Awesome
iCarl2.0
This is an on-going pytorch implementation of iCarl[1].
Interface to run experiments
runExperiment.py [-h] [--batch-size N] [--test-batch-size N]
[--epochs N] [--lr LR]
[--schedule SCHEDULE [SCHEDULE ...]]
[--gammas GAMMAS [GAMMAS ...]] [--momentum M]
[--no-cuda] [--no-distill] [--no-random]
[--no-herding] [--oversampling] [--seed S]
[--log-interval N] [--model-type MODEL_TYPE]
[--name NAME] [--sortby SORTBY] [--decay DECAY]
[--step-size STEP_SIZE]
[--memory-budget MEMORY_BUDGET]
[--epochs-class EPOCHS_CLASS] [--classes CLASSES]
[--depth DEPTH] [--dataset DATASET]
Dependencies
- Pytorch 0.3.0.post4
- Python 3.6
- torchnet (https://github.com/pytorch/tnt)
Ideas to implement and test
- Auto-encoder to preserve information of a class
- Generative models for data seen in the past to remove dependency on examplers. =======