Learning_From_Mistakes

Learning From Mistakes: A Multi-level Optimization Framework (Official Pytorch implementation for applications on Neural Architecture Search and Data Reweighting).

Setups

The requiring environment is as bellow:

• Linux
• Python 3+
• PyTorch
• Torchvision

Or, you can use the following command to build the environment and get started:

conda env create -f environment.yml

Running application to NAS on benchmark datasets (CIFAR-10 and CIFAR-100).

Here is an example about running the architecture search stage of DARTS on CIFAR-10:

python train_search_lfm.py --is_cifar100 0 --gpu 0 --unrolled --save darts-cifar10

Here is an example about running the evaluation stage of architecture searched on CIFAR-10:

python train.py --gpu 0 --auxiliary --cutout --arch [searched architecture]

Running application to DR on benchmark datasets (CIFAR-10 and CIFAR-100).

Here is an example about running the experiment on class imbalance dataset with 100 imbalance factor

python dr-lfm-imbalance.py --dataset cifar100 --num_classes 100 --imb_factor 0.01

Checkpoints that related to the results showed in the paper

Checkpoints of the Application to NAS:

• Evaluation checkpoints for DARTS on CIFAR-10
• Evaluation checkpoints for P-DARTS on CIFAR-10
• Evaluation checkpoints on ImageNet with architecture search by DARTS on CIFAR-10
• Evaluation checkpoints on ImageNet with architecture search by P-DARTS on CIFAR-10

Checkpoints of the Application to DR (Class Imbalance):

• Checkpoints for experiment on CIFAR-10 with 10 Imbalance Factor
• Checkpoints for experiment on CIFAR-10 with 20 Imbalance Factor
• Checkpoints for experiment on CIFAR-10 with 50 Imbalance Factor
• Checkpoints for experiment on CIFAR-10 with 100 Imbalance Factor
• Checkpoints for experiment on CIFAR-10 with 200 Imbalance Factor
• Checkpoints for experiment on CIFAR-100 with 10 Imbalance Factor
• Checkpoints for experiment on CIFAR-100 with 20 Imbalance Factor
• Checkpoints for experiment on CIFAR-100 with 50 Imbalance Factor
• Checkpoints for experiment on CIFAR-100 with 100 Imbalance Factor
• Checkpoints for experiment on CIFAR-100 with 200 Imbalance Factor

Checkpoints of the Application to DR (Label Noisy):

• Checkpoints for experiment on CIFAR-10 with 0.2 Flip Noisy
• Checkpoints for experiment on CIFAR-10 with 0.4 Flip Noisy
• Checkpoints for experiment on CIFAR-100 with 0.2 Flip Noisy
• Checkpoints for experiment on CIFAR-100 with 0.4 Flip Noisy

License

This work is licensed under MIT license. See the LICENSE for details.

Acknowledgement

We appreciate the developers of Differentiable Architecture Search and Meta-Weight-Net, and we express our gratitude to these awesome projects.