Generative Invariance Transfer [ICLR 2022]

Env setup

This will create a Conda environment called "tail."

conda env create -f environment.yml

Kuzushiji-49-LT

Data preparation

Download the following files four files into ./data/raw_kmnist_data (originally source is here):

File	Examples	Download (NumPy format)
Training images	232,365	k49-train-imgs.npz (63MB)
Training labels	232,365	k49-train-labels.npz (200KB)
Testing images	38,547	k49-test-imgs.npz (11MB)
Testing labels	38,547	k49-test-labels.npz (50KB)

Now to make long-tailed versions with rotation, background variation, and dilation/erosion run:

python -m core.make_kmnist_lt

This should produce a folder ./data/proc_kmnist_data with 30 long-tailed datasets for each transformation.

Train generative model, then classifiers.

There is a bash script for automatically training the GIT generative models on each of the 30 instances of K49-DIL-LT and K49-BG-LT, then training classifiers with or without GIT.

# Train generative models (MUNIT)
./scripts/train_k49_munits.sh
# Train classifiers
./scripts/train_k49.sh

The MUNIT model produces checkpoints in ./outputs, which may be used for training classifiers. Classifier checkpoints will be in ./outputs/baseline.

CIFAR-LT

Train generative model

This is based on MUNIT code. For CIFAR10 and CIFAR100, respectively:

python -m core.train_munit --config core/munit/configs/cifar10_0.01.yaml
python -m core.train_munit --config core/munit/configs/cifar100_0.01.yaml

Train classifier

Example training command:

# Cross entropy loss with Delayed ReSampling
python -m core.train_baselines --config core/ldram_drw/configs/cifar10_0.01.yaml --loss_type CE --train_rule DRS_Simple
# Cross entropy loss with Delayed ReSampling with GIT using MUNIT model.
python -m core.train_baselines --config core/ldram_drw/configs/cifar10_0.01.yaml --loss_type CE --train_rule DRS_Simple --use_munit

Here are common options for the arguments:

--config:

• core/ldam_drw/configs/cifar10_0.01.yaml: (CIFAR10-LT)
• core/ldam_drw/configs/cifar100_0.01.yaml: (CIFAR100-LT)

--loss_type:

• CE: Cross-entropy loss
• LDAM: LDAM loss

--train_rule:

• None: Standard ERM training
• DRS_Simple: Delayed ReSampling

--use_munit: Include this to use the GIT generative model as augmentation during training. Check that the paths for munit_ckpt and munit_config are correct in the --config file.

Credits

The generative model code is based off of MUNIT. Classifier training code is based off of LDAM-DRW.