Deep Clustering

(For ECS 271.) Let's compare the KL Divergence loss propounded by Xie et al. 2016 on to some alternatives, starting with features taken from a pre-trained image classifier instead of the SAE's used by Xie et al.

Main approach

1. Get pretrained Resnet-18

model = primary.net.new()

1. Download STL-10 dataset

python data/stl10_input.py

1. Partition dataset into train and test sets

python -m data.dataset

1. Run trainer to fine-tune Resnet-18

NAME=my_session
python -m primary.train \
	--save_path "saves/$NAME.pth" \
	--log_path "logs/$NAME.log"

SAE approach

1. Prepare dataset

python -m sae.dataset

1. Pretrain SAE

NAME=sae-pretrain
rm "logs/$NAME.log"
python -m sae.pretrain \
	--ep 5000 \
	--lr 1e-1 \
	--test_every 0 \
	--print_every 0 \
	--save_path "saves/$NAME.pth" \
	--log_path "logs/$NAME.log"

1. Fine-tune SAE's encoder

NAME=sae-finetune
python -m sae.finetune \
	--ep 5000 \
	--lr 1e-1 \
	--test_every 0 \
	--print_every 0 \
	--load_path "saves/sae-pretrain.pth" \
	--save_path "saves/$NAME.pth" \
	--log_path "logs/$NAME.log"

TODO

1. Subclass trainer to give a _loss function that doesn't use fine-tuning (KL Div loss) but instead performs reconstruction loss for SAE

DONE

1. Add SAE net, trainer to project
2. Extract HOG features
3. Make dataset for SAE training
4. Show ACC in trainer
5. Hyperparam search
6. Split dataset
7. Write dataset module
8. Write ACC computation
9. Write trainer to fine tune resnet-18 feature extractor