Contributing to Olympus¶

Adding new Basic Blocks¶

Models, Optimizers, LRSchedules, Datasets all use factories. To insert them you simply need to create a new file inside their respective modeule. olympus/models/.. for Models and register the model constructor.

Models¶

Create a new olympus/models/<my_model>.py

See Custom Model

import torch.nn as nn

class MyCustomModel(nn.Module):
    def __init__(self, input_size, output_size):
        self.main = nn.Linear(input_size[0], output_size[0])

    def forward(self, x):
        return self.main(x)

# Register my model
builders = {'my_model': MyCustomModel}

Model Optimizer¶

Create a new olympus/optimizers/<my_optimizer>.py

See Custom Optimizer

import torch.optim as optim

class MyCustomOptimizer(optim.Optimizer):
    pass

# Register my Optimizer
builders = {'my_optimizer': MyCustomOptimizer}

Weight Initialization¶

Create a new olympus/models/inits/<my_init>.py

Schedule¶

See Custom Schedule

Create a new olympus/optimizers/schedules/<my_optimizer>.py

Tasks¶

Task describe generic setup like classification

Create a new olympus/tasks/<my_task>.py

Baselines¶

Baselines are the top level scripts used to run a given tasks

Create a new olympus/baselines/<my_baseline>.py

Datasets¶

Create a new olympus/datasets/<my_baseline>.py

Dataset Sampling¶

Create a new olympus/datasets/sampling/<my_sampler>.py

Metrics¶

Create a new olympus/metrics/<my_metric>.py

Observers¶

Create a new olympus/observers/<my_observer>.py

Hyper-parameter Optimizer¶

Create a new olympus/hpo/<my_hpo>.py

Specifying hyper-parameters¶

To add new hyper parameters you simply need to override the static method get_space()

See Custom Model with NAS