Implementation:Fastai Fastbook ClassificationInterpretation

Knowledge Sources	fastbook fastai docs ClassificationInterpretation
Domains	Computer_Vision, Deep_Learning, Model_Evaluation
Last Updated	2026-02-09 17:00 GMT

Overview

Concrete tools for analyzing classification model performance and cleaning training data provided by fastai.interpret.ClassificationInterpretation and fastai.vision.widgets.ImageClassifierCleaner.

Description

ClassificationInterpretation computes predictions on the validation set and provides methods to visualize the confusion matrix, display the highest-loss images, and list the most confused category pairs. ImageClassifierCleaner is an interactive Jupyter widget that lets practitioners review and relabel or delete problematic images directly in the notebook.

Usage

Create a ClassificationInterpretation object after training is complete. Use its visualization methods to identify error patterns, then optionally launch the ImageClassifierCleaner widget to fix data quality issues before retraining.

Code Reference

Source Location

Repository: fastbook
File: translations/cn/02_production.md (lines 419-454), translations/cn/05_pet_breeds.md (lines 572-586)

Signature

# Create interpretation object from a trained Learner
ClassificationInterpretation.from_learner(learn)

# Plot the confusion matrix
ClassificationInterpretation.plot_confusion_matrix(
    normalize=False,     # Show raw counts or normalized percentages
    title='Confusion matrix',
    cmap='Blues',        # Colormap
    norm_dec=2,          # Decimal places for normalized values
    figsize=None,
    **kwargs
)

# Plot the images with highest losses
ClassificationInterpretation.plot_top_losses(
    k=9,                 # Number of top-loss images to show
    largest=True,        # If True, show highest losses; if False, lowest
    figsize=None,
    **kwargs
)

# Return list of most confused category pairs
ClassificationInterpretation.most_confused(
    min_val=1            # Minimum number of confusions to include
)

# Interactive data cleaner widget
ImageClassifierCleaner(learn)

Import

from fastai.vision.all import ClassificationInterpretation
from fastai.vision.widgets import ImageClassifierCleaner

I/O Contract

Inputs

Name	Type	Required	Description
learn	Learner	Yes	A trained Learner whose validation set will be evaluated
k	int	No	Number of top-loss images to display (default: 9)
largest	bool	No	Show highest losses if True, lowest if False (default: True)
normalize	bool	No	Normalize confusion matrix rows to percentages (default: False)
min_val	int	No	Minimum confusion count to include in most_confused output (default: 1)

Outputs

Name	Type	Description
interp	ClassificationInterpretation	Object holding all validation predictions, losses, and decoded outputs
confusion matrix	matplotlib figure	Heatmap showing predicted vs. actual class counts
top losses	matplotlib figure	Grid of images labeled with predicted/actual/loss/probability
most_confused	list of tuples	List of (actual, predicted, count) triples sorted by count descending
cleaner widget	ImageClassifierCleaner	Interactive Jupyter widget for reviewing and relabeling images

Usage Examples

Basic Usage: Full Interpretation Pipeline

from fastai.vision.all import *

# Assume 'learn' is a trained Learner
interp = ClassificationInterpretation.from_learner(learn)

# 1. Plot confusion matrix
interp.plot_confusion_matrix(figsize=(12, 12))

# 2. Show the 9 images with highest loss
interp.plot_top_losses(k=9, nrows=3)

# 3. List the most confused category pairs
most_confused = interp.most_confused(min_val=5)
print('Most confused pairs:')
for actual, predicted, count in most_confused:
    print(f'  {actual} -> {predicted}: {count} confusions')

Data Cleaning with ImageClassifierCleaner

from fastai.vision.widgets import ImageClassifierCleaner

# Launch interactive cleaner widget
cleaner = ImageClassifierCleaner(learn)
cleaner

# After reviewing images in the widget, apply changes:
# Delete images marked for deletion
for idx in cleaner.delete():
    cleaner.fns[idx].unlink()

# Move images marked for category change
for idx, cat in cleaner.change():
    shutil.move(str(cleaner.fns[idx]), path / cat)

Normalized Confusion Matrix

# Normalized view shows percentages instead of counts
# Useful when class sizes are imbalanced
interp.plot_confusion_matrix(normalize=True, figsize=(12, 12))

Examining Top Losses in Detail

# Get the raw top losses data
losses, idxs = interp.top_losses()

# Print details for the top 5
for i in range(5):
    idx = idxs[i]
    print(f'Loss: {losses[i]:.4f}')
    print(f'  Predicted: {interp.decoded[idx]}')
    print(f'  Actual: {interp.targs[idx]}')
    print(f'  File: {interp.dl.items[idx]}')
    print()

Related Pages

Implements Principle

Principle:Fastai_Fastbook_Model_Interpretation

Requires Environment

Page Connections

Double-click a node to navigate. Hold to expand connections.

Principle

Implementation

Heuristic

Environment