tsne+seaborn 绘制语义空间分布

import numpy as np
import matplotlib

matplotlib.use('Agg')
import matplotlib.pyplot as plt_sne
from sklearn import datasets
from sklearn.manifold import TSNE
import os

plt_sne.rcParams['font.sans-serif'] = ['SimHei']
plt_sne.rcParams['axes.unicode_minus'] = False

def plot_tsne(features, labels, epoch, fileNameDir=None):
    '''
    features:(N*m) N*m大小特征，其中N代表有N个数据，每个数据m维
    label:(N) 有N个标签
    '''
    print(features.shape, labels.shape)
    print(type(features), type(labels))
    print(np.any(np.isnan(features)), np.any(np.isinf(features)))
    features = np.nan_to_num(features)
    if not os.path.exists(fileNameDir):
        os.makedirs(fileNameDir)
    import pandas as pd
    tsne = TSNE(n_components=2, init='pca', random_state=0)
    import seaborn as sns

    # 查看标签的种类有几个
    class_num = len(np.unique(labels))  # 要分类的种类个数  eg:[0, 1, 2, 3]这个就是为4

    try:
        # x_min, x_max = features.min(0), features.max(0)
        # features = (features - x_min) / (x_max - x_min)  # 归一化,消除不同维度的量纲差别
        tsne_features = tsne.fit_transform(features)  # 将特征使用PCA降维至2维
    except:
        tsne_features = tsne.fit_transform(features)

    # 一个类似于表格的数据结构
    df = pd.DataFrame()
    df["y"] = labels
    df["comp1"] = tsne_features[:, 0]
    df["comp2"] = tsne_features[:, 1]

    # hue:根据y列上的数据种类，来生成不同的颜色；
    # style:根据y列上的数据种类，来生成不同的形状点；
    plt_sne.figure(figsize=(15,15))  #修改画布大小

    sns.scatterplot(x=df.comp1.tolist(), y=df.comp2.tolist(), hue=df.y.tolist(), style=df.y.tolist(),
                    palette=sns.color_palette("Set2", class_num),
                    data=df).set(title="T-SNE 语义空间分布图")

    # sns.scatterplot(x=df.comp1.tolist(), y=df.comp2.tolist(), hue=df.y.tolist(),
    #                 palette=sns.color_palette("Set2", class_num),
    #                 data=df).set(title="T-SNE 语义空间分布图")

    plt_sne.savefig(os.path.join(fileNameDir, "%s.jpg") % str(epoch), format="jpg")
    plt_sne.show()


if __name__ == '__main__':
    digits = datasets.load_digits(n_class=2)
    features, labels = digits.data, digits.target  #np.array()
    print(features.shape)
    print(labels.shape)

    plot_tsne(features, labels, "Set3", fileNameDir="test")

引用来源：https://blog.csdn.net/power_kaikaige/article/details/128324084