SPLIT

Original implementation of Separated Paths for Local and Global Information framework (SPLIT) in TensorFlow 2.

An Explicit Local and Global Representation Disentanglement Framework with Applications in Deep Clustering and Unsupervised Object Detection.

Rujikorn Charakorn, Yuttapong Thawornwattana, Sirawaj Itthipuripat, Poramate Manoonpong, and Nat Dilokthanakul

---

Installation

Tested on Ubuntu 18.04 and Linux Mint 19.2 with Python 3.6

pip install -r requirements.txt

Experiments

All results will be in output/ folder.

---

SPLIT-VAE

Generation (Fig. 4)

• SVHN

cd vae
python main.py --beta 40 --patch_size 1

• CelebA

cd vae
python main.py --beta 120 --patch_size 8 --dataset celeba64 -no_label

Style transfer (Fig. 5) and reconstruction accuracy table (Table 1)

• SVHN

cd vae
python main.py --beta 1 --patch_size 1

• CelebA

cd vae
python main.py --beta 30 --patch_size 8 --dataset celeba64 -no_label

---

GMVAE and SPLIT-GMVAE

Unsupervised clustering (Table 2)

• SVHN

cd vae
python main.py --model lggmvae --beta 40 --alpha 40 --y_size 30 --patch_size 4 --dataset svhn --training_steps 3000000

Cluster generation (Fig. 6) and Unseen data clustering visualisation (Fig. 7)

• SVHN

cd vae
python main.py --model lggmvae --beta 40 --alpha 40 --y_size 30 --patch_size 4 --dataset svhn --training_steps 3000000 -viz

• CelebA

cd vae
python main.py --model lggmvae --beta 120 --alpha 40 --y_size 30 --patch_size 8 --dataset celeba64 -no_label -viz --training_steps 3000000

---

SPAIR and SPLIT-SPAIR

Fig. 8 and 9

• Multi-Bird-Easy

GMVAE

cd spair
python main.py --dataset cub_solid_fixed --z_bg_beta 10 --latent_size 64 --bg_latent_size 4 --model bg_spair -dense_bg --training_steps 200000

SPLIT-VAE

cd spair
python main.py --dataset cub_solid_fixed --z_bg_beta 10 --patch_size 8 --latent_size 64 --bg_latent_size 4 --local_latent_size 4 --model lg_spair -split_z_l -concat_z_what -dense_local -dense_bg --training_steps 200000

• Multi-Bird-Hard

GMVAE

cd spair
python main.py --dataset cub_ckb_rot_6 --z_bg_beta 1 --latent_size 64 --bg_latent_size 64 --model bg_spair -dense_bg --training_steps 200000

SPLIT-VAE

cd spair
python main.py --dataset cub_ckb_rot_6 --z_bg_beta 1 --patch_size 8 --latent_size 64 --bg_latent_size 64 --local_latent_size 64 --model lg_spair -split_z_l --z_what_beta 0.5 -concat_z_what -dense_local -dense_bg --training_steps 200000