286
+4,50,11434.291992,16334.540039,0.351904,0.089116
-5,0,13335.293945,19050.201172,0.509557
+5,0,845658.312500,1208083.000000,0.922483,0.285983
-5,10,5531.156738,7901.527344,0.293409
+5,10,130569.406250,186527.515625,0.405198,0.227639
-5,20,3844.260010,5491.696777,0.241263
+5,20,245780.390625,351114.687500,0.301236,0.147030
-5,30,3619.303223,5170.297363,0.317237
+5,30,42017.671875,60025.066406,0.372036,0.117895
-5,40,3492.172852,4988.697754,0.281641
+5,40,11496.740234,16423.779297,0.286911,0.085701
-5,50,3069.457275,4384.815918,0.287269
+5,50,8891.728516,12702.317383,0.322913,0.099181
-6,0,8740.982422,12486.912109,0.483061
+6,0,617909.687500,882727.812500,0.834663,0.205251
-6,10,6110.571777,8729.239258,0.347929
+6,10,18171.734375,25959.302734,0.677465,0.190570
-6,20,3350.194092,4785.889160,0.239650
+6,20,423716.187500,605308.687500,0.464189,0.156125
-6,30,3008.237549,4297.353516,0.300327
+6,30,48133.507812,68761.914062,0.478786,0.221465
-6,40,2944.483887,4206.288574,0.273457
+6,40,20350.281250,29071.666016,0.343182,0.118739
-6,50,3018.033447,4311.365234,0.259749
+6,50,11372.219727,16245.889648,0.289177,0.101218
-7,0,6784.341309,9691.704102,0.495378
+7,0,534133.500000,763047.500000,0.732465,0.239962
-7,10,4946.872559,7066.822754,0.321541
+7,10,19025.574219,27178.703125,1.551094,0.183023
-7,20,2816.704102,4023.745361,0.274515
+7,20,58042.757812,82918.078125,0.298854,0.146532
-7,30,2991.350830,4273.229980,0.299238
+7,30,11810.656250,16872.201172,0.353035,0.092791
-7,40,3023.450684,4319.083984,0.307114
+7,40,9733.656250,13905.077148,0.312043,0.089972
-7,50,2944.912598,4206.909668,0.253002
+7,50,9069.159180,12955.803711,0.280795,0.129749
-8,0,7364.851562,10521.002930,0.498283
+8,0,650366.937500,929095.250000,0.787194,0.263791
-8,10,3897.301025,5567.439453,0.311517
+8,10,13121.248047,18744.031250,1.376233,0.141053
-8,20,3313.474854,4733.448242,0.203364
+8,20,19326.507812,27609.167969,0.243424,0.183426
-8,30,2697.139648,3852.940430,0.271355
+8,30,12904.376953,18434.667969,0.334577,0.092547
-8,40,2955.225342,4221.633301,0.273175
+8,40,9682.833984,13832.478516,0.296229,0.109518
-8,50,2932.081787,4188.575684,0.262651
+8,50,8866.144531,12665.773438,0.312481,0.099564
-9,0,4065.651611,5807.854004,0.513017
+9,0,265689.562500,379556.156250,0.738840,0.186122
-9,10,4358.108398,6225.728516,0.329111
+9,10,9944.607422,14206.222656,0.806142,0.096763
-9,20,3417.019043,4881.362305,0.218488
+9,20,22091.175781,31558.679688,0.285403,0.131950
-9,30,2917.136719,4167.226562,0.260805
+9,30,19346.830078,27638.111328,0.460825,0.131496
-9,40,2786.277832,3980.287109,0.256345
+9,40,10279.702148,14685.139648,0.314872,0.095758
-9,50,2602.660645,3717.971680,0.268723
+9,50,10420.340820,14886.050781,0.314008,0.112904

example/run_all.py

@@ -75,6 +75,7 @@ def main():
             run([sys.executable, str(base_dir / "evaluate_generated.py"), "--reference", str(ref)])
         else:
             run([sys.executable, str(base_dir / "evaluate_generated.py")])
+        run([sys.executable, str(base_dir / "summary_metrics.py")])
 
 
 if __name__ == "__main__":

example/sample.py

@@ -11,6 +11,7 @@ import torch.nn.functional as F
 
 from data_utils import load_split
 from hybrid_diffusion import HybridDiffusionModel, cosine_beta_schedule
+from hybrid_diffusion import TemporalGRUGenerator
 from platform_utils import resolve_device, safe_path, ensure_dir
 
