silk-road commited on
Commit
62faef9
1 Parent(s): a0cdf41

Create tsne.py

Browse files
Files changed (1) hide show
  1. tsne.py +259 -0
tsne.py ADDED
@@ -0,0 +1,259 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import numpy as np
2
+ import pandas as pd
3
+ from openTSNE import TSNE
4
+ import plotly.graph_objs as go
5
+ import matplotlib.pyplot as plt
6
+ import matplotlib.colors as mcolors
7
+ from sklearn.decomposition import PCA
8
+ from scipy.optimize import linear_sum_assignment
9
+
10
+ class TSNE_Plot():
11
+ def __init__(self, sentence, embed, label = None, n_clusters :int = 3, n_annotation_positions:int = 20):
12
+ assert n_clusters > 0, "N must be greater than 0"
13
+ self.N = n_clusters
14
+ self.test_X = pd.DataFrame({"text": sentence, "embed": [np.array(i) for i in embed]})
15
+ self.test_y = pd.DataFrame({'label':label}) if label is not None else pd.DataFrame({"label": self.cluster()})
16
+ self.embed = self.calculate_tsne()
17
+ self.init_df()
18
+
19
+ self.n_annotation_positions = n_annotation_positions
20
+ self.show_sentence = []
21
+ self.random_sentence()
22
+
23
+
24
+ self.annotation_positions = []
25
+ self.get_annotation_positions()
26
+ self.mapping = {}
27
+
28
+ def cluster(self):
29
+ from sklearn.cluster import KMeans
30
+ n_components = min(50, len(self.test_X))
31
+ pca = PCA(n_components=n_components)
32
+ compact_embedding = pca.fit_transform(np.array(self.test_X["embed"].tolist()))
33
+ kmeans = KMeans(n_clusters=self.N)
34
+ kmeans.fit(compact_embedding)
35
+ labels = kmeans.labels_
36
+ return labels
37
+
38
+ def generate_colormap(self, n_labels):
39
+ #创建一个均匀分布的颜色映射
40
+ color_norm = mcolors.Normalize(vmin=0, vmax=len(n_labels) - 1)
41
+ # 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
42
+ scalar_map = plt.cm.ScalarMappable(norm=color_norm, cmap='jet')
43
+
44
+ colormap = {}
45
+ for label in range(len(n_labels)):
46
+ # 将颜色值转换为十六进制
47
+ color_hex = mcolors.to_hex(scalar_map.to_rgba(label))
48
+ colormap[n_labels[label]] = color_hex
49
+ return colormap
50
+
51
+ def divide_hex_color_by_half(self, hex_color):
52
+ if len(hex_color) > 0 and hex_color[0] == "#":
53
+ hex_color = hex_color[1:]
54
+
55
+ red_hex, green_hex, blue_hex = hex_color[0:2], hex_color[2:4], hex_color[4:6]
56
+
57
+ red_half = int(red_hex, 16) // 10 + (255-25)
58
+ green_half = int(green_hex, 16) // 10 + (255-25)
59
+ blue_half = int(blue_hex, 16) // 10 + (255-25)
60
+
61
+ half_hex_color = "#{:02x}{:02x}{:02x}".format(red_half, green_half, blue_half)
62
+ return half_hex_color
63
+
64
+
65
+ def get_annotation_positions(self):
66
+ min_x, max_x = self.df['x'].min()-1, self.df['x'].max()+2
67
+ n = self.n_annotation_positions
68
+
69
+ y_min, y_max = self.df['y'].min() * 3, self.df['y'].max() * 3
70
+
71
+ add = 0 if n % 2 == 0 else 1
72
+ y_values = np.linspace(y_min, y_max, n//2+add)
73
+
74
+ left_positions = [(min_x, y) for y in y_values]
75
+ right_positions = [(max_x, y) for y in y_values]
76
+
77
+
78
+ self.annotation_positions = left_positions + right_positions
79
+
80
+
81
+ def euclidean_distance(self, p1, p2):
82
+ return np.sqrt((p1[0] - p2[0])**2 + (p1[1] - p2[1])**2)
83
+
84
+ def map_points(self):
85
+ # Get points from the dataframe using the show_sentence indices
86
+ points1 = [(self.embed[i][0], self.embed[i][1]) for i in self.show_sentence]
87
+
88
+ # Create a distance matrix between the points
89
+ distance_matrix = np.zeros((len(points1), len(self.annotation_positions)))
90
+
91
+ for i, point1 in enumerate(points1):
92
+ for j, point2 in enumerate(self.annotation_positions):
93
+ distance_matrix[i, j] = self.euclidean_distance(point1, point2)
94
+
95
+ # Apply linear_sum_assignment to find the optimal mapping
96
+ row_ind, col_ind = linear_sum_assignment(distance_matrix)
97
+
98
+ for i, j in zip(row_ind, col_ind):
99
+ self.mapping[self.show_sentence[i]] = self.annotation_positions[j]
100
+
101
+
102
+ def show_text(self, show_sentence, text):
103
+ sentence = []
104
+ for i in range(len(text)):
105
+ if i in show_sentence:
106
+ s = text[i][:10] + "..." + text[i][-10:]
107
+ sentence.append(s)
108
+ else:
109
+ sentence.append("")
110
+ return sentence
111
+
112
+ def init_df(self):
113
+ X, Y = np.split(self.embed, 2, axis=1)
114
+ data = {
115
+ "x": X.flatten(),
116
+ "y": Y.flatten(),
117
+ }
118
+
119
+ self.df = pd.DataFrame(data)
120
+
121
+
122
+ def format_data(self):
123
+ sentence = self.show_text(self.show_sentence, self.test_X["text"])
124
+ X, Y = np.split(self.embed, 2, axis=1)
125
+ n = len(self.test_X)
