上传单个文件: file_name.extension

CylinderKnMa/process_full_dataset.py

@@ -0,0 +1,273 @@
+import os
+import glob
+import re
+import numpy as np
+import torch
+from scipy.interpolate import griddata
+from matplotlib.path import Path
+from tqdm import tqdm
+import gc # 垃圾回收模块
+
+# ================= ⚙️ 核心配置区域 (Configuration) =================
+
+# 1. 路径配置
+BASE_DIR = os.path.dirname(os.path.abspath(__file__))
+# 原始数据根目录 (现在指向 data/raw，脚本会自动搜索内部的 Kn 子文件夹)
+RAW_DATA_ROOT = os.path.join(BASE_DIR, "", "") 
+OUTPUT_DIR = os.path.join(BASE_DIR, "data", "processed")
+os.makedirs(OUTPUT_DIR, exist_ok=True)
+
+# 2. 统一网格设置 (保持不变)
+GRID_H, GRID_W = 128, 192 
+X_RANGE = (-3.0, 5.5)
+Y_RANGE = (-3.0, 3.0)
+
+# 3. 物理量提取设置
+COL_INDICES = [5, 6, 7, 10] # rho, u, v, T
+VAR_NAMES = ['rho', 'u', 'v', 'T']
+NUM_CHANNELS = 4
+
+# 4. 全局参数范围 (用于归一化)
+KN_MIN, KN_MAX = 0.06, 1.0
+MA_MIN, MA_MAX = 2.0, 6.9
+
+# ================= 🛠️ 工具函数 (Utils) =================
+
+def parse_folder_params(folder_name):
+    """从文件夹名提取 Kn 和 Ma"""
+    kn_match = re.search(r"Kn(\d+)p(\d+)", folder_name)
+    ma_match = re.search(r"Ma(\d+)p(\d+)", folder_name)
+    
+    if kn_match and ma_match:
+        kn_val = float(f"{kn_match.group(1)}.{kn_match.group(2)}")
+        ma_val = float(f"{ma_match.group(1)}.{ma_match.group(2)}")
+        return kn_val, ma_val
+    return None, None
+
+def parse_surf_file(filepath):
+    """读取几何轮廓"""
+    points = []
+    start_reading = False
+    try:
+        with open(filepath, 'r') as f:
+            lines = f.readlines()
+        for line in lines:
+            clean_line = line.strip()
+            if not clean_line: continue
+            if not start_reading:
+                if clean_line.lower() == "points": start_reading = True
+                continue
+            parts = clean_line.split()
+            if len(parts) == 3:
+                try:
+                    points.append([float(parts[1]), float(parts[2])])
+                except ValueError: break
+            else: break
+        return np.array(points, dtype=np.float32)
+    except Exception as e:
+        print(f"Error reading surf {filepath}: {e}")
+        return None
+
+def parse_grid_file(filepath):
+    """读取流场数据"""
+    try:
+        data = np.loadtxt(filepath, skiprows=9)
+        coords = data[:, 1:3]         
+        values = data[:, COL_INDICES] 
+        return coords, values
+    except Exception as e:
+        print(f"Error reading grid {filepath}: {e}")
+        return None, None
+
+def create_geometry_mask(grid_coords, poly_points):
+    """生成掩码"""
+    flat_grid = grid_coords.reshape(-1, 2)
+    path = Path(poly_points)
+    is_inside = path.contains_points(flat_grid, radius=1e-9)
+    mask_flat = ~is_inside
+    return mask_flat.reshape(GRID_H, GRID_W).astype(np.float32)
+
+# ================= 🚀 主程序 (Main Pipeline) =================
+
+def main():
+    print(f"🚀 [Full Dataset] 开始处理全量数据 (2400组)")
+    print(f"   根目录: {RAW_DATA_ROOT}")
+    
+    # --- 1. 扫描所有待处理的文件夹 ---
+    print("🔍 正在扫描文件结构...")