 BASE_DIR = Path(__file__).resolve().parent
@@ -56,6 +57,13 @@ def main():
     model_ff_mult = int(cfg.get("model_ff_mult", 2))
     model_pos_dim = int(cfg.get("model_pos_dim", 64))
     model_use_pos = bool(cfg.get("model_use_pos_embed", True))
+    model_use_feature_graph = bool(cfg.get("model_use_feature_graph", False))
+    feature_graph_scale = float(cfg.get("feature_graph_scale", 0.1))
+    feature_graph_dropout = float(cfg.get("feature_graph_dropout", 0.0))
+    use_temporal_stage1 = bool(cfg.get("use_temporal_stage1", False))
+    temporal_hidden_dim = int(cfg.get("temporal_hidden_dim", 256))
+    temporal_num_layers = int(cfg.get("temporal_num_layers", 1))
+    temporal_dropout = float(cfg.get("temporal_dropout", 0.0))
 
     split = load_split(str(SPLIT_PATH))
     time_col = split.get("time_column", "time")
@@ -83,6 +91,9 @@ def main():
         ff_mult=model_ff_mult,
         pos_dim=model_pos_dim,
         use_pos_embed=model_use_pos,
+        use_feature_graph=model_use_feature_graph,
+        feature_graph_scale=feature_graph_scale,
+        feature_graph_dropout=feature_graph_dropout,
         cond_vocab_size=cond_vocab_size,
         cond_dim=cond_dim,
         use_tanh_eps=use_tanh_eps,
@@ -92,6 +103,20 @@ def main():
         model.load_state_dict(torch.load(str(MODEL_PATH), map_location=DEVICE, weights_only=True))
     model.eval()
 
+    temporal_model = None
+    if use_temporal_stage1:
+        temporal_model = TemporalGRUGenerator(
+            input_dim=len(cont_cols),
+            hidden_dim=temporal_hidden_dim,
+            num_layers=temporal_num_layers,
+            dropout=temporal_dropout,
+        ).to(DEVICE)
+        temporal_path = BASE_DIR / "results" / "temporal.pt"
+        if not temporal_path.exists():
+            raise SystemExit(f"missing temporal model file: {temporal_path}")
+        temporal_model.load_state_dict(torch.load(str(temporal_path), map_location=DEVICE, weights_only=True))
+        temporal_model.eval()
+
     betas = cosine_beta_schedule(timesteps).to(DEVICE)
     alphas = 1.0 - betas
     alphas_cumprod = torch.cumprod(alphas, dim=0)
@@ -110,19 +135,26 @@ def main():
             raise SystemExit("use_condition enabled but no files matched data_glob")
         cond = torch.randint(0, cond_vocab_size, (batch_size,), device=DEVICE, dtype=torch.long)
 
+    trend = None
+    if temporal_model is not None:
+        trend = temporal_model.generate(batch_size, seq_len, DEVICE)
+
     for t in reversed(range(timesteps)):
         t_batch = torch.full((batch_size,), t, device=DEVICE, dtype=torch.long)
         eps_pred, logits = model(x_cont, x_disc, t_batch, cond)
 
+        # Continuous reverse step (DDPM): x_{t-1} mean
+        a_t = alphas[t]
+        a_bar_t = alphas_cumprod[t]
         if cont_target == "x0":
             x0_pred = eps_pred
             if cont_clamp_x0 > 0:
                 x0_pred = torch.clamp(x0_pred, -cont_clamp_x0, cont_clamp_x0)
             eps_pred = (x_cont - torch.sqrt(a_bar_t) * x0_pred) / torch.sqrt(1.0 - a_bar_t)
-
+        elif cont_target == "v":
-        # Continuous reverse step (DDPM): x_{t-1} mean
+            v_pred = eps_pred
-        a_t = alphas[t]
+            x0_pred = torch.sqrt(a_bar_t) * x_cont - torch.sqrt(1.0 - a_bar_t) * v_pred
-        a_bar_t = alphas_cumprod[t]
+            eps_pred = torch.sqrt(1.0 - a_bar_t) * x_cont + torch.sqrt(a_bar_t) * v_pred
         coef1 = 1.0 / torch.sqrt(a_t)
         coef2 = (1 - a_t) / torch.sqrt(1 - a_bar_t)
         mean = coef1 * (x_cont - coef2 * eps_pred)
@@ -146,6 +178,8 @@ def main():
                     sampled = torch.multinomial(probs.view(-1, probs.size(-1)), 1).view(BATCH_SIZE, SEQ_LEN)
                     x_disc[:, :, i][mask] = sampled[mask]
 
