From 47d00106b1cfd85f56a81ceebe5fd28d0bb518d8 Mon Sep 17 00:00:00 2001 From: 1294271022 <1294271022@qq.com> Date: Thu, 28 May 2026 19:57:38 +0800 Subject: [PATCH] =?UTF-8?q?1=E3=80=81c++=E8=87=AA=E5=8A=A8=E6=8B=9F?= =?UTF-8?q?=E5=90=88=E8=8C=83=E5=9B=B4=E9=80=82=E9=85=8D?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- .../src/models/time_conditioned_surrogate.py | 44 +++++++++++++------ .../nmCalculation/nmCalculationAutoFitPSO.cpp | 6 +-- 2 files changed, 33 insertions(+), 17 deletions(-) diff --git a/ML/nmWTAI-ML/src/models/time_conditioned_surrogate.py b/ML/nmWTAI-ML/src/models/time_conditioned_surrogate.py index 7e93160..cde54e4 100644 --- a/ML/nmWTAI-ML/src/models/time_conditioned_surrogate.py +++ b/ML/nmWTAI-ML/src/models/time_conditioned_surrogate.py @@ -1,12 +1,12 @@ # -*- coding: utf-8 -*- -"""Time-conditioned surrogate network. +"""时间条件代理模型网络结构。 -The model predicts one point on the log-pressure and log-derivative curves: +TimeConditionedSurrogate 不一次性输出完整曲线，而是把“物理参数 + 流量制度 + 某个 +时间点特征”作为输入，预测该时间点的 log_pressure 和 log_derivative。它适合用于 +更灵活的时间采样、局部曲线重建以及需要按时间点加权的训练策略。 - f(params, schedule, time_point) -> [log_pressure(t), log_derivative(t)] - -It is intended for flexible time grids, point-wise training, and later PSO -candidate screening where only selected curves need full solver evaluation. +模型主体由参数编码器、制度编码器、时间编码器、融合投影和若干残差块组成。残差 +块让较深的全连接网络更容易优化，同时保留原始融合特征。 """ from __future__ import annotations @@ -21,7 +21,7 @@ from torch import nn @dataclass(frozen=True) class TimeConditionedSurrogateConfig: - """Network dimensions and architecture options.""" + """时间条件代理模型配置。""" param_dim: int schedule_dim: int @@ -33,9 +33,10 @@ class TimeConditionedSurrogateConfig: class ResidualBlock(nn.Module): - """Fully connected residual block used in the fusion trunk.""" + """时间条件模型使用的全连接残差块，用于在较深网络中稳定传播特征。""" def __init__(self, dim: int, dropout: float = 0.0): + """构造两层全连接残差块，并根据 dropout 参数决定是否启用随机失活。""" super().__init__() self.net = nn.Sequential( nn.LayerNorm(dim), @@ -47,12 +48,18 @@ class ResidualBlock(nn.Module): self.act = nn.GELU() def forward(self, x: torch.Tensor) -> torch.Tensor: - """Apply a residual correction and activation.""" + """在保留原始隐藏表示的基础上叠加两层非线性修正。""" + # 残差连接让块学习修正量，而不是每层都重新表示完整特征。 return self.act(x + self.net(x)) class TimeConditionedSurrogate(nn.Module): - """Point-wise time-conditioned pressure/derivative predictor.""" + """逐时间点预测的时间条件代理模型。 + + 与 ForwardSurrogate 一次输出整条曲线不同，该模型每次只预测一个时间点的 + log_pressure 和 log_derivative。训练数据通常由 PointCurveDataset 将 [N, T] 曲线 + 展开为 N*T 个样本，因此 batch 中每一行都带有自己的 time_x。 + """ def __init__( self, @@ -66,6 +73,7 @@ class TimeConditionedSurrogate(nn.Module): dropout: float = 0.05, use_schedule: bool = True, ): + """按配置组装时间条件代理模型的编码、融合和输出层。""" super().