JIDECA

Jointly Improved Deep Embedded Clustering for Android activity

Data processing files

Download Link and extract it.

$ unzip data_processing.zip

Create train and test dataset

(We already upload dataset with a batch_size of 2400.. Please, check code.)

If necessary, change num_workers (line 28 and 40) and batch_size (line 50) in data_processing.py. The default values are 8 and 2400, respectively.

$ cd jideca
$ python data_processing.py

output

data/train_re.pkl
data/train_se.pkl
data/test_re_23.pkl
data/test_re_34.pkl
data/test_se_23.pkl
data/test_se_34.pkl

Train

If necessary, change num_workers (line 137) and batch_size (line 67) in train.py. The default values are 64 and 2400, respectively. Set the batch_size value to the same value used in train and test dataset.

option:
-t: data types [real, semmantic_annotations]
-c: number of classes (default: 34)
-e: number of epochs (default: 5000)
-gid: GPU id (default: 0)
-r: resume
-js: use JS Divergence (use alignment loss)
-beta: hyperparameter for reconstruction term
-gamma: hyperpapameter for alignment term
-l : hyperparameter for lambda

$ cat train.sh
## JIDECA
#echo 'jideca real training..'
#python train.py -t real -gid 0 -e 2500 -js --off_scheduling

## JIDECA, change beta (reconstruction), gamma (alignment)
#echo 'jideca real training..'
#python train.py -t real -gid 0 -e 2500 -js -beta 1 -gamma 1 --off_scheduling

## JIDECA w/o alingment (== multimodal IDEC), no -js option
#echo 'multimodal idec real training..'
#python train.py -t real -gid 0 -e 2500 -gamma 0 --off_scheduling

## JIDECA w/o reconstruction (== multimodal DEC w/ alignment)
#echo 'multimodal dec w/ alignment real training..'
#python train.py -t real -gid 0 -e 2500 -js -beta 0 --off_scheduling

## JIDECA w/o reconstruction and alignment (== multimodal DEC)
#echo 'multimodal dec real training..'
#python train.py -t real -gid 0 -e 2500 -beta 0 -gamma 0 --off_scheduling

echo 'jideca semantic_annotations training..'
python train.py -t semantic_annotations -gid 0 -e 2500 -js --off_scheduling

## DEC
#echo 'dec re training..'
#python train.py -t real -gid 0 -e 2500 -l 1 -beta 0  --off_scheduling -c 23

#echo 'dec se training..'
#python train.py -t semantic_annotations -gid 0 -e 2500 -l 1 -beta 0  --off_scheduling -c 23

#echo 'dec dnn training..'
#python train.py -t real -gid 0 -e 2500 -l 0 -beta 0  --off_scheduling -c 23

## IDEC
#echo 'idec re training..'
#python train.py -t real -gid 0 -e 2500 -l 1 -beta 1  --off_scheduling -c 23

#echo 'idec se training..'
#python train.py -t semantic_annotations -gid 0 -e 2500 -l 1 -beta 1  --off_scheduling -c 23

#echo 'idec dnn training..'
#python train.py -t real -gid 0 -e 2500 -l 0 -beta 1  --off_scheduling -c 23

Modify train.sh to suit your experiment and run it.

. ./train.sh

output (example)

cd ..
cd clustering
result/idec_se_34_conv_dnn_jsd/

Evaluation

cd ../evaluation

Please, change proper number and names in the test_evaluation.py.

103     run("idec_se_34_conv_dnn_jsd", "34")

Then,

$ . ./test_evaluation.sh

Evaluation (including purity)

Same process as Evaluation, but use test_evaluation_purity.py and test_evaluation_purity.sh.

For repeat test

Comment out the resulting part in jideca/train.py

427 #    trainer.result(img_train_loader,
428 #                   img_test_dataset,
429 #                   text_train_loader,
430 #                   w2v_test_data,
431 #                   text_test_names)

Then, use (modify and execute) result.sh in jideca directory.

For evaluation, use test_evaluation_iter.sh or test_evaluation_purity_iter.sh in evaluation directory.