+    if trend is not None:
+        x_cont = x_cont + trend
     print("sampled_cont_shape", tuple(x_cont.shape))
     print("sampled_disc_shape", tuple(x_disc.shape))
 

example/summary_metrics.py

@@ -0,0 +1,40 @@
+#!/usr/bin/env python3
+"""Print average metrics from eval.json for quick tracking."""
+
+import json
+from datetime import datetime
+from pathlib import Path
+
+
+def mean(values):
+    return sum(values) / len(values) if values else None
+
+
+def main():
+    base_dir = Path(__file__).resolve().parent
+    eval_path = base_dir / "results" / "eval.json"
+    if not eval_path.exists():
+        raise SystemExit(f"missing eval.json: {eval_path}")
+
+    obj = json.loads(eval_path.read_text(encoding="utf-8"))
+    ks = list(obj.get("continuous_ks", {}).values())
+    jsd = list(obj.get("discrete_jsd", {}).values())
+    lag = list(obj.get("continuous_lag1_diff", {}).values())
+
+    avg_ks = mean(ks)
+    avg_jsd = mean(jsd)
+    avg_lag1 = mean(lag)
+
+    print("avg_ks", avg_ks)
+    print("avg_jsd", avg_jsd)
+    print("avg_lag1_diff", avg_lag1)
+
+    history_path = base_dir / "results" / "metrics_history.csv"
+    if not history_path.exists():
+        history_path.write_text("timestamp,avg_ks,avg_jsd,avg_lag1_diff\n", encoding="utf-8")
+    with history_path.open("a", encoding="utf-8") as f:
+        f.write(f"{datetime.utcnow().isoformat()},{avg_ks},{avg_jsd},{avg_lag1}\n")
+
+
+if __name__ == "__main__":
+    main()

example/train.py

@@ -14,6 +14,7 @@ import torch.nn.functional as F
 from data_utils import load_split, windowed_batches
 from hybrid_diffusion import (
     HybridDiffusionModel,
+    TemporalGRUGenerator,
     cosine_beta_schedule,
     q_sample_continuous,
     q_sample_discrete,
@@ -58,12 +59,25 @@ DEFAULTS = {
     "model_ff_mult": 2,
     "model_pos_dim": 64,
     "model_use_pos_embed": True,
+    "model_use_feature_graph": True,
+    "feature_graph_scale": 0.1,
+    "feature_graph_dropout": 0.0,
+    "use_temporal_stage1": True,
+    "temporal_hidden_dim": 256,
+    "temporal_num_layers": 1,
+    "temporal_dropout": 0.0,
+    "temporal_loss_weight": 1.0,
     "disc_mask_scale": 0.9,
     "shuffle_buffer": 256,
     "cont_loss_weighting": "none",  # none | inv_std
     "cont_loss_eps": 1e-6,
-    "cont_target": "eps",  # eps | x0
+    "cont_target": "eps",  # eps | x0 | v
     "cont_clamp_x0": 0.0,
+    "quantile_loss_weight": 0.0,
+    "quantile_points": [0.05, 0.25, 0.5, 0.75, 0.95],
+    "quantile_loss_warmup_steps": 200,
+    "quantile_loss_clip": 6.0,
+    "quantile_loss_huber_delta": 1.0,
 }
 
