JP-Amboage / qtml-hybrid-workflow

Hybrid quantum-classical hyperparameter optimization workflow using performance prediction.

HYBRID WORKFLOW REPO for QTML

• File structure
• How to add a new target model
• Main files args
• Other relevant repositories

File structure

• MPI_regular_hyperband.py: main script for running regular hyperband.
• MPI_swift_hyperband.py: main script for running swift-hyperband.
• utils_MPI.py: common funcions of MPI_regular_hyperband.py and MPI_swift_hyperband.py.
• template_methods.py: template for the model specific methods.
• <whatever model>_methods.py: contains the specific ConfigGenerator class and needed methods for <whatever model>.
• start_script_<whatever model>_regular.sh: batch script to run regular hyperband for <whatever model>.
• start_script_<whatever model>_swift.sh: batch script to run swift hyperband for <whatever model>.
• predictor.py: classical performance predictor.
• qpredictor.py: quantum performance predictor.
• QSVR_extended.py: quantum SVR implementation. Needed by qpredictor.py.
• embeddings.py: file to save and re-use embeddings for the annealer. Needed by qpredictor.py.
• utility.py: Needed by QSVR_extended.py.
• create_env-dwave.sh: create the required environment to run the code.

How to add a new target model: MyModel

1. Create the file mymodel_methods.py.
2. Code the class ConfigGeneratorMyModel inside the file mymodel_methods.py following the declaration in template_methods.py. Taking a look at ConfigGeneratorLSTM in lstm_methods.py may also be useful.
3. Code the method train_myModel inside the file mymodel_methods.py following the declaration in template_methods.py. Taking a look at train_lstm in lstm_methods.py may also be useful.
4. Code the additional methods that you may need for saving/loading/building the model and loading the data.
5. Edit MPI_regular_hyperband.py and MPI_swift_hyperband.py between the comments that state CHANGE THIS TO ADD MORE MODELS. More precisely, add a new case in the if else structure as follows:

if model_name == 'some_model':
    ...
elif model_name == 'some_other_model':
    ...
elif model_name == 'mymodel':
    from mymodel_methods import ConfigGeneratorMyModel
    ConfigGenerator_class = ConfigGeneratorMyModel
    from mymodel_methods import train_myModel
    train_model_method = train_myModel
else:
    ...

6. Write the batch files start_script_mymodel_regular.sh and start_script_mymodel_swift.sh to run the algorithms for your model.

Main files args

• MPI_regular_hyperband.py accepts the following flags:

  • --dir_name: (default saved_models_regular) name of the folder where models are saved during training. Must be empty before starting the script.
  • --seed: (default 0) random seed for generating configurations.
  • model_name: (default lstm) name of the target model.
  • save_models: (default False) keep the trained models in dir_name once the script is finished.
  • --r: R parameter from hyperband algorithm, max epoch that any model is trained.
  • --eta: (default 2 the less agressive possible value) eta parameter from hyperband algorithm, controls the discarding rate.

• MPI_swift_hyperband.py accepts the following flags:

  • --dir_name: (default saved_models_swift) name of the folder where models are saved during training. Must be empty before starting the script.
  • --seed: (default 0) random seed for generating configurations.
  • model_name: (default lstm) name of the target model.
  • save_models: (default False) keep the trained models in dir_name once the script is finished.
  • --r: R parameter from hyperband algorithm, max epoch that any model is trained.
  • --eta: (default 2 the less agressive possible value) eta parameter from hyperband algorithm, controls the discarding rate.
  • --d: (default 15) minimum training size for the performance predictors.
  • --k: (default 0.5) proportion of models to train, taken from fast-hyperband.
  • --known_curve: (default 0.5) position of the performance prediction decision points inside each round.
  • --pred_type: (default quantum) quantum | classical.

Other relevant repositories

• Sequential modular implementation of Swift-Hyperband that also allows for simulated runs (not public).
• Distributed implementation of Swift-Hyperband for Rusty and Swift-ASHA draft (not public).