Chapter 20: Classification

• See also
No related topics
Availability
Not available in MIL-Lite


Available in MIL
Previous
Next

This chapter explains how to use the MIL classification module.

• MIL Classification module
• Steps to performing classification
• • Basic work-flow
• Basic concepts for the MIL Classification module
• • Classifiers
  • Datasets
  • Training
  • Predicting
• Classifiers and how they work
• • CNN
  • • Image classification
  • Tree ensemble
  • • Feature classification
• Datasets
• • Steps to build the datasets
  • Proper data collection, classes, and entries
  • • Data foresight
  • Classes and labeling (the ground truth)
  • • Organized data
  • Splitting the source dataset
  • Training dataset
  • • Training dataset for a CNN
    • Training dataset for a tree ensemble
  • Development dataset
  • Testing dataset
  • An example of distributing data among datasets
  • Importing data from a CSV file to a dataset context
  • • General CSV file format
    • Headers for authors
    • Headers for class definitions
    • Headers for entries
  • UUID
  • • MIL_UUID utility macros for C users
• Training: in general
• • Steps to train
  • Training fundamentals
• Training: CNN
• • Predefined CNN classifiers to train
  • • FCNet - small
    • FCNet - medium
    • FCNet - extra large
    • Selecting a predefined CNN
    • Input image sizes
  • Training modes
  • • Complete
    • Transfer learning
    • Fine tuning
    • Summary and comparison
  • Training mode controls
  • • Learning rate
    • Maximum number of epochs
    • Mini-batch size
    • Schedule type
    • Comparing default training mode controls
• Training: tree ensemble
• • Feature importance
  • • Modes
• Training: analyze and adjust
• • Results
  • Recognizing a properly trained classifier
  • Expected trends and fluctuations
  • Examination
  • Adjustments
  • • Bias and variance analysis
    • High training dataset error
    • High development dataset error
  • Beware of useless pursuits
• Predicting
• • Steps to predict
  • Timeout and stop
  • Results
  • • Drawing
  • Assisted labeling
• Advanced techniques
• • Coarse segmentation
  • • Performing coarse segmentation
    • Training images
    • Coarse segmentation results (classification map)
    • Drawing coarse segmentation results
  • Augmentation
  • Advanced analysis of your training
  • • Confusion matrix
    • Score distribution
    • Improving a deployed network
  • Shuffling
  • Extracting
  • Preprocessing
  • Pitfalls
  • • Data quality
    • Data distribution
    • Unbalanced data
• Requirements, recommendations, and troubleshooting
• • Computer requirements for a predefined CNN classifier
  • Troubleshooting for a predefined CNN classifier
  • • GPU, CUDA, and CuDNN installation
• Classification examples