 
@@ -188,12 +202,27 @@ def main():
         ff_mult=int(config.get("model_ff_mult", 2)),
         pos_dim=int(config.get("model_pos_dim", 64)),
         use_pos_embed=bool(config.get("model_use_pos_embed", True)),
+        use_feature_graph=bool(config.get("model_use_feature_graph", False)),
+        feature_graph_scale=float(config.get("feature_graph_scale", 0.1)),
+        feature_graph_dropout=float(config.get("feature_graph_dropout", 0.0)),
         cond_vocab_size=cond_vocab_size,
         cond_dim=int(config.get("cond_dim", 32)),
         use_tanh_eps=bool(config.get("use_tanh_eps", False)),
         eps_scale=float(config.get("eps_scale", 1.0)),
     ).to(device)
+    temporal_model = None
+    if bool(config.get("use_temporal_stage1", False)):
+        temporal_model = TemporalGRUGenerator(
+            input_dim=len(cont_cols),
+            hidden_dim=int(config.get("temporal_hidden_dim", 256)),
+            num_layers=int(config.get("temporal_num_layers", 1)),
+            dropout=float(config.get("temporal_dropout", 0.0)),
+        ).to(device)
     opt = torch.optim.Adam(model.parameters(), lr=float(config["lr"]))
+    if temporal_model is not None:
+        opt_temporal = torch.optim.Adam(temporal_model.parameters(), lr=float(config["lr"]))
+    else:
+        opt_temporal = None
     ema = EMA(model, float(config["ema_decay"])) if config.get("use_ema") else None
 
     betas = cosine_beta_schedule(int(config["timesteps"])).to(device)
@@ -203,7 +232,11 @@ def main():
     os.makedirs(config["out_dir"], exist_ok=True)
     out_dir = safe_path(config["out_dir"])
     log_path = os.path.join(out_dir, "train_log.csv")
+    use_quantile = float(config.get("quantile_loss_weight", 0.0)) > 0
     with open(log_path, "w", encoding="utf-8") as f:
+        if use_quantile:
+            f.write("epoch,step,loss,loss_cont,loss_disc,loss_quantile\n")
+        else:
             f.write("epoch,step,loss,loss_cont,loss_disc\n")
     with open(os.path.join(out_dir, "config_used.json"), "w", encoding="utf-8") as f:
         json.dump(config, f, indent=2)
@@ -235,10 +268,20 @@ def main():
             x_cont = x_cont.to(device)
             x_disc = x_disc.to(device)
 
+            temporal_loss = None
+            x_cont_resid = x_cont
+            trend = None
+            if temporal_model is not None:
+                trend, pred_next = temporal_model.forward_teacher(x_cont)
+                temporal_loss = F.mse_loss(pred_next, x_cont[:, 1:, :])
+                temporal_loss = temporal_loss * float(config.get("temporal_loss_weight", 1.0))
+                trend = trend.detach()
+                x_cont_resid = x_cont - trend
+
             bsz = x_cont.size(0)
             t = torch.randint(0, int(config["timesteps"]), (bsz,), device=device)
 
-            x_cont_t, noise = q_sample_continuous(x_cont, t, alphas_cumprod)
+            x_cont_t, noise = q_sample_continuous(x_cont_resid, t, alphas_cumprod)
 
             mask_tokens = torch.tensor(vocab_sizes, device=device)
             x_disc_t, mask = q_sample_discrete(
@@ -253,10 +296,14 @@ def main():
 
             cont_target = str(config.get("cont_target", "eps"))
             if cont_target == "x0":
-                x0_target = x_cont
+                x0_target = x_cont_resid
                 if float(config.get("cont_clamp_x0", 0.0)) > 0:
                     x0_target = torch.clamp(x0_target, -float(config["cont_clamp_x0"]), float(config["cont_clamp_x0"]))
                 loss_base = (eps_pred - x0_target) ** 2
+            elif cont_target == "v":
+                a_bar_t = alphas_cumprod[t].view(-1, 1, 1)
+                v_target = torch.sqrt(a_bar_t) * noise - torch.sqrt(1.0 - a_bar_t) * x_cont_resid
+                loss_base = (eps_pred - v_target) ** 2
             else:
                 loss_base = (eps_pred - noise) ** 2
 
