A Pytorch and Lightning based framework for research and ml pipeline automation.

1. $\text{Hyperparameter space}.$
2. $\text{Genetic algorithms(single-objective/multi-objective)}$
3. $\text{Best hyperparameters in config.yaml}$
4. $\text{Training session}$

from lightorch.htuning.optuna import htuning
from ... import NormalModule
from ... import FourierVAE

def objective(trial) -> Dict[str, float]:
    ... # define hyperparameters
    return hyperparameters

if __name__ == '__main__':
    htuning(
        model_class = FourierVAE,
        hparam_objective = objective,
        datamodule = NormalModule,
        valid_metrics = [f"Training/{name}" for name in [
            "Pixel",
            "Perceptual",
            "Style",
            "Total variance",
            "KL Divergence"]],
        directions = ['minimize', 'minimize', 'minimize', 'minimize', 'minimize'],
        precision = 'medium',
        n_trials = 150,
    )

exec: python3 -m htuning

trainer: # trainer arguments
  logger: true 
  enable_checkpointing: true
  max_epochs: 250
  accelerator: cuda
  devices:  1
  precision: 32
  
model:
  class_path: utils.FourierVAE #model relative path
  dict_kwargs: #**hparams
    encoder_lr: 2e-2
    encoder_wd: 0
    decoder_lr: 1e-2
    decoder_wd: 0
    alpha:
      - 0.02
      - 0.003
      - 0.003
      - 0.01
    beta: 0.00001
    optimizer: adam

data: # Dataset arguments
  class_path: data.DataModule
  init_args:
    type_dataset: mnist 
    batch_size: 12
    pin_memory: true
    num_workers: 8

from lightorch.training.cli import trainer

if __name__ == '__main__':
    trainer()

exec: python3 -m training -c config.yaml

• Built in Module class for:
  • Adversarial training.
  • Supervised, Self-supervised training.
• Multi-Objective and Single-Objective optimization and Hyperparameter tuning with optuna.

• Fourier Convolution.
• Fourier Deconvolution.
• Partial Convolution. (Optimized implementation)
• Grouped Query Attention, Multi Query Attention, Multi Head Attention. (Interpretative usage) (with flash-attention option)
• Self Attention, Cross Attention.
• Normalization methods.
• Positional encoding methods.
• Embedding methods.
• Useful criterions.
• Useful utilities.
• Built-in Default Feed Forward Networks.
• Adaptation for $\mathbb{C}$ modules.
• Interpretative Deep Neural Networks.
• Monte Carlo forward methods.

• Linkedin
• GitHub
• Email: jorged.encyso@gmail.com

@misc{lightorch,
  author = {Jorge Enciso},
  title = {LighTorch: Automated Deep Learning framework for researchers},
  howpublished = {\url{https://github.com/Jorgedavyd/LighTorch}},
  year = {2024}
}