AIBox

My AI toolbox of helper functions and other assorted utils

Configuration Folder Info

• Model configurations go into models folder
• Experiment configurations go into experiments folder
• Configuration folders can have subdirectories
  • for example: a model at the path models/cnn/classifier1.toml has the model configuration name cnn/classifier1
• any default.toml in a folder or subfolder will be loaded first before the specified configuration name
  • the specified config will override any parameters specified with the same key path from the default config

Refer to TOML spec for help on syntax: TOML.io

Project Structure

├── configs
│ ├── experiments
│ │  ├── vae.toml
│ │  ├── default.toml
│ │  ├── mnist.toml
│ ├── models
│ │  ├── ae/vae.toml
│ │  ├── pix2pix.toml
│ │  └── default.toml
├── src
│  └── project_package
├── pyproject.toml
└── README.md

• configs folder can be anywhere, can override via CLI args
• can specify multiple configurations and OmegaConf merges them

Examples

Models

[model]
_target_ = "project.models.CNN" # required
in_channels = 3

[[model.optimizers]] # makes an array of objects at the models.args.optimizers path
_target_ = 'torch.optim.Adam'
lr = 1e-3

Experiment

[data]
_target_ = "torch"

[data.args]
# ...

[loss]

[trainer]

[tuner] # optional params for ray tune

[tuner.values]
param.values = [0, 1, 28]
param.search_space = 'choice' # see raytune for search_space names