clana

clana是一个库和命令行应用程序，用于可视化具有大量类别的分类器的混淆矩阵。clana的两个主要贡献是，如卷积神经网络架构的分析和优化的第5章中解释的混淆矩阵排序（CMO），以及实现它的优化算法。CMO技术可以应用于任何多类分类器，有助于了解哪些类别的组最相似。

安装

安装clana的推荐方法是

$ pip install clana --user --upgrade

如果您想安装最新版本

$ git clone https://github.com/MartinThoma/clana.git; cd clana
$ pip install -e . --user

使用

$ clana --help
Usage: clana [OPTIONS] COMMAND [ARGS]...

  Clana is a toolkit for classifier analysis.

  See https://arxiv.org/abs/1707.09725, Chapter 4.

Options:
  --version  Show the version and exit.
  --help     Show this message and exit.

Commands:
  distribution   Get the distribution of classes in a dataset.
  get-cm         Generate a confusion matrix from predictions and ground...
  get-cm-simple  Generate a confusion matrix.
  visualize      Optimize and visualize a confusion matrix.

可视化命令会给出这样的图像

MNIST示例

$ cd docs/
$ python mnist_example.py  # creates `train-pred.csv` and `test-pred.csv`
$ clana get-cm --gt gt-train.csv  --predictions train-pred.csv --n 10
2019-09-14 09:47:30,655 - root - INFO - cm was written to 'cm.json'
$ clana visualize --cm cm.json --zero_diagonal
Score: 13475
2019-09-14 09:49:41,593 - root - INFO - n=10
2019-09-14 09:49:41,593 - root - INFO - ## Starting Score: 13475.00
2019-09-14 09:49:41,594 - root - INFO - Current: 13060.00 (best: 13060.00, hot_prob_thresh=100.0000%, step=0, swap=False)
[...]
2019-09-14 09:49:41,606 - root - INFO - Current: 9339.00 (best: 9339.00, hot_prob_thresh=100.0000%, step=238, swap=False)
Score: 9339
Perm: [0, 6, 5, 8, 3, 2, 1, 7, 9, 4]
2019-09-14 09:49:41,639 - root - INFO - Classes: [0, 6, 5, 8, 3, 2, 1, 7, 9, 4]
Accuracy: 93.99%
2019-09-14 09:49:41,725 - root - INFO - Save figure at '/home/moose/confusion_matrix.tmp.pdf'
2019-09-14 09:49:41,876 - root - INFO - Found threshold for local connection: 398
2019-09-14 09:49:41,876 - root - INFO - Found 9 clusters
2019-09-14 09:49:41,877 - root - INFO - silhouette_score=-0.012313948323292875
    1: [0]
    1: [6]
    1: [5]
    1: [8]
    1: [3]
    1: [2]
    1: [1]
    2: [7, 9]
    1: [4]

这将给出

标签操作

准备一个具有强制的表头的labels.csv文件

$ clana visualize --cm cm.json --zero_diagonal --labels mnist/labels.csv

数据分布

$ clana distribution --gt gt.csv --labels labels.csv [--out out/] [--long]

每个标签打印一行，例如：

60% cat (56789 elements)
20% dog (12345 elements)
 5% mouse (1337 elements)
 1% tux (314 elements)

如果指定了--out，则创建一个水平条形图。第一个条形是最常见的类别，第二个条形是第二常见的类别，...

它使用简短标签，除非在命令中添加了--long。

可视化

查看可视化

作为库的使用

>>> import numpy as np
>>> arr = np.array([[9, 4, 7, 3, 8, 5, 2, 8, 7, 6],
                    [4, 9, 2, 8, 5, 8, 7, 3, 6, 7],
                    [7, 2, 9, 1, 6, 3, 0, 8, 5, 4],
                    [3, 8, 1, 9, 4, 7, 8, 2, 5, 6],
                    [8, 5, 6, 4, 9, 6, 3, 7, 8, 7],
                    [5, 8, 3, 7, 6, 9, 6, 4, 7, 8],
                    [2, 7, 0, 8, 3, 6, 9, 1, 4, 5],
                    [8, 3, 8, 2, 7, 4, 1, 9, 6, 5],
                    [7, 6, 5, 5, 8, 7, 4, 6, 9, 8],
                    [6, 7, 4, 6, 7, 8, 5, 5, 8, 9]])
>>> from clana.optimize import simulated_annealing
>>> result = simulated_annealing(arr)
>>> result.cm
array([[9, 8, 7, 6, 5, 4, 3, 2, 1, 0],
       [8, 9, 8, 7, 6, 5, 4, 3, 2, 1],
       [7, 8, 9, 8, 7, 6, 5, 4, 3, 2],
       [6, 7, 8, 9, 8, 7, 6, 5, 4, 3],
       [5, 6, 7, 8, 9, 8, 7, 6, 5, 4],
       [4, 5, 6, 7, 8, 9, 8, 7, 6, 5],
       [3, 4, 5, 6, 7, 8, 9, 8, 7, 6],
       [2, 3, 4, 5, 6, 7, 8, 9, 8, 7],
       [1, 2, 3, 4, 5, 6, 7, 8, 9, 8],
       [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]])
>>> result.perm
array([2, 7, 0, 4, 8, 9, 5, 1, 3, 6])

您可以使用result.cm进行可视化，并使用result.perm以相同的顺序获取您的标签。

# Just some example labels
# ['0', '1', '2', '3', '4', '5', '6', '7', '8', '9', '10']
>>> labels = [str(el) for el in range(11)]
>>> np.array(labels)[result.perm]
array(['2', '7', '0', '4', '8', '9', '5', '1', '3', '6'], dtype='<U2')