这里有一个使用split=True和虚拟数据强制分割空半部分的方法。对于左半部分，metric被设置为1用于真实的数据，2用于虚拟数据。对于右半部分，反之亦然。我们需要确保所有 Dataframe 对modality列使用相同的分类顺序，以避免混淆。

from matplotlib import pyplot as plt
import seaborn as sns
import pandas as pd
import numpy as np

sns.set_style('white')
df1 = pd.DataFrame({'modality': pd.Categorical.from_codes(np.random.randint(0, 3, 30), ['a', 'b', 'c']),
                    'Value': np.random.rand(30) * 25 + 50})
df1['metric'] = 1
df1_dummy = pd.DataFrame({'modality': pd.Categorical.from_codes([0], ['a', 'b', 'c']), 'Value': [np.nan]})
df1_dummy['metric'] = 2

df2 = pd.DataFrame({'modality': pd.Categorical.from_codes(np.random.randint(0, 3, 100), ['a', 'b', 'c']),
                    'Value': np.random.randn(100).cumsum() / 10 + 1})
df2['metric'] = 2
df2_dummy = pd.DataFrame({'modality': pd.Categorical.from_codes([0], ['a', 'b', 'c']), 'Value': [np.nan]})
df2_dummy['metric'] = 1

ax1 = sns.violinplot(y='Value', x='modality', hue='metric', palette=['turquoise', 'red'],
                     inner="stick", split=True, data=pd.concat([df1, df1_dummy]))
ax1.legend_.remove()
ax1.set_ylabel('metric 1')

ax2 = ax1.twinx()
sns.violinplot(y='Value', x='modality', hue='metric', palette=['turquoise', 'red'],
               inner="stick", split=True, data=pd.concat([df2, df2_dummy]), ax=ax2)
ax2.set_ylabel('metric 2')

plt.tight_layout()
plt.show()

PS：这是原始代码的一个可能的改编：

• 使用plt.Rectangle((x0+width/2, y0), width/2, height)将小提琴夹在ax2上
• 使用sns.violinplot()的ax=参数指示正确的子图
• 不更改ax的xlim和ylim
• 确保两个 Dataframe 对modality使用相同的分类顺序
• 剪切“内部”行，对于ax1：循环遍历这些行，得到它们的x0和x1，然后将行缩短为x0和(x0+x1)/2
• 类似于ax2：循环遍历这些行，得到它们的x0和x1，然后将行缩短为(x0+x1)/2和x1
• 更新ax2的图例，将其与ax1的图例合并，然后删除ax1的图例

from matplotlib import pyplot as plt
import seaborn as sns
import pandas as pd
import numpy as np

df1 = pd.DataFrame({'modality': pd.Categorical.from_codes(np.random.randint(0, 3, 30), ['a', 'b', 'c']),
                    'Value': np.random.rand(30) * 25 + 50})
df1['metric'] = 1
df2 = pd.DataFrame({'modality': pd.Categorical.from_codes(np.random.randint(0, 3, 100), ['a', 'b', 'c']),
                    'Value': np.random.randn(100).cumsum() / 10 + 1})
df2['metric'] = 2

fig, ax1 = plt.subplots()

sns.set_style('white')
palette1 = 'Set2'
palette2 = 'Set1'

sns.violinplot(y=df1.Value, x=df1.modality, hue=df1.metric, palette=palette1, inner="stick", ax=ax1)
for violin in ax1.collections:
    bbox = violin.get_paths()[0].get_extents()
    x0, y0, width, height = bbox.bounds
    violin.set_clip_path(plt.Rectangle((x0, y0), width / 2, height, transform=ax1.transData))
for line in ax1.lines:
    x = line.get_xdata()
    line.set_xdata([x[0], np.mean(x)])

ax1.set_ylabel("metric1")
ax1.set_xlabel("Modality")

ax2 = ax1.twinx()
sns.violinplot(y=df2.Value, x=df2.modality, hue=df2.metric, palette=palette2, inner="stick", ax=ax2)
ylim = ax2.get_ylim()
for violin in ax2.collections:
    bbox = violin.get_paths()[0].get_extents()
    x0, y0, width, height = bbox.bounds
    violin.set_clip_path(plt.Rectangle((x0 + width / 2, y0), width / 2, height, transform=ax2.transData))
for line in ax2.lines:
    x = line.get_xdata()
    line.set_xdata([np.mean(x), x[1]])
ax2.set_ylabel("Metric2")
ax2.set_xticklabels(ax2.get_xticklabels(), rotation=45, ha='right')
ax2.legend(handles=ax1.legend_.legendHandles + ax2.legend_.legendHandles, title='Metric')
ax1.legend_.remove()

fig.tight_layout()
plt.show()