Context-Aware Doubly-Robust Semi-Supervised Learning

Project structure

• config/: hydra hyperparameter configuration files. All hyperparameters can be directly modified in the CLI using hydra's override syntax
• datasets/: dataset data files
• src/: main codebase; implements the components necessary to semi-supervised training, such as :
  • data/: data loaders
  • models/: trainable model classes
  • pretrained/: pre-trained teacher models
  • train/: training components:
    • bias_estimate_schedule.py: curriculum learning schedules; specifies how teacher bias estimates are introduced in semi-supervised losses
    • criterion.py: loss functions and metrics
    • loss.py: Doubly-Robust (DR) and Context-aware DR (CDR) semi-supervised training losses
    • tuning.py: tuned DR (TDR) and CDR optimal tuning parameters computation
• study/: codebase containing the experiments presented in the paper
  • protocols/: grid-search protocols:
    • commands/: list of commands per protocol
    • protocol_confs/: user-specified protocol configurations
    • slurm/: execute protocol commands via Slurm batch scripts (requires Slurm install)
  • train_model_cdr.py: TDR/CDR training
  • train_model_dr.py: DR training
  • train_model_supervised_erm.py: Supervised Empirical Risk Minimization (ERM) training
  • train_model_pseudo_erm.py: Pseudo-ERM (P-ERM) training
  • plots_toy_example.py: create toy-example experiment plots (see Reproduce Experiments)
  • plots_beamforming.py: create beamforming experiment plots (see Reproduce Experiments)
  • create_protocol.py: create custom grid-search protocols (see (see Extend experiments)
• logs/: experimental data is stored in this folder by default. Can be changed by setting the keyword argument logs_dir=<DIR> to a custom directory <DIR> when executing experiments
• scripts/generate_beamforming_dataset.py: script used to generate the beamforming dataset from the ray-tracing data in the "Environment Aware Communications" repository. The ray-tracing data raw.zip file must be unzipped before executing this script.

Reproduce Experiments

Setup Python Environment

• Download and install miniconda
• Install conda environment by running the command:
  conda env create -f ./environment.yaml
• Source the environment by running:
  conda activate cdr

Toy Example

• Execute toy-example protocols directly as:

bash ./study/protocols/commands/toy_example_supervised_erm.sh
bash ./study/protocols/commands/toy_example_pseudo_erm.sh
bash ./study/protocols/commands/toy_example_dr.sh
bash ./study/protocols/commands/toy_example_cdr.sh

• OR send protocols commands to Slurm cluster via:

sbatch ./study/protocols/slurm/toy_example_supervised.sh
sbatch ./study/protocols/slurm/toy_example_pseudo_erm.sh
sbatch ./study/protocols/slurm/toy_example_dr.sh
sbatch ./study/protocols/slurm/toy_example_cdr.sh

• Generate toy-example plots in ./logs/figures as:

python ./study/plots_toy_example.py +experiment=toy_example logs_subdir=toy_example

Beamforming

• Execute beamforming protocols directly as:

bash ./study/protocols/commands/beamforming_supervised_erm.sh
bash ./study/protocols/commands/beamforming_pseudo_erm.sh
bash ./study/protocols/commands/beamforming_dr.sh
bash ./study/protocols/commands/beamforming_tdr.sh
bash ./study/protocols/commands/beamforming_cdr.sh

• OR send protocols commands to Slurm cluster via:

sbatch ./study/protocols/slurm/toy_example_supervised.sh
sbatch ./study/protocols/slurm/toy_example_pseudo_erm.sh
sbatch ./study/protocols/slurm/toy_example_dr.sh
sbatch ./study/protocols/slurm/toy_example_tdr.sh
sbatch ./study/protocols/slurm/toy_example_cdr.sh

• Generate beamforming plots in ./logs/figures as:

python ./study/plots_beamforming.py +experiment=beamforming logs_subdir=beamforming