+    # 匹配 Kn0p06 到 Kn1p00 的所有父文件夹
+    kn_dirs = sorted(glob.glob(os.path.join(RAW_DATA_ROOT, "Kn*")))
+    
+    all_sim_folders = []
+    for kn_dir in kn_dirs:
+        # 在每个 Kn 文件夹下找 simulation*
+        sims = glob.glob(os.path.join(kn_dir, "simulation*"))
+        all_sim_folders.extend(sims)
+    
+    total_files = len(all_sim_folders)
+    print(f"📂 共发现 {len(kn_dirs)} 个 Kn 组, 合计 {total_files} 个仿真案例。")
+    
+    if total_files == 0:
+        print("❌ 未找到数据，请检查路径结构！")
+        return
+
+    # --- 2. 准备网格与内存预分配 (Pre-allocation) ---
+    # 预分配内存是处理大量数据的关键，避免 list append 导致的内存碎片和扩容开销
+    print("💾 预分配内存空间...")
+    
+    x = np.linspace(X_RANGE[0], X_RANGE[1], GRID_W)
+    y = np.linspace(Y_RANGE[0], Y_RANGE[1], GRID_H)
+    grid_x, grid_y = np.meshgrid(x, y)
+    grid_coords_template = np.stack((grid_x, grid_y), axis=-1) # (128, 192, 2)
+
+    # 预分配大数组 [N, ...]
+    # 使用 float32 节省一半内存
+    dataset_params = np.zeros((total_files, 2), dtype=np.float32)        # [Kn, Ma]
+    dataset_flows  = np.zeros((total_files, 128, 192, 4), dtype=np.float32) # Flow
+    dataset_masks  = np.zeros((total_files, 128, 192), dtype=np.float32)    # Mask
+    
+    valid_count = 0
+    valid_indices = [] # 记录成功的索引
+
+    # --- 3. 核心循环 (Processing Loop) ---
+    print("⚙️ 开始插值与处理 (这可能需要一段时间)...")
+    
+    # 使用 tqdm 显示进度条
+    pbar = tqdm(total=total_files, unit="sample")
+    
+    for i, folder in enumerate(all_sim_folders):
+        try:
+            folder_name = os.path.basename(folder)
+            
+            # A. 解析参数
+            kn, ma = parse_folder_params(folder_name)
+            if kn is None: 
+                pbar.update(1)
+                continue
+
+            # B. 寻找文件
+            grid_search = glob.glob(os.path.join(folder, "output", "grid*.dat"))
+            surf_path = os.path.join(folder, "surf", "geom.surf")
+            
+            if not grid_search or not os.path.exists(surf_path):
+                pbar.update(1)
+                continue
+                
+            grid_path = grid_search[0] # 默认取第一个
+            
+            # C. 读取与插值
+            raw_coords, raw_values = parse_grid_file(grid_path)
+            geo_points = parse_surf_file(surf_path)
+            
+            if raw_coords is None or geo_points is None:
+                pbar.update(1)
+                continue
+            
+            # 核心耗时步: 插值
+            # 注意: 如果不同case的原始网格raw_coords是完全固定的，这里可以优化。
+            # 但为了安全起见(适应自适应网格)，我们对每个case单独插值。
+            interp_values = griddata(
+                raw_coords, 
+                raw_values, 
+                (grid_x, grid_y), 
+                method='linear', 
+                fill_value=0.0
+            )
+            
+            # D. Masking
+            fluid_mask = create_geometry_mask(grid_coords_template, geo_points)
+            fluid_mask_4ch = np.repeat(fluid_mask[:, :, np.newaxis], 4, axis=2)
+            cleaned_flow = interp_values * fluid_mask_4ch
+            
+            # E. 存入预分配数组
+            dataset_params[valid_count] = [kn, ma]
+            dataset_flows[valid_count] = cleaned_flow
+            dataset_masks[valid_count] = fluid_mask
+            
+            valid_indices.append(i) # 记录这个 slot 是有效的
+            valid_count += 1
+            
+        except Exception as e:
+            print(f"\n⚠️ 处理出错 {folder_name}: {e}")
+        
+        pbar.update(1)
+        
+        # 每处理100个手动触发垃圾回收，防止内存泄漏
+        if i % 100 == 0:
+            gc.collect()
+
+    pbar.close()
+    
+    # --- 4. 裁剪数组 (去除预分配但未使用的空槽) ---
+    if valid_count < total_files:
+        print(f"⚠️ 只有 {valid_count}/{total_files} 组数据有效，正在裁剪数组...")
