import numpy as np import pandas as pd from openTSNE import TSNE import plotly.graph_objs as go import matplotlib.pyplot as plt import matplotlib.colors as mcolors from sklearn.decomposition import PCA from scipy.optimize import linear_sum_assignment class TSNE_Plot(): def __init__(self, sentence, embed, label = None, n_clusters :int = 3, n_annotation_positions:int = 20): assert n_clusters > 0, "N must be greater than 0" self.N = n_clusters self.test_X = pd.DataFrame({"text": sentence, "embed": [np.array(i) for i in embed]}) self.test_y = pd.DataFrame({'label':label}) if label is not None else pd.DataFrame({"label": self.cluster()}) self.embed = self.calculate_tsne() self.init_df() self.n_annotation_positions = n_annotation_positions self.show_sentence = [] self.random_sentence() self.annotation_positions = [] self.get_annotation_positions() self.mapping = {} def cluster(self): from sklearn.cluster import KMeans n_components = min(50, len(self.test_X)) pca = PCA(n_components=n_components) compact_embedding = pca.fit_transform(np.array(self.test_X["embed"].tolist())) kmeans = KMeans(n_clusters=self.N) kmeans.fit(compact_embedding) labels = kmeans.labels_ return labels def generate_colormap(self, n_labels): #创建一个均匀分布的颜色映射 color_norm = mcolors.Normalize(vmin=0, vmax=len(n_labels) - 1) # 使用 plt.cm 中预先定义的colormap,你可以自由选择其他colormap如"hsv", "hot", "cool", "viridis"等 scalar_map = plt.cm.ScalarMappable(norm=color_norm, cmap='jet') colormap = {} for label in range(len(n_labels)): # 将颜色值转换为十六进制 color_hex = mcolors.to_hex(scalar_map.to_rgba(label)) colormap[n_labels[label]] = color_hex return colormap def divide_hex_color_by_half(self, hex_color): if len(hex_color) > 0 and hex_color[0] == "#": hex_color = hex_color[1:] red_hex, green_hex, blue_hex = hex_color[0:2], hex_color[2:4], hex_color[4:6] red_half = int(red_hex, 16) // 10 + (255-25) green_half = int(green_hex, 16) // 10 + (255-25) blue_half = int(blue_hex, 16) // 10 + (255-25) half_hex_color = "#{:02x}{:02x}{:02x}".format(red_half, green_half, blue_half) return half_hex_color def get_annotation_positions(self): min_x, max_x = self.df['x'].min()-1, self.df['x'].max()+2 n = self.n_annotation_positions y_min, y_max = self.df['y'].min() * 3, self.df['y'].max() * 3 add = 0 if n % 2 == 0 else 1 y_values = np.linspace(y_min, y_max, n//2+add) left_positions = [(min_x, y) for y in y_values] right_positions = [(max_x, y) for y in y_values] self.annotation_positions = left_positions + right_positions def euclidean_distance(self, p1, p2): return np.sqrt((p1[0] - p2[0])**2 + (p1[1] - p2[1])**2) def map_points(self): # Get points from the dataframe using the show_sentence indices points1 = [(self.embed[i][0], self.embed[i][1]) for i in self.show_sentence] # Create a distance matrix between the points distance_matrix = np.zeros((len(points1), len(self.annotation_positions))) for i, point1 in enumerate(points1): for j, point2 in enumerate(self.annotation_positions): distance_matrix[i, j] = self.euclidean_distance(point1, point2) # Apply linear_sum_assignment to find the optimal mapping row_ind, col_ind = linear_sum_assignment(distance_matrix) for i, j in zip(row_ind, col_ind): self.mapping[self.show_sentence[i]] = self.annotation_positions[j] def show_text(self, show_sentence, text): sentence = [] for i in range(len(text)): if i in show_sentence: s = text[i][:10] + "..." + text[i][-10:] sentence.append(s) else: sentence.append("") return sentence def init_df(self): X, Y = np.split(self.embed, 2, axis=1) data = { "x": X.flatten(), "y": Y.flatten(), } self.df = pd.DataFrame(data) def format_data(self): sentence = self.show_text(self.show_sentence, self.test_X["text"]) X, Y = np.split(self.embed, 2, axis=1) n = len(self.test_X) data = { "x": X.flatten(), "y": Y.flatten(), "label": self.test_y["label"], "sentence" : sentence, "size" : [20 