How to change size of confusion matrix in fastai?

user_1603480 0
Viewed 608

I am drawing a Confusion Matrix in fastai with following code:

interp = ClassificationInterpretation.from_learner(learn)
interp.plot_confusion_matrix()

But I end up with a super small matrix because I have around 20 categories:

small confusion matrix

I have found the related question for sklearns but don't know how to apply it to fastai (because we don't use pyplot directly.

confusion-matrix fast-ai data-visualization machine-learning python
1 Answers

If you check the code of the function ClassificationInterpretation.plot_confusion_matrix (in file fastai / interpret.py), this is what you see:

    def plot_confusion_matrix(self, normalize=False, title='Confusion matrix', cmap="Blues", norm_dec=2,
                              plot_txt=True, **kwargs):
        "Plot the confusion matrix, with `title` and using `cmap`."
        # This function is mainly copied from the sklearn docs
        cm = self.confusion_matrix()
        if normalize: cm = cm.astype('float') / cm.sum(axis=1)[:, np.newaxis]
        fig = plt.figure(**kwargs)
        plt.imshow(cm, interpolation='nearest', cmap=cmap)
        plt.title(title)
        tick_marks = np.arange(len(self.vocab))
        plt.xticks(tick_marks, self.vocab, rotation=90)
        plt.yticks(tick_marks, self.vocab, rotation=0)

The key here is the line fig = plt.figure(**kwargs), so basically, the function plot_confusion_matrix will propagate its parameters to the plot function.

So you could use either one of these:

  • dpi=xxx (e.g. dpi=200)
  • figsize=(xxx, yyy)

See this post on StackOverflow about the relations they have with each other: https://stackoverflow.com/a/47639545/1603480

So in your case, you could just do:

interp.plot_confusion_matrix(figsize=(12, 12))

And the Confusion Matrix would look like:

big confusion matrix

FYI: this also applies to other plot functions, like

interp.plot_top_losses(20, nrows=5, figsize=(25, 25))
user_16034800
Related