+        dataset_params = dataset_params[:valid_count]
+        dataset_flows = dataset_flows[:valid_count]
+        dataset_masks = dataset_masks[:valid_count]
+    else:
+        print(f"✅ 所有 {valid_count} 组数据处理成功。")
+
+    # --- 5. 归一化 (Normalization) ---
+    print("📊 正在执行全局归一化...")
+    
+    # Input Normalization (Kn, Ma) -> [-1, 1]
+    norm_params = np.zeros_like(dataset_params)
+    norm_params[:, 0] = 2 * (dataset_params[:, 0] - KN_MIN) / (KN_MAX - KN_MIN) - 1.0
+    norm_params[:, 1] = 2 * (dataset_params[:, 1] - MA_MIN) / (MA_MAX - MA_MIN) - 1.0
+    
+    # Output Normalization (Flow) -> [-1, 1] based on current dataset stats
+    norm_flows = np.zeros_like(dataset_flows)
+    flow_stats = {}
+    
+    # 扩展 Mask 维度用于广播: [N, H, W] -> [N, H, W, 1]
+    mask_broad = dataset_masks[..., np.newaxis]
+    
+    for c in range(NUM_CHANNELS):
+        channel_data = dataset_flows[..., c]
+        # 仅利用 fluid 区域计算 min/max
+        # 提取所有 mask=1 的像素点
+        valid_pixels = channel_data[dataset_masks == 1.0]
+        
+        _min, _max = float(valid_pixels.min()), float(valid_pixels.max())
+        flow_stats[VAR_NAMES[c]] = {'min': _min, 'max': _max}
+        
+        print(f"   通道 {VAR_NAMES[c]}: min={_min:.4f}, max={_max:.4f}")
+        
+        if _max > _min:
+            # 归一化公式
+            norm_channel = 2 * (channel_data - _min) / (_max - _min) - 1.0
+            # 再次应用 Mask 确保固体内部为 0 (或者 -1/0，取决于你的习惯，这里保持0更安全)
+            norm_flows[..., c] = norm_channel * dataset_masks
+        else:
+            norm_flows[..., c] = 0.0
+
+    # --- 6. 转为 PyTorch Tensor 并保存 ---
+    print("📦 正在打包 Tensor...")
+    
+    tensor_params = torch.tensor(norm_params, dtype=torch.float32) # [2400, 2]
+    tensor_flow = torch.tensor(norm_flows, dtype=torch.float32)    # [2400, 128, 192, 4]
+    tensor_mask = torch.tensor(dataset_masks, dtype=torch.float32).unsqueeze(1) # [2400, 1, 128, 192]
+    
+    # 网格坐标重复 N 次 (如果显存不够，这一步可以在 Dataset 里动态生成，但这里为了方便先存好)
+    # [2400, 128, 192, 2]
+    tensor_grid = torch.tensor(np.tile(grid_coords_template[np.newaxis, ...], (valid_count, 1, 1, 1)), dtype=torch.float32)
+
+    final_data = {
+        "input_params": tensor_params,
+        "flow_label": tensor_flow,
+        "mask": tensor_mask,
+        "grid_coords": tensor_grid,
+        "flow_stats": flow_stats, # 保存统计量用于反归一化
+        "description": "Unified Grid 128x192, Full Dataset (Kn 0.06-1.0)"
+    }
+
+    save_path = os.path.join(OUTPUT_DIR, "cylinder_full_2400.pt")
+    print(f"💾 正在保存至 {save_path} ...")
+    torch.save(final_data, save_path)
+    
+    print(f"\n🎉 全量处理完成! 文件大小预估: ~1GB")
+    print(f"   Input Shape: {tensor_params.shape}")
+    print(f"   Label Shape: {tensor_flow.shape}")
+
+if __name__ == "__main__":
+    main()