@@ -282,17 +329,72 @@ def main():
 
             lam = float(config["lambda"])
             loss = lam * loss_cont + (1 - lam) * loss_disc
+
+            q_weight = float(config.get("quantile_loss_weight", 0.0))
+            quantile_loss = 0.0
+            if q_weight > 0:
+                warmup = int(config.get("quantile_loss_warmup_steps", 0))
+                if warmup > 0:
+                    q_weight = q_weight * min(1.0, (total_step + 1) / float(warmup))
+                q_points = config.get("quantile_points", [0.05, 0.25, 0.5, 0.75, 0.95])
+                q_tensor = torch.tensor(q_points, device=device, dtype=x_cont.dtype)
+                # Use normalized space for stable quantiles on x0.
+                x_real = x_cont_resid
+                a_bar_t = alphas_cumprod[t].view(-1, 1, 1)
+                if cont_target == "x0":
+                    x_gen = eps_pred
+                elif cont_target == "v":
+                    v_pred = eps_pred
+                    x_gen = torch.sqrt(a_bar_t) * x_cont_t - torch.sqrt(1.0 - a_bar_t) * v_pred
+                else:
+                    # eps prediction
+                    x_gen = (x_cont_t - torch.sqrt(1.0 - a_bar_t) * eps_pred) / torch.sqrt(a_bar_t)
+                q_clip = float(config.get("quantile_loss_clip", 0.0))
+                if q_clip > 0:
+                    x_real = torch.clamp(x_real, -q_clip, q_clip)
+                    x_gen = torch.clamp(x_gen, -q_clip, q_clip)
+                x_real = x_real.view(-1, x_real.size(-1))
+                x_gen = x_gen.view(-1, x_gen.size(-1))
+                q_real = torch.quantile(x_real, q_tensor, dim=0)
+                q_gen = torch.quantile(x_gen, q_tensor, dim=0)
+                q_delta = float(config.get("quantile_loss_huber_delta", 0.0))
+                q_diff = q_gen - q_real
+                if q_delta > 0:
+                    quantile_loss = torch.nn.functional.smooth_l1_loss(q_gen, q_real, beta=q_delta)
+                else:
+                    quantile_loss = torch.mean(torch.abs(q_diff))
+                loss = loss + q_weight * quantile_loss
+
             opt.zero_grad()
             loss.backward()
             if float(config.get("grad_clip", 0.0)) > 0:
                 torch.nn.utils.clip_grad_norm_(model.parameters(), float(config["grad_clip"]))
             opt.step()
+            if opt_temporal is not None:
+                opt_temporal.zero_grad()
+                temporal_loss.backward()
+                if float(config.get("grad_clip", 0.0)) > 0:
+                    torch.nn.utils.clip_grad_norm_(temporal_model.parameters(), float(config["grad_clip"]))
+                opt_temporal.step()
             if ema is not None:
                 ema.update(model)
 
             if step % int(config["log_every"]) == 0:
                 print("epoch", epoch, "step", step, "loss", float(loss))
                 with open(log_path, "a", encoding="utf-8") as f:
+                    if use_quantile:
+                        f.write(
+                            "%d,%d,%.6f,%.6f,%.6f,%.6f\n"
+                            % (
+                                epoch,
+                                step,
+                                float(loss),
+                                float(loss_cont),
+                                float(loss_disc),
+                                float(quantile_loss),
+                            )
+                        )
+                    else:
                         f.write(
                             "%d,%d,%.6f,%.6f,%.6f\n"
                             % (epoch, step, float(loss), float(loss_cont), float(loss_disc))
@@ -308,11 +410,15 @@ def main():
                 }
                 if ema is not None:
                     ckpt["ema"] = ema.state_dict()
+                if temporal_model is not None:
+                    ckpt["temporal"] = temporal_model.state_dict()
                 torch.save(ckpt, os.path.join(out_dir, "model_ckpt.pt"))
 
     torch.save(model.state_dict(), os.path.join(out_dir, "model.pt"))
     if ema is not None:
         torch.save(ema.state_dict(), os.path.join(out_dir, "model_ema.pt"))
+    if temporal_model is not None:
+        torch.save(temporal_model.state_dict(), os.path.join(out_dir, "temporal.pt"))
 
 
 if __name__ == "__main__":

