0

I am trying to create a color map for my 10x10 confusion matrix that is provided by sklearn. I would like to be able to customize the color map to be normalized between [0,1] but I have had no success. I am trying to use ax_ and matplotlib.colors.Normalize but am struggling to get something to work since ConfusionMatrixDisplay is a sklearn object that creates a different than usual matplotlib plot.

My code is the following:

from sklearn.metrics import ConfusionMatrixDisplay, confusion_matrix

train_confuse_matrix = confusion_matrix(y_true = ytrain, y_pred = y_train_pred_labels)
print(train_confuse_matrix)
cm_display = ConfusionMatrixDisplay(train_confuse_matrix, display_labels = ['airplane', 'automobile', 'bird', 'cat', 'deer', 'dog', 'frog', 'horse', 'ship', 'truck'])
print(cm_display)
cm_display.plot(cmap = 'Greens')
plt.show()
plt.clf()


[[3289   56   84   18   55    7   83   61   48  252]
 [   2 3733    0    1    2    1   16    1    3  220]
 [  81   15 3365   64   81   64  273   18    6   17]
 [  17   37   71 3015  127  223  414   44    6   64]
 [   3    1   43   27 3659   24  225   35    0    3]
 [   5   23   38  334  138 3109  224   80    4   25]
 [   3    1   19   10   12    7 3946    1    1    5]
 [   4    7   38   69  154   53   89 3615    2   27]
 [  62   67   12    7   25    3   62    4 3595  153]
 [   2   30    1    2    4    0   15    2    0 3957]]

enter image description here

Coldchain9
  • 1,373
  • 11
  • 31
  • 2
    What exactly are you expecting? – Quang Hoang Oct 27 '20 at 16:54
  • Maybe `from matplotlib.colors import LogNorm` and something like `cm_display.plot(cmap='Greens', norm=LogNorm(vmin=0.9, vmax=max(cm_display)))`? – JohanC Oct 27 '20 at 17:16
  • @QuangHoang A normalized colormap scale from 0 to 1. Such that the values don't range from 0 to 3500+ on the color scale. I would like to represent a proportion of my values and their color heat. e.g. closer to 1 would be that the classifier predicted a high amount of values in that particular cell of the confusion matrix. – Coldchain9 Oct 27 '20 at 17:58
  • All color maps work that way. My question is more like how do you present the confusion matrix values, e.g 3289 in your case? – Quang Hoang Oct 27 '20 at 17:59
  • @QuangHoang 3289 airplanes were predicted to be airplanes out of 3953 total airplanes. This is about 83% accuracy for airplanes so the color scale should be ~ .83 between [0,1]. – Coldchain9 Oct 27 '20 at 18:04

2 Answers2

1

Let's try imshow and annotate manually:

accuracies = conf_mat/conf_mat.sum(1)
fig, ax = plt.subplots(figsize=(10,8))
cb = ax.imshow(accuracies, cmap='Greens')
plt.xticks(range(len(classes)), classes,rotation=90)
plt.yticks(range(len(classes)), classes)

for i in range(len(classes)):
    for j in range(len(classes)):
        color='green' if accuracies[i,j] < 0.5 else 'white'
        ax.annotate(f'{conf_mat[i,j]}', (i,j), 
                    color=color, va='center', ha='center')

plt.colorbar(cb, ax=ax)
plt.show()

Output:

enter image description here

Quang Hoang
  • 146,074
  • 10
  • 56
  • 74
0

I would comment on above great answer by @quang-hoang, but do not have enough reputation. The annotation position needs to be swappeed to (j,i) since the output from imshow is transposed.

Code:

classes = ['A','B','C']
accuracies = np.random.random((3,3))

fig, ax = plt.subplots(figsize=(10,8))
cb = ax.imshow(accuracies, cmap='Greens')
plt.xticks(range(len(classes)), classes,rotation=90)
plt.yticks(range(len(classes)), classes)

for i in range(len(classes)):
    for j in range(len(classes)):
        color='green' if accuracies[i,j] < 0.5 else 'white'
        ax.annotate(f'{accuracies[i,j]:.2f}', (j,i), 
                    color=color, va='center', ha='center')

plt.colorbar(cb, ax=ax)
plt.show()

Output

user17069135
  • 1
  • 2