Python自定義指標聚類範例程式碼

def Eu_dist(data, center):
    """ 以 歐氏距離 為聚類準則的距離計算函數
        data: 形如 [n_sample, n_feature] 的 tensor
        center: 形如 [n_cluster, n_feature] 的 tensor"""
    data = data.unsqueeze(1)
    center = center.unsqueeze(0)
    dist = ((data - center) ** 2).sum(dim=2)
    return dist

import torch
import numpy as np
import matplotlib.pyplot as plt
Adam = torch.optim.Adam
 
def get_progress(current, target, bar_len=30):
    """ current: 當前完成任務數
        target: 任務總數
        bar_len: 進度條長度
        return: 進度條字串"""
    assert current <= target
    percent = round(current / target * 100, 1)
    unit = 100 / bar_len
    solid = int(percent / unit)
    hollow = bar_len - solid
    return "■" * solid + "□" * hollow + f" {current}/{target}({percent}%)"
 
 
class Cluster:
    """ 聚類器
        n_cluster: 簇中心數
        dist_fun: 距離計算函數
            kwargs:
                data: 形如 [n_sample, n_feather] 的 tensor
                center: 形如 [n_cluster, n_feature] 的 tensor
            return: 形如 [n_sample, n_cluster] 的 tensor
        init: 初始簇中心
        max_iter: 最大迭代輪數
        lr: 中心點座標學習率
        stop_thresh: 停止訓練的loss浮動閾值
        cluster_centers_: 聚類中心
        labels_: 聚類結果"""
 
    def __init__(self, n_cluster, dist_fun, init=None, max_iter=300, lr=0.08, stop_thresh=1e-4):
        self._n_cluster = n_cluster
        self._dist_fun = dist_fun
        self._max_iter = max_iter
        self._lr = lr
        self._stop_thresh = stop_thresh
        # 初始化引數
        self.cluster_centers_ = None if init is None else torch.FloatTensor(init)
        self.labels_ = None
        self._bar_len = 20
 
    def fit(self, data):
        """ data: 形如 [n_sample, n_feature] 的 tensor
            return: loss浮動紀錄檔"""
        if self.cluster_centers_ is None:
            self._init_cluster(data, self._max_iter // 5)
        log = self._train(data, self._max_iter, self._lr)
        # 開始若干輪次的訓練，得到loss浮動紀錄檔
        return log
 
    def classify(self, data, show=False):
        """ data: 形如 [n_sample, n_feature] 的 tensor
            show: 繪製分類結果
            return: 分類標籤"""
        dist = self._dist_fun(data, self.cluster_centers_)
        self.labels_ = dist.argmin(axis=1)
        # 將標籤載入到範例屬性
        if show:
            for idx in range(self._n_cluster):
                container = data[self.labels_ == idx]
                plt.scatter(container[:, 0], container[:, 1], alpha=0.7)
            plt.scatter(self.cluster_centers_[:, 0], self.cluster_centers_[:, 1], c="gold", marker="p", s=50)
            plt.show()
        return self.labels_
 
    def _init_cluster(self, data, epochs):
        self.cluster_centers_ = data.mean(dim=0).reshape(1, -1)
        for idx in range(1, self._n_cluster):
            dist = np.array(self._dist_fun(data, self.cluster_centers_).min(dim=1)[0])
            new_cluster = data[np.random.choice(range(data.shape[0]), p=dist / dist.sum())].reshape(1, -1)
            # 取新的中心點
            self.cluster_centers_ = torch.cat([self.cluster_centers_, new_cluster], dim=0)
            progress = get_progress(idx, self._n_cluster, bar_len=self._n_cluster if self._n_cluster <= self._bar_len else self._bar_len)
            print(f"rCluster Init: {progress}", end="")
            self._train(data, epochs, self._lr * 2.5, init=True)
            # 初始化簇中心時使用較大的lr
 
    def _train(self, data, epochs, lr, init=False):
        center = self.cluster_centers_.cuda()
        center.requires_grad = True
        data = data.cuda()
        optimizer = Adam([center], lr=lr)
        # 將中心資料載入到 GPU 上
        init_patience = int(epochs ** 0.5)
        patience = init_patience
        update_log = []
        min_loss = np.inf
        for epoch in range(epochs):
            # 對樣本分類並更新中心點
            sample_dist = self._dist_fun(data, center).min(dim=1)
            self.labels_ = sample_dist[1]
            loss = sum([sample_dist[0][self.labels_ == idx].mean() for idx in range(len(center))])
            # loss 函數: 所有樣本到中心點的最小距離和 - 中心點間的最小間隔
            loss.backward()
            optimizer.step()
            optimizer.zero_grad()
            # 反向傳播梯度更新中心點
            loss = loss.item()
            progress = min_loss - loss
            update_log.append(progress)
            if progress > 0:
                self.cluster_centers_ = center.cpu().detach()
                min_loss = loss
                # 脫離計算圖後記錄中心點
            if progress < self._stop_thresh:
                patience -= 1
                # 耐心值減少
                if patience < 0:
                    break
                    # 耐心值歸零時退出
            else:
                patience = init_patience
                # 恢復耐心值
            progress = get_progress(init_patience - patience, init_patience, bar_len=self._bar_len)
            if not init:
                print(f"rCluster: {progress}titer: {epoch + 1}", end="")
        if not init:
            print("")
        return torch.FloatTensor(update_log)

def DIoU_dist(boxes, anchor):
    """ 以 DIoU 為聚類準則的距離計算函數
        boxes: 形如 [n_sample, 2] 的 tensor
        anchor: 形如 [n_cluster, 2] 的 tensor"""
    n_sample = boxes.shape[0]
    n_cluster = anchor.shape[0]
    dist = Eu_dist(boxes, anchor)
    # 計算歐式距離
    union_inter = torch.prod(boxes, dim=1).reshape(-1, 1) + torch.prod(anchor, dim=1).reshape(1, -1)
    boxes = boxes.unsqueeze(1).repeat(1, n_cluster, 1)
    anchor = anchor.unsqueeze(0).repeat(n_sample, 1, 1)
    compare = torch.stack([boxes, anchor], dim=2)
    # 組合檢測框與 anchor 的資訊
    diag = torch.sum(compare.max(dim=2)[0] ** 2, dim=2)
    dist /= diag
    # 計算外接矩形的對角線長度
    inter = torch.prod(compare.min(dim=2)[0], dim=2)
    iou = inter / (union_inter - inter)
    # 計算 IoU
    dist += 1 - iou
    return dist