wudihjq / ood-ka

Code for paper "OOD-KA: Amalgamating Knowledge in the Wild"

OOD-KA

Code for paper "OOD-KA: Amalgamating Knowledge in the Wild"

Requirements

To install requirements:

pip install -r requirements.txt

Reproduce our results

Image Classification

1. Split CIFAR100

cd engine/datasets
python split_cifar100.py

2. Train Teachers

python train_scratch.py --lr 0.1 --batch-size 128 --model wrn16_2 --dataset cifar100_part0 --gpu 0
python train_scratch.py --lr 0.1 --batch-size 128 --model wrn16_2 --dataset cifar100_part1 --gpu 0

3. OOD-KA

python OOD_KA.py \
--model wrn16_2 \
--unlabeled cifar10 \
--lr 1e-3 \
--lr_g 1e-3 \
--z_dim 100 \
--oh 1.0 \
--bn 1.0 \
--local 1.0 \
--adv 1.0 \
--sim 1.0 \
--balance 10.0

Semantic Segmentation

1. Train Teachers with DeepLabV3

python train_deeplab.py --model deeplabv3plus_mobilenet --dataset voc --output_stride 16 --lr 0.01 --batch_size 16
python train_deeplab.py --model deeplabv3plus_mobilenet --dataset nyu --output_stride 16 --lr 0.01 --batch_size 16

1. OOD-KA for semantic segmentation

python PFA_deeplab.py \
--model deeplabv3plus_mobilenet \
--output_stride 16 \
--batch_size 16 \
--lr 1e-2 \
--lr_g 1e-3 \
--z_dim 256 \
--oh 1.0 \
--bn 0.5 \
--local 0.5 \
--adv 1.0 \
--sim 1.0 \
--balance 1.0