Deep Learning in Computer Vision Final Project

The objective was to compare the performance of Dirichlet output layers and Temperature Scaling in the calibration of Deeplab v3+ on CamVid dataset.

Training a model:

• Type in Command Line: python train_model.py [type_of_training] [L2 reg] [dropout rate] [loss] [weights (optional)]
  • [type_of_training] must be 'data_augmentation' or 'finetuning'
  • [loss] must be 'dirichlet' or 'crossentropy'
  • [weights] must be the name of the weight file and be in Weights folder

Calibration Analysis of a model:

• Type in Command Line: python calibration_analysis.py [type_of_model] [desired_T] [weights]
  • [type_of_model] must be 'd' for dirichlet model or 't' for baseline model with temperature scaling
  • [desired_T] must be a number bigger than zero (this option is only available for baseline model)
  • [weights] must be the name of the weight file and be in Weights folder

Temperature Calculation for a model:

• Type in Command Line: python temp_calculation.py [loss] [weights]
  • [loss] must be 'nll'(Negative Log Likelihood), 'mce' (Maximum Calibration Error) or 'ece'(Expected Calibration Error)
  • [weights] must be the name of the weight file and be in Weights folder