- Refactored code
- Added dependencies in environment.txt (but don't use it for creating new env it contains dependencies carried over from a last project)
- Added dropout layers + config and weight decay.
- Added a script to generate commands multiple configs
- Removed wandb loggers and added custom logger
- Added class imbalanced training
- Added Curriculum learning with pretrained ResNet18
python train.py --batchsize 1028 --workers 2 --epochs 80 --opt adam --arch 'mlp[16384,16384,512]' --reg wd --iid
See exps.sh more details.
Code for the paper The Deep Bootstrap Framework: Good Online Learners are Good Offline Generalizers.
The main training code is here, and a sample configuration of hyperparameter sweep (using Caliban) is here.
The CIFAR-5m dataset is released at: https://github.com/preetum/cifar5m