if i in self.show_sentence else 10 for i in range(n)], "pos" : [{"x_offset": self.mapping.get(i, (0, 0))[0], "y_offset": self.mapping.get(i, (0, 0))[1]} for i in range(n)], "annotate" : [True if i in self.show_sentence else False for i in range(n)], } self.df = pd.DataFrame(data) def calculate_tsne(self): embed = np.array(self.test_X["embed"].tolist()) n_components = min(50, len(self.test_X)) pca = PCA(n_components=n_components) compact_embedding = pca.fit_transform(embed) tsne = TSNE( perplexity=30, metric="cosine", n_jobs=8, random_state=42, verbose=False, ) embedding_train = tsne.fit(compact_embedding) embedding_train = embedding_train.optimize(n_iter=1000, momentum=0.8) return embedding_train def random_sentence(self): #多次随机可能会影响可视化结果 n_samples = len(self.test_y) show_sentence = [] while len(show_sentence) < self.n_annotation_positions: show_sentence.append(np.random.randint(0, n_samples)) show_sentence = list(set(show_sentence)) # 确保每个标签至少有一个句子,用在show_sentence中最多的标签的句子来补充 label_count = self.test_y["label"].value_counts() max_label = label_count.index[0] max_count = label_count[0] for i in range(max_count): for j in range(len(label_count)): if label_count[j] == i: show_sentence.append(self.test_y[self.test_y["label"] == label_count.index[j]].index[0]) self.show_sentence = list(set(show_sentence)) def plot(self, return_fig=False): min_x, max_x = self.df['x'].min()-1, self.df['x'].max()+2 fig = go.Figure() fig = go.Figure(layout=go.Layout( autosize=False, # 禁止图像自动调整大小 height=800, # 您可以根据需要调整这个值 width=1500, # 您可以根据需要调整这个值 # plot_bgcolor="#262626", )) label_colors = self.generate_colormap(self.df['label'].unique()) line_legend_group = "lines" # 为每个类别的点创建散点图 for label, color in label_colors.items(): mask = self.df["label"] == label fig.add_trace(go.Scatter(x=self.df[mask]['x'], y=self.df[mask]['y'], mode='markers', marker=dict(color=color, size=self.df[mask]['size']), # 添加 size 参数 showlegend=True, legendgroup=line_legend_group, name = str(label)) ) # 为每个句子创建注释 for x, y, label, sentence, pos, annotate in zip(self.df.x, self.df.y, self.df.label, self.df.sentence, self.df.pos, self.df.annotate): if not sentence: continue if not annotate: continue # pos在左边 criteria = (pos["x_offset"] - min_x) < 1e-2 sentence_annotation = dict( x=pos["x_offset"], y=pos["y_offset"], xref="x", yref="y", text=sentence, showarrow=False, xanchor="right" if criteria else 'left', yanchor='middle', font=dict(color="black"), bordercolor=label_colors.get(label, "black"), borderpad=2, bgcolor=self.divide_hex_color_by_half(label_colors.get(label, "black")) ) fig.add_annotation(sentence_annotation) x_start = x - 1 if criteria else x + 1 x_turn = x - 0.5 if criteria else x + 0.5 y_turn = y fig.add_trace(go.Scatter(x=[pos["x_offset"], x_start, x_turn, x], y=[pos["y_offset"], pos["y_offset"], y_turn, y], mode='lines', line=dict(color=label_colors.get(label, "black")), showlegend=False, legendgroup=line_legend_group)) # 取消坐标轴的数字 fig.update_xaxes(tickvals=[]) fig.update_yaxes(tickvals=[]) if not return_fig: fig.show() else: return fig def tsne_plot(self, n_sentence = 20, return_fig=False): # 计算t-SNE,返回降维后的数据,每个元素为一个二维向量 embedding_train = self.calculate_tsne() # 随机抽取显示文本, n为抽取的数量,show_sentence为一个列表,每个元素为显示文本的索引 if self.n_annotation_positions != min(n_sentence, len(self.test_y)): self.n_annotation_positions = min(n_sentence, len(self.test_y)) self.random_sentence() self.get_annotation_positions() # find the optimal sentence positions self.map_points() # 格式化数据,输出为一个pandas的DataFrame,包含x, y, label, sentence, sentence_pos, size # x, y为降维后的坐标,label为类别,sentence为显示的文本,sentence_pos为文本的位置("left", "right"),size为被选中文本的大小 self.format_data() # self.df = self.df.sort_values('y').reset_index(drop=True) if not return_fig: # 绘制图像 self.plot() else: return self.plot(return_fig=return_fig)