pokouqiege/utils/data_process/show_feature.py

import os
import torch
import numpy as np
import matplotlib.pyplot as plt
from datetime import datetime


def visualize_feature_map(feature_logits,
                          save_root=r"/home/zhaoyinghan/py_ws/code/pokou/MultiVisionModels/models/line_detect/train_results/feature",
                          max_saved=5):
    """
    Visualize 5 feature maps stacked together. Saves result inside a timestamp folder.
    Keeps only the last `max_saved` folders.

    Args:
        feature_logits: Tensor [B, 1, H, W]
        save_root: root folder to save images
        max_saved: maximum number of timestamp folders to keep
    """

    with torch.no_grad():
        os.makedirs(save_root, exist_ok=True)

        # -----------------------
        # timestamp folder
        # -----------------------
        timestamp = datetime.now().strftime("%Y%m%d_%H%M%S")
        save_dir = os.path.join(save_root, timestamp)
        os.makedirs(save_dir, exist_ok=True)

        # -----------------------
        # cleanup old folders
        # -----------------------
        all_dirs = [d for d in os.listdir(save_root) if os.path.isdir(os.path.join(save_root, d))]
        all_dirs.sort()  # °´Ãû×ÖÅÅÐò£¬Ê±¼ä´Á˳Ðò
        if len(all_dirs) > max_saved:
            for d in all_dirs[:-max_saved]:
                try:
                    full_path = os.path.join(save_root, d)
                    import shutil
                    shutil.rmtree(full_path)
                    print(f"[visualize_feature_map] Removed old folder: {full_path}")
                except Exception as e:
                    print(f"[visualize_feature_map] Failed to remove {full_path}: {e}")

        # -----------------------
        # take first 5 feature maps
        # -----------------------
        B = feature_logits.size(0)
        num = min(5, B)
        maps = feature_logits[:num, 0, ...].detach().cpu()  # shape (num, H, W)

        # -----------------------
        # stack & normalize
        # -----------------------
        stacked = maps.sum(dim=0).numpy()
        stacked = (stacked - stacked.min()) / (stacked.max() - stacked.min() + 1e-6)

        # -----------------------
        # save image
        # -----------------------
        save_path = os.path.join(save_dir, "stacked_feature.png")
        plt.figure(figsize=(6, 6))
        plt.imshow(stacked, cmap="viridis")
        plt.axis("off")
        plt.savefig(save_path, bbox_inches="tight", pad_inches=0)
        plt.close()

        print(f"[visualize_feature_map] Saved stacked feature ¡ú {save_path}")