mdlearn.utils

Configurations and utilities for model building and training.

Functions

`get_torch_optimizer`(name, hparams, parameters)	Construct a PyTorch optimizer specified by `name` and `hparams`.
`get_torch_scheduler`(name, hparams, optimizer)	Construct a PyTorch lr_scheduler specified by `name` and `hparams`.
`log_checkpoint`(checkpoint_file, epoch, ...)	Write a torch .pt file containing the epoch, model, optimizer, and scheduler.
`parse_args`()	Parse command line arguments using argparse library
`resume_checkpoint`(checkpoint_file, model, ...)	Modifies `model`, `optimizer`, and `scheduler` with values stored in torch .pt file `checkpoint_file` to resume from a previous training checkpoint.

pydantic settings mdlearn.utils.BaseSettings

Show JSON schema

{
   "title": "BaseSettings",
   "description": "Base class for settings, allowing values to be overridden by environment variables.\n\nThis is useful in production for secrets you do not wish to save in code, it plays nicely with docker(-compose),\nHeroku and any 12 factor app design.",
   "type": "object",
   "properties": {},
   "additionalProperties": false
}

dump_yaml(cfg_path: Union[str, pathlib.Path])

classmethod from_yaml(filename: Union[str, pathlib.Path]) → mdlearn.utils._T

pydantic settings mdlearn.utils.OptimizerConfig

pydantic schema for PyTorch optimizer which allows for arbitrary optimizer hyperparameters.

Show JSON schema

{
   "title": "OptimizerConfig",
   "description": "pydantic schema for PyTorch optimizer which allows\nfor arbitrary optimizer hyperparameters.",
   "type": "object",
   "properties": {
      "name": {
         "title": "Name",
         "default": "Adam",
         "env_names": "{'name'}",
         "type": "string"
      },
      "hparams": {
         "title": "Hparams",
         "default": {},
         "env_names": "{'hparams'}",
         "type": "object"
      }
   }
}

Config

extra: str = allow

Fields

hparams (Dict[str, Any])
name (str)

field hparams: Dict[str, Any] = {}

field name: str = 'Adam'

pydantic settings mdlearn.utils.SchedulerConfig

pydantic schema for PyTorch scheduler which allows for arbitrary scheduler hyperparameters.

Show JSON schema

{
   "title": "SchedulerConfig",
   "description": "pydantic schema for PyTorch scheduler which allows for arbitrary\nscheduler hyperparameters.",
   "type": "object",
   "properties": {
      "name": {
         "title": "Name",
         "default": "ReduceLROnPlateau",
         "env_names": "{'name'}",
         "type": "string"
      },
      "hparams": {
         "title": "Hparams",
         "default": {},
         "env_names": "{'hparams'}",
         "type": "object"
      }
   }
}

Config

extra: str = allow

Fields

hparams (Dict[str, Any])
name (str)

field hparams: Dict[str, Any] = {}

field name: str = 'ReduceLROnPlateau'

pydantic settings mdlearn.utils.WandbConfig

Show JSON schema

{
   "title": "WandbConfig",
   "description": "Base class for settings, allowing values to be overridden by environment variables.\n\nThis is useful in production for secrets you do not wish to save in code, it plays nicely with docker(-compose),\nHeroku and any 12 factor app design.",
   "type": "object",
   "properties": {
      "wandb_project_name": {
         "title": "Wandb Project Name",
         "env_names": "{'wandb_project_name'}",
         "type": "string"
      },
      "wandb_entity_name": {
         "title": "Wandb Entity Name",
         "env_names": "{'wandb_entity_name'}",
         "type": "string"
      },
      "model_tag": {
         "title": "Model Tag",
         "env_names": "{'model_tag'}",
         "type": "string"
      }
   },
   "additionalProperties": false
}

Fields

model_tag (Optional[str])
wandb_entity_name (Optional[str])
wandb_project_name (Optional[str])

field model_tag: Optional[str] = None

field wandb_entity_name: Optional[str] = None

field wandb_project_name: Optional[str] = None

init(cfg: mdlearn.utils.BaseSettings, model: torch.nn.Module, wandb_path: Union[str, pathlib.Path]) → Optional[wandb.config]

Initialize wandb with model and config.

Parameters

cfg (BaseSettings) – Model configuration with hyperparameters and training settings.
model (torch.nn.Module) – Model to train, passed to wandb.watch(model) for logging.
wandb_path (PathLike) – Path to write wandb/ directory containing training logs.

Returns

Optional[wandb.config] – wandb config object or None if wandb_project_name is None.

mdlearn.utils.get_torch_optimizer(name: str, hparams: Dict[str, Any], parameters) → torch.optim.Optimizer: Construct a PyTorch optimizer specified by name and hparams.

mdlearn.utils.get_torch_scheduler(name: Optional[str], hparams: Dict[str, Any], optimizer: torch.optim.Optimizer) → Optional[torch.optim.lr_scheduler._LRScheduler]

Construct a PyTorch lr_scheduler specified by name and hparams.

Parameters

name (Optional[str]) – Name of PyTorch lr_scheduler class to use. If name is None, simply return None.
hparams (Dict[str, Any]) – Hyperparameters to pass to the lr_scheduler.
optimizer (torch.optim.Optimizer) – The initialized optimizer.

Returns

Optional[torch.optim.lr_scheduler._LRScheduler] – The initialized PyTorch scheduler, or None if name is None.

mdlearn.utils.log_checkpoint(checkpoint_file: Union[str, pathlib.Path], epoch: int, model: torch.nn.Module, optimizers: Dict[str, torch.optim.Optimizer], scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None)

Write a torch .pt file containing the epoch, model, optimizer, and scheduler.

Parameters

checkpoint_file (PathLike) – Path to save checkpoint file.
epoch (int) – The current training epoch.
model (torch.nn.Module) – The model whose parameters are saved.
optimizers (Dict[str, torch.optim.Optimizer]) – The optimizers whose parameters are saved.
scheduler (Optional[torch.optim.lr_scheduler._LRScheduler]) – Optional scheduler whose parameters are saved.

mdlearn.utils.parse_args() → argparse.Namespace

Parse command line arguments using argparse library

Returns: argparse.Namespace – Dict like object containing a path to a YAML file accessed via the config property.

Example

>>> from mdlearn.utils import parse_args
>>> args = parse_args()
>>> # MyConfig should inherit from BaseSettings
>>> cfg = MyConfig.from_yaml(args.config)

mdlearn.utils.resume_checkpoint(checkpoint_file: Union[str, pathlib.Path], model: torch.nn.Module, optimizers: Dict[str, torch.optim.Optimizer], scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None) → int

Modifies model, optimizer, and scheduler with values stored in torch .pt file checkpoint_file to resume from a previous training checkpoint.

Parameters

checkpoint_file (PathLike) – Path to checkpoint file to resume from.
model (torch.nn.Module) – Module to update the parameters of.
optimizers (Dict[str, torch.optim.Optimizer]) – Optimizers to update.
scheduler (Optional[torch.optim.lr_scheduler._LRScheduler]) – Optional scheduler to update.

Returns

int – The epoch the checkpoint is saved plus one i.e. the current training epoch to start from.