__init__() config = self._coerce_config( config=config, @@ -121,7 +129,7 @@ class TimeConditionedSurrogate(nn.Module): dropout: float, use_schedule: bool, ) -> TimeConditionedSurrogateConfig: - """Support both config-object and legacy keyword constructors.""" + """兼容配置对象式构造和旧版关键字参数式构造。""" if config is not None: return config if param_dim is None or schedule_dim is None or time_dim is None: @@ -141,6 +149,7 @@ class TimeConditionedSurrogate(nn.Module): @staticmethod def _build_encoder(in_dim: int, out_dim: int) -> nn.Sequential: + """构建 Linear-LayerNorm-GELU 编码器。""" return nn.Sequential( nn.Linear(in_dim, out_dim), nn.LayerNorm(out_dim), @@ -153,18 +162,25 @@ class TimeConditionedSurrogate(nn.Module): time_x: torch.Tensor, schedule_x: torch.Tensor | None = None, ) -> torch.Tensor: - """Fuse sample-level and point-level features into a two-channel output.""" + """融合样本级特征和点级时间特征，输出双通道响应。 + + params_x 与 schedule_x 在同一条曲线的所有时间点上相同，time_x 则随时间点变化。 + 三类特征编码后拼接进入残差主干，最后输出 [B, 2]，分别对应标准化空间中的 + log_pressure 和 log_derivative。 + """ + # params_x 和 schedule_x 是样本级特征；time_x 是展开后的点级特征。 param_feat = self.param_encoder(params_x) time_feat = self.time_encoder(time_x) if self.use_schedule: if schedule_x is None: - raise ValueError("use_schedule=True, but schedule_x was not provided") + raise ValueError("use_schedule=True，但 forward 没有传入 schedule_x") schedule_feat = self.schedule_encoder(schedule_x) fused = torch.cat([param_feat, schedule_feat, time_feat], dim=-1) else: fused = torch.cat([param_feat, time_feat], dim=-1) + # 融合后的特征经过残差主干，输出 log_pressure 和 log_derivative 两个通道。 hidden = self.input_proj(fused) hidden = self.blocks(hidden) return self.head(hidden) @@ -173,5 +189,5 @@ class TimeConditionedSurrogate(nn.Module): def build_time_conditioned_surrogate( config: TimeConditionedSurrogateConfig, ) -> TimeConditionedSurrogate: - """Create a time-conditioned surrogate from a config object.""" + """根据配置创建时间条件代理模型。""" return TimeConditionedSurrogate(config) diff --git a/Src/nmNum/nmCalculation/nmCalculationAutoFitPSO.cpp b/Src/nmNum/nmCalculation/nmCalculationAutoFitPSO.cpp index 610a466..23cadba 100644 --- a/Src/nmNum/nmCalculation/nmCalculationAutoFitPSO.cpp +++ b/Src/nmNum/nmCalculation/nmCalculationAutoFitPSO.cpp @@ -47,8 +47,8 @@ static const int kSurrogateScoringMaxAttempts = 2; static const bool kUseFixedPsoSeed = true; static const unsigned int kFixedPsoSeed = 1792008679u; -static const double kSurrogateKMin = 1.0e-4; -static const double kSurrogateKMax = 100.0; +static const double kSurrogateKMin = 1.0e-3; +static const double kSurrogateKMax = 10.0; static const double kSurrogateSkinMin = -10.0; static const double kSurrogateSkinMax = 10.0; static const double kSurrogateWellboreCMin = 1.0e-4; @@ -56,7 +56,7 @@ static const double kSurrogateWellboreCMax = 2.0; static const double kSurrogatePhiMin = 1.0e-2; static const double kSurrogatePhiMax = 0.50; static const double kSurrogateHMin = 2.0; -static const double kSurrogateHMax = 100.0; +static const double kSurrogateHMax = 50.0; static const double kSurrogateCfFixed = 4.315e-4; static const double kSurrogateCfRelTol = 0.25; static const int kSurrogateMinProdSections = 3;