config.py

from src.utils import Container, DataName, ActivType, LossType
import torch

"""
DATASETS 
DIGITS  : 1797  |
ENERGY  : 768   | 8->h->2
GRID    : 10000 | 12->h->1
HOUSE   : 19794 | 8->h->1
IRIS    : 150   |
MNIST   : 10000 |
"""

cfg = Container() # struct-like class to contain run configurations  
cfg.online = True # W&B local or global

# Run Parameters

cfg.seed = (42, 15, 26)
cfg.dev = torch.device("cpu")
cfg.save_weights = False

# Dataset/Task Parameters

cfg.experiment_name = "final" 
# the additional strings will be added to the experiment name:
# <dataset_name>[<hidden_sizes>]<activation_type>_<learning:l/r>
#cfg.dataset_name = DataName.ENERGY

# Network Parameters

cfg.batch_size = 100
cfg.lr = 0.005

cfg.learn = (True, False)
cfg.rand_bound = 1
cfg.rand_step_size= 0.1

#cfg.maxiter = if commented, gets the maximum: SIZE // batch_size

cfg.n_epochs = 100
cfg.per_epoch_test = 10

cfg.activ_type = (ActivType.RELU, ActivType.SIGMOID, ActivType.GELU)
#cfg.loss_type = automatically defined by dataset_name / task_type

#cfg.input_size = automatically defined by src.utils.default_config 
cfg.hidden_size = (128, [128, 16], [128, 64, 16]) 
#cfg.output_size = automatically defined by src.utils.default_config 

# Scheduler Parameters for Learning via SGD

cfg.sch__use = False

if cfg.sch__use:
    cfg.sch__gamma = 0.9
    cfg.sch__step_size = 20

# if given as a dictionary in the format below
# dataset is generated randomly 

dataset1 = {
    'input_size'    : 10,   # input dimension
    'output_size'   : 1,    # output dimension
    'sample_size'   : 1000, # number of samples
    'model_par_std' : 0.1,  # standard deviation of the optimum model parameters
    'noise_std'     : 0.02  # standard deviation of the noise induced to the input dataset
    }

dataset2 = {
    'input_size'    : 20,   # input dimension
    'output_size'   : 2,    # output dimension
    'sample_size'   : 2000, # number of samples
    'model_par_std' : 0.2,  # standard deviation of the optimum model parameters
    'noise_std'     : 0.05  # standard deviation of the noise induced to the input dataset
    }

cfg.dataset_name = DataName.GRID
cfg.start_from = 25