Skip to main content

​
ExperimentConfig

Dataclass containing all experiment parameters for edge AI hardware optimization. Includes training settings, hardware constraints, and optimization configurations.

​
Fields

​
seed
int
required
Random seed for reproducibility across all random number generators.
​
dataset
str
required
Dataset name. Supported values: "mnist", "fashion-mnist".
​
batch_size
int
required
Batch size for training and validation data loaders.
​
epochs
int
required
Number of training epochs.
​
learning_rate
float
required
Learning rate for the optimizer.
​
train_subset
int | None
required
Maximum number of training samples to use. None uses the full training set.
​
val_subset
int | None
required
Maximum number of validation samples to use. None uses the full validation set.
​
power_watts
float
required
Power budget constraint in watts for edge device simulation.
​
pruning_levels
list[float]
required
List of pruning ratios to evaluate (e.g., [0.0, 0.3, 0.5] for 0%, 30%, 50% pruning).
​
precisions
list[str]
required
List of numeric precisions to evaluate (e.g., ["fp32", "fp16", "int8"]).
​
calibration_batches
int
required
Number of batches to use for quantization calibration.
​
output_dir
str
required
Directory path for saving experiment results and artifacts.
​
memory_budgets_mb
list[float]
required
List of memory budget constraints in megabytes to evaluate.
​
active_memory_budget_mb
float
required
Active memory budget in megabytes for runtime constraints.
​
cpu_frequency_scale
float
required
CPU frequency scaling factor (e.g., 0.8 for 80% of max frequency).
​
dataloader_seed
int
required
Seed for data loader shuffling. Defaults to seed if not specified in config file.
​
num_workers
int
required
Number of worker processes for data loading. Defaults to 2 if not specified.
​
benchmark_repeats
int
required
Number of times to repeat benchmarks for averaging. Defaults to 5 if not specified.
​
memory_bandwidth_gbps
float
required
Memory bandwidth in GB/s for hardware simulation. Defaults to 12.8 if not specified.

​
load_config

Loads and parses experiment configuration from a text file format. 
def load_config(path: str | Path) -> ExperimentConfig
​
path
str | Path
required
Path to the configuration file. Accepts string paths or pathlib.Path objects.
​
config
ExperimentConfig
Parsed configuration object with all experiment parameters.

​
Configuration File Format

The configuration file uses a simple key: value format:
  • Lines starting with # are comments
  • Empty lines are ignored
  • Each line contains a key-value pair separated by :
  • Lists use bracket notation: [item1, item2, item3]
  • Supports automatic type inference for strings, integers, floats, booleans, null values, and lists

​
Supported Value Types

  • Integers: 42, 100
  • Floats: 0.001, 3.14
  • Strings: mnist, "output/results"
  • Booleans: true, false
  • Null: null, none
  • Lists: [0.0, 0.3, 0.5], [fp32, fp16, int8]
Some fields have default values if not specified: dataloader_seed defaults to seed, num_workers defaults to 2, benchmark_repeats defaults to 5, and memory_bandwidth_gbps defaults to 12.8.

​
Usage Example

from edge_opt.config import load_config
from pathlib import Path

# Load from string path
config = load_config("experiments/baseline.conf")

# Or use pathlib
config_path = Path("experiments/baseline.conf")
config = load_config(config_path)

# Access fields
print(f"Dataset: {config.dataset}")
print(f"Batch size: {config.batch_size}")
print(f"Pruning levels: {config.pruning_levels}")

​
Error Handling

The function will raise exceptions for:
  • FileNotFoundError: Config file doesn’t exist
  • KeyError: Required field missing from config file
  • ValueError: Invalid value format or type conversion failure

​
See Also

  • Model - Use config.seed with set_deterministic()
  • Data - Pass config fields to build_loaders()

Build docs developers (and LLMs) love

Get started for freeTalk to us