report.md

@@ -0,0 +1,365 @@
+# Hybrid Diffusion for ICS Traffic (HAI 21.03) — Project Report
+# 工业控制系统流量混合扩散生成（HAI 21.03）— 项目报告
+
+## 1. Project Goal / 项目目标
+Build a **hybrid diffusion-based generator** for industrial control system (ICS) traffic features, targeting **mixed continuous + discrete** feature sequences. The output is **feature-level sequences**, not raw packets. The generator should preserve:
+- **Distributional fidelity** (continuous value ranges and discrete frequencies)
+- **Temporal consistency** (time correlation and sequence structure)
+- **Protocol/field consistency** (for discrete fields)
+
+构建一个用于工业控制系统（ICS）流量特征的**混合扩散生成模型**，面向**连续+离散混合特征序列**。输出为**特征级序列**而非原始报文。生成结果需要同时保持：
+- **分布一致性**（连续值范围与离散取值频率）
+- **时序一致性**（时间相关性与序列结构）
+- **字段/协议一致性**（离散字段的逻辑一致）
+
+This project is aligned with the STOUTER idea of **structure-aware diffusion** for spatiotemporal data, but applied to **ICS feature sequences** rather than cellular traffic.
+
+本项目呼应 STOUTER 的**结构先验+扩散**思想，但应用于**ICS 特征序列**而非蜂窝流量。
+
+---
+
+## 2. Data and Scope / 数据与范围
+**Dataset used in the current implementation:** HAI 21.03 (CSV feature traces)
+
+**当前实现使用的数据集：** HAI 21.03（CSV 特征序列）
+
+**Data location (default in config):**
+- `dataset/hai/hai-21.03/train*.csv.gz`
+
+**数据位置（config 默认）：**
+- `dataset/hai/hai-21.03/train*.csv.gz`
+
+**Feature split (fixed schema):**
+- Defined in `example/feature_split.json`
+- **Continuous features:** sensor/process values
+- **Discrete features:** binary/low-cardinality status/flag fields
+- `time` column is excluded from modeling
+
+**特征拆分（固定 schema）：**
+- `example/feature_split.json`
+- **连续特征：** 传感器/过程值
+- **离散特征：** 二值/低基数状态字段
+- `time` 列不参与训练
+
+---
+
+## 3. End-to-End Pipeline / 端到端流程
+
+**One command pipeline:**
+```
+python example/run_all.py --device cuda
+```
+
+**一键流程：**
+```
+python example/run_all.py --device cuda
+```
+
+### Pipeline stages / 流程阶段
+1) **Prepare data** (`example/prepare_data.py`)
+2) **Train model** (`example/train.py`)
+3) **Generate samples** (`example/export_samples.py`)
+4) **Evaluate** (`example/evaluate_generated.py`)
+5) **Summarize metrics** (`example/summary_metrics.py`)
+
+1) **数据准备**（统计量与词表）
+2) **训练模型**
+3) **生成样本并导出**
+4) **评估指标**
+5) **汇总指标**
+
+---
+
+## 4. Technical Architecture / 技术架构
+
+### 4.1 Hybrid Diffusion Model (Core) / 混合扩散模型（核心）
+Defined in `example/hybrid_diffusion.py`.
+
+**Key components:**
+- **Continuous branch**: Gaussian diffusion (DDPM style)
+- **Discrete branch**: Mask diffusion for categorical tokens
+- **Shared backbone**: GRU + residual MLP + LayerNorm
+- **Embedding inputs**:
+  - continuous projection
+  - discrete embeddings per column
+  - time embedding (sinusoidal)
+  - positional embedding (sequence index)
+  - optional condition embedding (`file_id`)
+
+**Outputs:**
+- Continuous head: predicts target (`eps`, `x0`, or `v`)
+- Discrete heads: predict logits for each discrete column
+
+**核心组成：**
+- **连续分支：** 高斯扩散（DDPM）
+- **离散分支：** Mask 扩散
+- **共享主干：** GRU + 残差 MLP + LayerNorm
+- **输入嵌入：**
+  - 连续投影
+  - 离散字段嵌入
+  - 时间嵌入（正弦）
+  - 位置嵌入（序列索引）
+  - 条件嵌入（可选，`file_id`）
+
+**输出：**
+- 连续 head：预测 `eps/x0/v`
+- 离散 head：各字段 logits
+
+---
+
+### 4.2 Feature Graph Mixer (Structure Prior) / 特征图混合器（结构先验）
+Implemented in `example/hybrid_diffusion.py` as `FeatureGraphMixer`.
