fastai_helpers -- Fastai utility functions

Key Files

stats.py

• StatsRepo()
  • .add(), .show_results(), .restore(), .save()

train.py

• _get_learner()
  • higher level function to create learner, inclusing cupping of pre-trained model
  • model_cutter(): used for custom layer clipping of pre-trained model
• lr_find()
  • lrfinder + plots
  • Learner.Recorder.plot2(): returns smoothed plot with suggestions
• _train()
  • supports both one-cycle and flat annealing
• Learner.svalidate()
  • silently validate -- compute validation results without updating Jupyter cell output
• get_best_stats()
  • Returns best result from recorder history
• get_val_stats()
  • Returns final result from recorder history
• Thresholded metrics for multi-label classification:
  • error_02(), accuracy_02(), error_05(), accuracy_05()
• mishify()
  • modifies resnet and xresnet models to use MISH activation function in place of RELU.
  • note: Includes version of MISH with application (mish source: https://github.com/lessw2020/mish)

interpret.py

• analyze_interp()
  • Summary results for nterpreter including plots
• combine_predictions()
  • Generates composite prediction based on multiple interpreters
• plot_confusion_matrix()
  • plots confusion matrix with clean-up of distplay boundaries
• plot_confusion_matrix_thresh()
  • plots confusion matrix after thresholding values
  • threshold_confusion_matrix(): computes confusion matric using threshold
• interpretation_summary()
  • Outputs various stats (summary, cm, threshold analysis)
• get_accuracy()
  • Computes accuract stat for interpreter
• analyze_confidence()
  • Plots confidence histograms
• accuracy_vs_threshold()
  • Plots accuract vs threshold value
• show_incremental_accuracy()
  • TBS
• analyze_low_confidence()
  • TBS

other_helpers.py

• verify_gpu()
  • checks GPU status
• reset_seeds()
  • initialized all random seeds to known value

Files requiring reorganization

helpers.py

• _get_interp()
  • ISSUE: Code may be too implementation specific