vis/umap_plot.py

import torch
import numpy as np
import pandas as pd
from umap import UMAP
import matplotlib.pyplot as plt
from matplotlib.colors import LinearSegmentedColormap

# base_path = '../run-cls/rn18-diff-16-mamba-pcs-768-ce-32-0.001-50'
# img_path = 'umap_before.png'
base_path = '../run-cls/rn18-diff-16-mamba-pcs-768-ce-32-0.001-50/uda_r2'
img_path = 'umap_after.png'

# —— 0. 你的两个函数 —— #
def get_label_dict():
    """
    返回一个字典 { 描述(str): 标签(int, 0 或 1) }
    比如:
      return {
        "image_0001": 1,
        "image_0234": 0,
        # ...
      }
    """
    df = pd.read_excel('../dataset/r2_case.xlsx')
    df = df[~pd.isna(df['TAG'])]

    tags = {}

    for _, row in df.iterrows():
        desc = row['SEQUENCE - D-type amino acid substitution']
        tag = row['TAG']
        if tag.lower() == 'improved':
            tags[desc] = 1
        elif tag.lower() == 'not improved':
            tags[desc] = 0

    return tags


def get_score_dict():
    """
    返回 { 描述: Float in [0,1] }，
    用于背景点的颜色映射。
    """
    df = pd.read_csv(f'{base_path}/feature_preds.csv')
    desc = df['seq'].values
    scores = df['pred'].values
    minimum, maximum = np.min(scores), np.max(scores)
    scores = (scores - minimum) / (maximum - minimum)
    tags = {
        desc: score
        for desc, score in zip(desc, scores)
    }

    return tags

# —— 1. 加载特征 —— #
data_dict = torch.load(f'{base_path}/features.pth', map_location='cpu')
descs = list(data_dict.keys())
features = np.vstack([
    data_dict[d].cpu().numpy() if isinstance(data_dict[d], torch.Tensor)
    else np.array(data_dict[d])
    for d in descs
])
print(f"共 {features.shape[0]} 个样本，特征维度 {features.shape[1]}")

# —— 2. 拿到标签与分数字典 —— #
label_dict = get_label_dict()
score_dict = get_score_dict()

# —— 3. 用 UMAP 降到 2 维 —— #
umap = UMAP(n_components=2, metric='euclidean')
points_2d = umap.fit_transform(features)

# —— 4. 分别收集索引 —— #
idx1     = [i for i,d in enumerate(descs) if d in label_dict and label_dict[d]==1]
idx0     = [i for i,d in enumerate(descs) if d in label_dict and label_dict[d]==0]
idx_rest = [i for i,d in enumerate(descs) if d not in label_dict]

# —— 5. 构造渐变色 —— #
# skyblue ('#87CEEB') → peachpuff ('#FFDAB9')
cmap = LinearSegmentedColormap.from_list('bg_cmap', ["#6EB1EC", "#E69C98"])
# 背景点用它们对应的 score 决定颜色，缺失则给 0.5
bg_scores = np.array([ score_dict.get(descs[i], 0.5) for i in idx_rest ])
bg_colors = cmap(bg_scores)

# —— 6. 绘图 —— #
plt.figure(figsize=(3,3))

# 背景点（渐变色）
plt.scatter(
    points_2d[idx_rest,0],
    points_2d[idx_rest,1],
    s=20,
    c=bg_colors,
    alpha=0.8,
    label='Background'
)

# 标签 0
plt.scatter(
    points_2d[idx0,0],
    points_2d[idx0,1],
    s=60,
    c="#218BE7",
    alpha=1.0,
    label='Not Improved'
)

# 标签 1
plt.scatter(
    points_2d[idx1,0],
    points_2d[idx1,1],
    s=60,
    c="#E76760",
    alpha=1.0,
    label='Improved'
)

# # 标注被高亮的点
# for i in idx0 + idx1:
#     plt.annotate(
#         descs[i],
#         (points_2d[i,0], points_2d[i,1]),
#         textcoords='offset points',
#         xytext=(2,2),
#         fontsize=8,
#         color='black'
#     )

# plt.title('UMAP Visualization', fontsize=18)
plt.axis('off')
# plt.legend(loc='upper left', frameon=False, fontsize=14)
plt.tight_layout()
# —— 可选：保存图像 —— #
plt.savefig(img_path, bbox_inches='tight', pad_inches=0, dpi=300)
# plt.show()