On Contrastive Representations of Stochastic Processes

This project is based on PyTorch, Hydra and PyTorch Lightning.

Project organisation

config/ -> Hydra project configurations

src/ -> Everything that relates to models

utils/ -> Helper functions

experiments/ -> Where experiments are saved (logs, checkpoints, configs, etc)

Install

virtualenv -p python3.8 venv
source venv/bin/activate
pip install -r requirements.txt

Experiments

Sinusoids

• Run the different models with:

python main.py -m +experiment=sine model=ssl,cnp self_attn=id,on  # Untargeted CReSP, FCLR, ANCP, CNP

• Figure 2b results can be obtained by varying the distance between the modes:

python main.py -m +experiment=sine model=cnp,ssl self_attn=on dataset.eps=0.0,0.5,1.,2.,5.,8.,10

• Figure 2c results can be obtained by varying the number of training views:

python main.py -m +experiment=sine model=cnp,ssl self_attn=on n_views_train=2,5,10,20,50

ShapeNet

• Download and decompress the r2n2 dataset:

wget http://cvgl.stanford.edu/data2/ShapeNetRendering.tgz -P data
tar zxvf data/ShapeNetRendering.tgz -C data

• Run the different models with:

python main.py +experiment=shapenet model=ssl targeted=True  # Targeted CReSP
python main.py -m +experiment=shapenet model=ssl,cnp self_attn=id,on  # Untargeted CReSP, FCLR, ANCP, CNP
python main.py +experiment=shapenet model=sup fix_clf_train=True  # Supervised

• Figure 4a and 4b results can be obtained by varying the colour distortion strength:

python main.py -m +experiment=nocolour self_attn=on,id targeted=False,True model=ssl,cnp
python main.py -m +experiment=shapenet self_attn=on,id targeted=False,True model=ssl,cnp dataset.distortion_s=0.5,1.0,1.5

• Figure 4c results can be obtained by varying the number of training views:

python main.py -m +experiment=shapenet self_attn=on targeted=False self_attn=on,id model=ssl n_views_train=6,12,24

• Figure 5a can be obtained by varying the fraction of labels available:

python main.py -m +experiment=shapenet targeted=False model=ssl,cnp clf.prop=0.01,0.02,0.04,0.1,0.2,0.4,1.0

• Figure 5b can be obtained by varying the number of test views:

python main.py -m +experiment=shapenet targeted=False model=ssl,cnp n_views_test=1,2,4,10,20

Snooker

• Table 3 results can be obtained with:

python main.py +experiment=snooker model=cnp  # CNP
python main.py +experiment=snooker targeted=false  # Untargeted CReSP
python main.py +experiment=snooker  # Targeted CReSP
python main.py +experiment=snooker agg=kernel enc=simple  # MetaCDE

Miscellaneous

Logging

Can select the logger with logger=tensorboard, by default it's using logger=csv.

• Tensorboard: execute the following and forward the port to access live logs:

tensorboard --logdir experiments/

• CSV: The metrics additionally saved under .../NAME_OF_EXPERIMENT/RUN_ID/logs/metrics.csv