+
+Purpose: inject **learnable feature-dependency prior** without dataset-specific hardcoding.
+
+**Mechanism:**
+- Learns a dense feature relation matrix `A`
+- Applies: `x + x @ A`
+- Symmetric stabilizing constraint: `(A + A^T)/2`
+- Controlled by scale and dropout
+
+**Config:**
+```
+"model_use_feature_graph": true,
+"feature_graph_scale": 0.1,
+"feature_graph_dropout": 0.0
+```
+
+**目的：**在不写死特定数据集关系的情况下，引入**可学习特征依赖先验**。
+
+**机制：**
+- 学习稠密关系矩阵 `A`
+- 特征混合：`x + x @ A`
+- 对称化稳定：`(A + A^T)/2`
+- 通过 scale/dropout 控制强度
+
+---
+
+### 4.3 Two-Stage Temporal Backbone / 两阶段时序骨干
+Stage-1 uses a **GRU temporal generator** to model sequence trend in normalized space. Stage-2 diffusion then models the **residual** (x − trend). This decouples temporal consistency from distribution alignment.
+
+第一阶段使用 **GRU 时序生成器**在归一化空间建模序列趋势；第二阶段扩散模型学习**残差**（x − trend），实现时序一致性与分布对齐的解耦。
+
+---
+
+## 5. Diffusion Formulations / 扩散建模形式
+
+### 5.1 Continuous Diffusion / 连续扩散
+Forward process:
+```
+x_t = sqrt(a_bar_t) * x_0 + sqrt(1 - a_bar_t) * eps
+```
+
+Targets supported:
+- **eps prediction** (standard DDPM)
+- **x0 prediction** (direct reconstruction)
+- **v prediction** (v = sqrt(a_bar)*eps − sqrt(1-a_bar)*x0)
+
+Current config default:
+```
+"cont_target": "v"
+```
+
+Sampling uses the target to reconstruct `eps` and apply standard DDPM reverse update.
+
+**前向扩散：**如上公式。
+
+**支持的目标：**
+- `eps`（噪声预测）
+- `x0`（原样本预测）
+- `v`（v‑prediction）
+
+**当前默认：**`cont_target = v`
+
+**采样：**根据目标反解 `eps` 再执行标准 DDPM 反向步骤。
+
+---
+
+### 5.2 Discrete Diffusion (Mask) / 离散扩散（Mask）
+Forward process: replace tokens with `[MASK]` using cosine schedule:
+```
+p(t) = 0.5 * (1 - cos(pi * t / T))
+```
+Optional scale: `disc_mask_scale`
+
+Reverse process: cross-entropy on masked positions only.
+
+**前向：**按 cosine schedule 进行 Mask。
+**反向：**仅在 mask 位置计算交叉熵。
+
+---
+
+## 6. Loss Design (Current) / 当前损失设计
+Total loss:
+```
+L = λ * L_cont + (1 − λ) * L_disc + w_q * L_quantile
+```
+
+### 6.1 Continuous Loss / 连续损失
+Depending on `cont_target`:
+- eps target: MSE(eps_pred, eps)
+- x0 target: MSE(x0_pred, x0)
+- v target: MSE(v_pred, v_target)
+
+Optional inverse-variance weighting:
+```
+cont_loss_weighting = "inv_std"
+```
+
+### 6.2 Discrete Loss / 离散损失
+Cross-entropy on masked positions only.
+
+### 6.3 Quantile Loss (Distribution Alignment) / 分位数损失（分布对齐）
+Added to improve KS (distribution shape alignment):
+- Compute quantiles on generated vs real x0
+- Loss = Huber or L1 difference on quantiles
+
+Stabilization:
+```
+quantile_loss_warmup_steps
+quantile_loss_clip
+quantile_loss_huber_delta
+```
+
+---
+
+## 7. Training Strategy / 训练策略
+Defined in `example/train.py`.
+
+**Key techniques:**
+- EMA of model weights
+- Gradient clipping
+- Shuffle buffer to reduce batch bias
+- Optional feature graph prior
+- Quantile loss warmup for stability
+- Optional stage-1 temporal GRU (trend) + residual diffusion
+
+**Config highlights (example/config.json):**
+```
+timesteps: 600
+batch_size: 128
+seq_len: 128
+epochs: 10
+max_batches: 4000
+lambda: 0.7
+cont_target: "v"
+quantile_loss_weight: 0.1
+model_use_feature_graph: true
+use_temporal_stage1: true
+```
+
+---
+
+## 8. Sampling & Export / 采样与导出
+Defined in:
+- `example/sample.py`
+- `example/export_samples.py`
+
+**Export steps:**