126
+ data = {
127
+ "x": X.flatten(),
128
+ "y": Y.flatten(),
129
+ "label": self.test_y["label"],
130
+ "sentence" : sentence,
131
+ "size" : [20 if i in self.show_sentence else 10 for i in range(n)],
132
+ "pos" : [{"x_offset": self.mapping.get(i, (0, 0))[0], "y_offset": self.mapping.get(i, (0, 0))[1]} for i in range(n)],
133
+ "annotate" : [True if i in self.show_sentence else False for i in range(n)],
134
+ }
135
+ self.df = pd.DataFrame(data)
136
+
137
+ def calculate_tsne(self):
138
+ embed = np.array(self.test_X["embed"].tolist())
139
+ n_components = min(50, len(self.test_X))
140
+ pca = PCA(n_components=n_components)
141
+ compact_embedding = pca.fit_transform(embed)
142
+ tsne = TSNE(
143
+ perplexity=30,
144
+ metric="cosine",
145
+ n_jobs=8,
146
+ random_state=42,
147
+ verbose=False,
148
+ )
149
+ embedding_train = tsne.fit(compact_embedding)
150
+ embedding_train = embedding_train.optimize(n_iter=1000, momentum=0.8)
151
+ return embedding_train
152
+
153
+ def random_sentence(self):
154
+ #多次随机可能会影响可视化结果
155
+ n_samples = len(self.test_y)
156
+
157
+ show_sentence = []
158
+ while len(show_sentence) < self.n_annotation_positions:
159
+ show_sentence.append(np.random.randint(0, n_samples))
160
+ show_sentence = list(set(show_sentence))
161
+
162
+ # 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
163
+ label_count = self.test_y["label"].value_counts()
164
+ max_label = label_count.index[0]
165
+ max_count = label_count[0]
166
+ for i in range(max_count):
167
+ for j in range(len(label_count)):
168
+ if label_count[j] == i:
169
+ show_sentence.append(self.test_y[self.test_y["label"] == label_count.index[j]].index[0])
170
+ self.show_sentence = list(set(show_sentence))
171
+
172
+ def plot(self, return_fig=False):
173
+ min_x, max_x = self.df['x'].min()-1, self.df['x'].max()+2
174
+ fig = go.Figure()
175
+ fig = go.Figure(layout=go.Layout(
176
+ autosize=False, # 禁止图像自动调整大小
177
+ height=800, # 您可以根据需要调整这个值
178
+ width=1500, # 您可以根据需要调整这个值
179
+ # plot_bgcolor="#262626",
180
+ ))
181
+
182
+ label_colors = self.generate_colormap(self.df['label'].unique())
183
+
184
+ line_legend_group = "lines"
185
+
186
+ # 为每个类别的点创建散点图
187
+ for label, color in label_colors.items():
188
+ mask = self.df["label"] == label
189
+ fig.add_trace(go.Scatter(x=self.df[mask]['x'], y=self.df[mask]['y'], mode='markers',
190
+ marker=dict(color=color, size=self.df[mask]['size']), # 添加 size 参数
191
+ showlegend=True, legendgroup=line_legend_group,
192
+ name = str(label))
193
+ )
194
+
195
+
196
+ # 为每个句子创建注释
197
+ 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):
198
+ if not sentence:
199
+ continue
200
+ if not annotate:
201
+ continue
202
+ # pos在左边
203
+ criteria = (pos["x_offset"] - min_x) < 1e-2
204
+
205
+ sentence_annotation = dict(
206
+ x=pos["x_offset"],
207
+ y=pos["y_offset"],
208
+ xref="x",
209
+ yref="y",
210
+ text=sentence,
211
+ showarrow=False,
212
+ xanchor="right" if criteria else 'left',
213
+ yanchor='middle',
214
+ font=dict(color="black"),
215
+ bordercolor=label_colors.get(label, "black"),
216
+ borderpad=2,
217
+ bgcolor=self.divide_hex_color_by_half(label_colors.get(label, "black"))
218
+ )
219
+ fig.add_annotation(sentence_annotation)
220
+
221
+ x_start = x - 1 if criteria else x + 1
222
+ x_turn = x - 0.5 if criteria else x + 0.5
223
+ y_turn = y
224
+
225
+ 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',
226
+ line=dict(color=label_colors.get(label, "black")), showlegend=False, legendgroup=line_legend_group))
227
+
228
+ # 取消坐标轴的数字
229
+ fig.update_xaxes(tickvals=[])
230
+ fig.update_yaxes(tickvals=[])
231
+
232
+ if not return_fig:
233
+ fig.show()
234
+ else:
235
+ return fig
236
+
237
+ def tsne_plot(self, n_sentence = 20, return_fig=False):
238
+ # 计算t-SNE,返回降维后的数据,每个元素为一个二维向量
239
+ embedding_train = self.calculate_tsne()
240
+
241
+ # 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
242
+ if self.n_annotation_positions != min(n_sentence, len(self.test_y)):
243
+ self.n_annotation_positions = min(n_sentence, len(self.test_y))
244
+ self.random_sentence()
245
+ self.get_annotation_positions()
246
+
247
+ # find the optimal sentence positions
248
+ self.map_points()
249
+
250
+ # 格式化数据,输出为一个pandas的DataFrame,包含x, y, label, sentence, sentence_pos, size
251
+ # 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
252
+ self.format_data()
253
+ # self.df = self.df.sort_values('y').reset_index(drop=True)
254
+
255
+ if not return_fig:
256
+ # 绘制图像
257
+ self.plot()
258
+ else:
259
+ return self.plot(return_fig=return_fig)