一些小修改

2025-11-19 12:23:17 +08:00
parent 95d719cc1e
commit 80ae03c8c1
25 changed files with 2291 additions and 17 deletions
--- a/GPUMD/t-SNE/t-SNE.py
+++ b/GPUMD/t-SNE/t-SNE.py
@@ -0,0 +1,140 @@
+from pathlib import Path
+import numpy as np
+import matplotlib
+matplotlib.use("Agg")
+import matplotlib.pyplot as plt
+from sklearn.manifold import TSNE
+from sklearn.decomposition import PCA
+
+def tsne_dir_shared_coords(
+    dir_path: str,
+    *,
+    metric: str = "euclidean",     # 可试 "cosine"；想保留尺度差异用 "euclidean"
+    perplexity: float = 50.0,      # 30k~50k 样本建议 30~50
+    n_iter: int = 1000,
+    early_exaggeration: float = 12.0,
+    learning_rate = "auto",
+    standardize: bool = False,
+    pca_dim: int | None = None,    # 先用 PCA 降到 pca_dim(如 20) 再跑 t-SNE，可提速
+    context: bool = True,
+    make_joint: bool = True,
+    init: str = "pca",
+    random_state: int = 42
+) -> None:
+    p = Path(dir_path)
+    if not p.is_dir():
+        raise ValueError(f"{dir_path!r} 不是有效文件夹")
+
+    files = sorted(p.glob("*.npy"))
+    if not files:
+        print(f"目录 {p} 中未找到 .npy 文件")
+        return
+
+    X_list, paths, counts = [], [], []
+    for f in files:
+        try:
+            data = np.load(f)
+            if data.ndim != 2:
+                print(f"[跳过] {f.name}: 期望二维数组，实际 shape={data.shape}")
+                continue
+
+            # 统一到 (n_samples, 30)
+            if data.shape[1] == 30:
+                X = data
+            elif data.shape[0] == 30:
+                X = data.T
+            else:
+                print(f"[跳过] {f.name}: shape={data.shape}, 未检测到 30 维特征")
+                continue
+
+            mask = np.isfinite(X).all(axis=1)
+            if not np.all(mask):
+                X = X[mask]
+                print(f"[提示] {f.name}: 移除了含 NaN/Inf 的样本行")
+
+            if X.shape[0] < 3:
+                print(f"[跳过] {f.name}: 样本数过少(n={X.shape[0]})")
+                continue
+
+            X_list.append(X)
+            paths.append(f)
+            counts.append(X.shape[0])
+        except Exception as e:
+            print(f"[错误] 读取 {f.name} 失败: {e}")
+
+    if not X_list:
+        print("未找到可用的数据文件")
+        return
+
+    X_all = np.vstack(X_list)
+
+    if standardize:
+        mean = X_all.mean(axis=0)
+        std = X_all.std(axis=0); std[std == 0] = 1.0
+        X_all = (X_all - mean) / std
+
+    if pca_dim is not None and pca_dim > 2:
+        X_all = PCA(n_components=pca_dim, random_state=random_state).fit_transform(X_all)
+
+    tsne = TSNE(
+        n_components=2,
+        metric=metric,
+        perplexity=float(perplexity),
+        early_exaggeration=float(early_exaggeration),
+        learning_rate=learning_rate,
+        init=init,
+        random_state=random_state,
+        method="barnes_hut",   # 适合大样本
+        angle=0.5,
+        verbose=0,
+    )
+    Z_all = tsne.fit_transform(X_all)
+
+    # 统一坐标轴范围
+    x_min, x_max = float(Z_all[:, 0].min()), float(Z_all[:, 0].max())
+    y_min, y_max = float(Z_all[:, 1].min()), float(Z_all[:, 1].max())
+    pad_x = 0.05 * (x_max - x_min) if x_max > x_min else 1.0
+    pad_y = 0.05 * (y_max - y_min) if y_max > y_min else 1.0
+
+    colors = [
+        "#1f77b4","#ff7f0e","#2ca02c","#d62728","#9467bd",
+        "#8c564b","#e377c2","#7f7f7f","#bcbd22","#17becf"
+    ]
+
+    # 分文件出图
+    start = 0
+    for i, (f, n) in enumerate(zip(paths, counts)):
+        Zi = Z_all[start:start + n]; start += n
+        fig, ax = plt.subplots(figsize=(6, 5), dpi=150)
+        if context:
+            ax.scatter(Z_all[:, 0], Z_all[:, 1], s=5, c="#cccccc", alpha=0.35, edgecolors="none", label="All")
+        ax.scatter(Zi[:, 0], Zi[:, 1], s=8, c=colors[i % len(colors)], alpha=0.9, edgecolors="none", label=f.name)
+        ax.set_title(f"{f.name} • t-SNE(shared) (perp={perplexity}, metric={metric})", fontsize=9)
+        ax.set_xlabel("t-SNE-1"); ax.set_ylabel("t-SNE-2")
+        ax.set_xlim(x_min - pad_x, x_max + pad_x); ax.set_ylim(y_min - pad_y, y_max + pad_y)
+        ax.grid(True, linestyle="--", linewidth=0.3, alpha=0.5)
+        if context: ax.legend(loc="best", fontsize=8, frameon=False)
+        fig.tight_layout()
+        out_png = f.with_suffix("").as_posix() + "_tsne_shared.png"
+        fig.savefig(out_png); plt.close(fig)
+        print(f"[完成] {f.name} -> {out_png}")
+
+    # 总览图
+    if make_joint:
+        start = 0
+        fig, ax = plt.subplots(figsize=(7, 6), dpi=150)
+        for i, (f, n) in enumerate(zip(paths, counts)):
+            Zi = Z_all[start:start + n]; start += n
+            ax.scatter(Zi[:, 0], Zi[:, 1], s=8, c=colors[i % len(colors)], alpha=0.85, edgecolors="none", label=f.name)
+        ax.set_title(f"t-SNE(shared) overview (perp={perplexity}, metric={metric})", fontsize=10)
+        ax.set_xlabel("t-SNE-1"); ax.set_ylabel("t-SNE-2")
+        ax.set_xlim(x_min - pad_x, x_max + pad_x); ax.set_ylim(y_min - pad_y, y_max + pad_y)
+        ax.grid(True, linestyle="--", linewidth=0.3, alpha=0.5)
+        ax.legend(loc="best", fontsize=8, frameon=False)
+        fig.tight_layout()
+        out_png = Path(dir_path) / "tsne_shared_overview.png"
+        fig.savefig(out_png.as_posix()); plt.close(fig)
+        print(f"[完成] 总览 -> {out_png}")
+
+if __name__ == "__main__":
+    tsne_dir_shared_coords("data")