+- Reverse diffusion with conditional sampling
+- Reverse normalize continuous values
+- Clamp to observed min/max
+- Restore discrete tokens from vocab
+- Write to CSV
+
+---
+
+## 9. Evaluation Metrics / 评估指标
+Implemented in `example/evaluate_generated.py`.
+
+### Continuous Metrics / 连续指标
+- **KS statistic** (distribution similarity per feature)
+- **Quantile errors** (q05/q25/q50/q75/q95)
+- **Lag‑1 correlation diff** (temporal structure)
+
+### Discrete Metrics / 离散指标
+- **JSD** over token frequency distribution
+- **Invalid token counts**
+
+### Summary Metrics / 汇总指标
+Auto-logged in:
+- `example/results/metrics_history.csv`
+- via `example/summary_metrics.py`
+
+---
+
+## 10. Automation / 自动化
+
+### One‑click pipeline / 一键流程
+```
+python example/run_all.py --device cuda
+```
+
+### Metrics logging / 指标记录
+Each run appends:
+```
+timestamp,avg_ks,avg_jsd,avg_lag1_diff
+```
+
+---
+
+## 11. Key Engineering Decisions / 关键工程决策
+
+### 11.1 Mixed-Type Diffusion / 混合类型扩散
+Continuous + discrete handled separately to respect data types.
+
+### 11.2 Structure Prior / 结构先验
+Learnable feature graph added to encode implicit dependencies.
+
+### 11.3 v‑prediction
+Chosen to stabilize training and improve convergence in diffusion.
+
+### 11.4 Distribution Alignment / 分布对齐
+Quantile loss introduced to directly reduce KS.
+
+---
+
+## 12. Known Issues / Current Limitations / 已知问题与当前局限
+- **KS remains high** in many experiments, meaning continuous distributions are still misaligned.
+- **Lag‑1 may degrade** when quantile loss is too strong.
+- **Loss spikes** observed when quantile loss is unstable (mitigated with warmup + clip + Huber).
+
+**当前问题：**
+- KS 高，说明连续分布仍未对齐
+- 分位数损失过强时会损害时序相关性
+- 分位数损失不稳定时会出现 loss 爆炸（已引入 warmup/clip/Huber）
+
+---
+
+## 13. Suggested Next Steps (Research Roadmap) / 下一步建议（研究路线）
+1) **SNR-weighted loss** (improve stability across timesteps)
+2) **Two-stage training** (distribution first, temporal consistency second)
+3) **Upgrade discrete diffusion** (D3PM-style transitions)
+4) **Structured conditioning** (state/phase conditioning)
+5) **Graph-based priors** (explicit feature/plant dependency graphs)
+
+---
+
+## 14. Code Map (Key Files) / 代码索引（关键文件）
+
+**Core model**
+- `example/hybrid_diffusion.py`
+
+**Training**
+- `example/train.py`
+
+**Sampling & export**
+- `example/sample.py`
+- `example/export_samples.py`
+
+**Pipeline**
+- `example/run_all.py`
+
+**Evaluation**
+- `example/evaluate_generated.py`
+- `example/summary_metrics.py`
+
+**Configs**
+- `example/config.json`
+
+---
+
+## 15. Summary / 总结
+This project implements a **hybrid diffusion model for ICS traffic features**, combining continuous Gaussian diffusion with discrete mask diffusion, enhanced with a **learnable feature-graph prior**. The system includes a full pipeline for preparation, training, sampling, exporting, and evaluation. Key research challenges remain in **distribution alignment (KS)** and **joint optimization of distribution fidelity vs temporal consistency**, motivating future improvements such as SNR-weighted loss, staged training, and stronger structural priors.
+
+本项目实现了用于 ICS 流量特征的**混合扩散模型**，将连续高斯扩散与离散 Mask 扩散结合，并引入**可学习特征图先验**。系统包含完整的数据准备、训练、采样、导出与评估流程。当前研究挑战集中在**连续分布对齐（KS）**与**分布/时序一致性之间的权衡**，后续可通过 SNR‑weighted loss、分阶段训练与更强结构先验继续改进。