Configuration Settings

This page documents all the default configuration values available in the SNN2 framework. These configurations are organized into multiple INI files located in the SNN2/default_ini/ directory.

Overview

The SNN2 framework uses a hierarchical configuration system with parameter interpolation. Configuration values can reference other values using the {parameter_name} syntax, allowing for flexible and reusable configurations.

Core Configuration (default.ini)

Basic folder structure and timing settings:

Folder Structure:

  • pkg_folder: Path to the package folder

  • main_folder: Main working directory

  • uxtime_sec: Unix timestamp in seconds for unique run identification

Model Configuration (model.ini)

Neural network model parameters:

Model Architecture:

  • weights_file: Path to pre-trained model weights

  • checkpoint_file: Path for saving model checkpoints

  • model: Model architecture specification

  • loss: Loss function configuration

  • metrics: List of metrics to track during training

  • optimizer: Optimizer settings

Training Parameters:

  • epochs: Number of training epochs

  • batch_size: Training batch size

  • learning_rate: Learning rate for optimization

  • validation_split: Fraction of data for validation

Input/Output Configuration (io.ini)

Paths and file management settings:

Output Paths:

  • results_path: Directory for storing results

  • checkpoints_path: Directory for model checkpoints

  • logs_path: Directory for training logs

  • tensorboard_path: TensorBoard logging directory

  • csv_net_evolution_file: CSV file for network evolution tracking

Data Paths:

  • dataset: Primary dataset path

  • test_dataset: Test dataset path

  • validation_dataset: Validation dataset path

Reinforcement Learning Configuration (reinforcement_model.ini)

Reinforcement learning specific parameters:

RL Algorithm Settings:

  • gamma: Discount factor (default: 0.99)

  • entropy_beta: Entropy regularization coefficient (default: 0.01)

  • learning_rate_rl: Learning rate for RL components

  • replay_buffer_size: Size of experience replay buffer

Environment Settings:

  • action_space_size: Number of possible actions

  • state_space_size: Dimensionality of state representation

  • reward_threshold: Threshold for episode success

  • max_episode_length: Maximum steps per episode

Data Flow Configuration (flow.ini)

Data processing and flow management:

Flow Types:

  • defaultFlow: Basic data flow configuration

  • NetCatFlow: Network traffic data flow

  • VMAFFlow: Video quality assessment flow

  • MVNOFlow: Mobile network operator data flow

Flow Parameters:

  • NetCat_columns: Column specification for NetCat data

  • NetCat_dataset: Dataset path for NetCat flow

  • VMAF_columns: Column specification for VMAF data

  • VMAF_dataset: Dataset path for VMAF flow

  • vmaf_threshold: Quality threshold for VMAF processing

  • window: Window size for data processing

Extended Flows:

  • MVNOFlow_inference: Inference-specific MVNO flow

  • MVNOFlow_smartMemory: Memory-optimized MVNO flow

  • MVNOFlow_extended: Extended MVNO flow with additional features

Flow Control:

  • inference_dst_name: Destination name for inference results

  • extra_dataset_base_name: Base name for additional datasets

  • use_random_samples: Flag for random sampling

  • gray_post_train_portion: Portion of data for post-training

  • gray_in_train_portion: Portion of data within training

Callbacks Configuration (callbacks.ini)

Training callbacks and monitoring:

Logging Callbacks:

  • csvLogger: CSV logging for training metrics

  • tensorBoard: TensorBoard visualization

  • saveEmbeddings: Save learned embeddings

  • saveObject: Save model objects

Training Control:
  • earlyStopping: Early stopping configuration

    • patience: 20 epochs

    • min_delta: 0.001 minimum improvement

  • modelCheckpoint: Model checkpoint saving

Testing and Validation:
  • testParam: Testing parameter callback

    • testParamEpochThreshold: 10 epochs

    • testParamEndTrainFlag: False

    • testParamDelta: 10.0

  • activate_metric: Metric activation

  • set_metric_attr: Metric attribute setting

Reinforcement Learning Callbacks:
  • reinforcement: Main RL callback

    • reinforceCBThresholdUnit: ‘ep’ (epochs)

    • reinforceCBThreshold: 10

    • reinforceCBInitialTrainingEpochs: 1

  • FastReplayMemGenerator: Fast replay memory generation

  • FakeReinforceFixedAction: Fixed action RL testing

    • fixedAction: 0

  • RL_manager: RL environment manager

    • RL_initial_skip: 0

  • RL_partial_manager: Partial RL environment manager

Accuracy Callbacks:
  • ott_accuracy: Over-the-top accuracy measurement

    • ottAcc_fxdM_flag: True

    • ottAcc_fxdM: 0.5

  • mno_accuracy: Mobile network operator accuracy

    • mnoAcc_fxdM_flag: True

    • mnoAcc_fxdM: 0.5

Advanced Callbacks:

  • controlledMargin: Margin control callback

  • controlledCrossEntropy: Cross-entropy control

    • categorical_threshold: 1

    • categorical_value: -1.0

    • categorical_triangular_flag: True

    • categorical_dario_norm: False

  • ACCCE: Advanced Controlled Cross Entropy

Additional Configuration Files

The SNN2 framework includes additional configuration files for specific components:

  • actions.ini: Action space definitions for reinforcement learning

  • optimizers.ini: Optimizer configurations (Adam, SGD, RMSprop, etc.)

  • losses.ini: Loss function definitions

  • metrics.ini: Custom metrics and evaluation functions

  • datasets.ini: Dataset loading and preprocessing configurations

  • preprocessors.ini: Data preprocessing pipeline configurations

  • postprocessors.ini: Post-processing and result formatting

  • rewards.ini: Reward function definitions for RL

  • cure_algorithm.ini: CURE clustering algorithm parameters

  • models.ini: Additional model architecture definitions

Configuration Usage

To use these configurations in your experiments:

  1. Default Values: All parameters have sensible defaults defined in the INI files

  2. Override Values: You can override any parameter by specifying it in your experiment configuration

  3. Parameter Interpolation: Use {parameter_name} to reference other configuration values

  4. Environment Variables: Some parameters can be set via environment variables

Example configuration override:

[model]
epochs = 100
batch_size = 64
learning_rate = 0.001

[reinforcement_model]
gamma = 0.95
entropy_beta = 0.02

For more details on specific parameters and their effects, refer to the individual configuration files in the SNN2/default_ini/ directory.