ninachang1107 / dsc_gan

Deep Subspace Clustering + GAN

Deep Subspace Clustering + GAN

Based on:

Pan Ji*, Tong Zhang*, Hongdong Li, Mathieu Salzmann, Ian Reid. Deep Subspace Clustering Networks. in NIPS'17.

Forked off: https://github.com/panji1990/Deep-subspace-clustering-networks

How to run:

CUDA_VISIBLE_DEVICES=0 python dsc_gan.py exp_name --lambda3 1 --epochs 3000

Dependencies:

pip install tensorflow-gpu==1.1
pip install tensorboard

tensorboard useful for visualization only.

Notes on parameter tuning

Training result seems to be highly dependent on quality of pre-trained model. For a particular pretrained model, finetuning achieves very consistent result. But if we repeat pre-training everytime, final finetuning result has significant variance.

TODO

[ ] Add COIL10 and COIL100 [ ] Allow tuning of number of conv-layers and disc size [ ] Nxr rxN

coil20 experiments

Without GAN

• coil20_1: pretrain 10000, epochs 1000, enableat 1000

• coil20_2: pretrain 1, epochs 1000, enableat 1000

• coil20_3: pretrain 10000, epochs 1000, enableat 1000, lambda4 100

• coil20_4: pretrain 10000, epochs 1000, enableat 1000, lambda1 100

• coil20_5: pretrain 10000, epochs 1000, enableat 1000, lambda1 10000

• coil20_6: pretrain 10000, epochs 1000, enableat 1000, lambda1 100000

• coil20_7: pretrain 10000, epochs 1000, enableat 1000, lambda1 1000000

• coil_m1: pretrain 3500, epochs 500, enableat1000, lambda2 20, alpha 0.12

• coil_m2: pretrain 3500, epochs 500, enableat1000, lambda2 20, alpha 0.12

• coil_m3: pretrain 3500, epochs 500, enableat1000, lambda2 20, alpha 0.12

• coil_m4: pretrain 3500, epochs 500, enableat1000, lambda2 20, alpha 0.12, lr0.0004

• coil_m5: pretrain 6000, epochs 700, enableat1000, lambda2 20, alpha 0.12, lr0.0004

• coil_m6: pretrain 6000, epochs 1000, enableat2000, lambda2 20, alpha 0.12, lr0.0004

• coil_m7: pretrain 6000, epochs 1000, enableat2000, lambda2 20, alpha 0.12, lr0.0002

• coil_m8: pretrain 5000, epochs 1000, enableat2000, lambda2 20, alpha 0.12, lr0.0002

• coil_m3raw: pretrain 3500, epochs 500, enableat1000, lambda2 20, alpha 0.12, COIL20RRRaw

• coil_m3dim: pretrain 3500, epochs 500, enableat1000, lambda2 20, alpha 0.12, COIL20RRdimension

• coil_k1: pretrain 3500, epochs 500, enable-at 1000, lambda2 20, alpha 0.12, kernel_size 5

ORL exps

submean, one2one, no-uni-norm all run with lr=0.001 orl_000: orl_001: orl_010: orl_100:

all run with lr=0.0002 orl_000_2: orl_001_2: orl_010_2: orl_100_2:

all run with lr=0.001, but enable-at 200 orl_000_3: orl_001_3: orl_010_3: orl_100_3:

all run with lr=0.001, but enable-at 100 orl_000_4: orl_001_4: orl_010_4: orl_100_4:

all run with lr=0.001, but enable-at 51 orl_000_5: orl_001_5: orl_010_5: orl_100_5:

Without uniform normalization, training seem somewhat unstable. There seems to be a bug in submean

So the best so far seems to be orl_010_4

orl_010_4_a0.05 to orl_010_4_a0.11 search for alpha. Turns out the default value 0f 0.1 was best