TSNE画特征图

TSNE的可视化特征结果

一、可视化特征或embeddings

1.1 二维的值

对bert输出层的可视化（这是个二维的，batch, hidden_state）
codeantenna

1.2 三维的值

对bert的last_state进行可视化（这个是三维的,batch,seq_length, hidden_state）

def tsne_plot_similar_words_png(title, embedding_clusters, a, filename=""):
    plt.figure(figsize=(16, 9))
    colors = cm.rainbow(np.linspace(0, 1, len(embedding_clusters)))
    i = 1
    for embeddings, color in zip(embedding_clusters, colors):
        x = embeddings[:, 0]
        y = embeddings[:, 1]
        plt.scatter(x, y, c=color, alpha=a)
        plt.text(x.mean(), y.mean(), color='white', weight='bold', fontsize=13, path_effects=[PathEffects.withStroke(linewidth=3,
                                                                                                                             foreground="black", alpha=0.7)])
        i += 1
    plt.title(title)
    plt.grid(True)
    plt.xlim(-200, 200)
    plt.ylim(-200, 200)
    plt.show()
    # plt.savefig(filename, format='png', dpi=150, bbox_inches='tight')


def batch_size_length_state():
    embed_cluster = outputs_train.last_hidden_state.detach().numpy()
    n, m, k = embed_cluster.shape

    print(n, m, k)
    perplexity = 30
    tsne = TSNE(perplexity=perplexity, n_components=2, init='pca', n_iter=3500, random_state=32)
    embeddings_en_2d_gif = np.array(tsne.fit_transform(embed_cluster.reshape(n * m, k))).reshape(n, m, 2)
    tsne_plot_similar_words_png('Vizualizing similar words from Google News using t-SNE (perplexity={})'.format(perplexity), embeddings_en_2d_gif, 0.6)

# batch_size_length_state()

参考

上述的代码微调一下都可以使用
github的开源代码
github上复杂的生成方法