gflownet.utils.common

Functions

`set_device`(device)	Get `torch` device from device.
`set_float_precision`(precision)	Get `torch` float type from precision.
`set_int_precision`(precision)	Get `torch` integer type from `int` precision.
`torch2np`(x)	Convert a torch tensor to a numpy array.
`download_file_if_not_exists`(path, url)	Download a file from google drive if path doestn't exist.
`resolve_path`(path)	Resolve a path by expanding environment variables, user home directory, and making
`find_latest_checkpoint`(ckpt_dir)	Find the latest checkpoint in the input directory.
`read_hydra_config`([rundir, config_name])
`gflownet_from_config`(config[, env])	Create GFlowNet from a Hydra OmegaConf config.
`load_gflownet_from_rundir`(rundir[, no_wandb, ...])	Load GFlowNet from a run path (directory with a `.hydra` directory inside).
`batch_with_rest`(start, stop, step[, tensor])	Yields batches of indices from start to stop with step size. The last batch may be
`tfloat`(x, device, float_type)	Convert input to a float tensor. If the input is a list of tensors, the tensors
`tlong`(x, device)	Convert input to a long tensor. If the input is a list of tensors, the tensors
`tint`(x, device, int_type)	Convert input to an integer tensor. If the input is a list of tensors, the tensors
`tbool`(x, device)	Convert input to a boolean tensor. If the input is a list of tensors, the tensors
`concat_items`(list_of_items[, indices])	Concatenates a list of items into a single tensor or array.
`extend`(orig, new)	Extends the original list or tensor with the new list or tensor.
`copy`(x)	Makes copy of the input tensor or list.
`bootstrap_samples`(tensor, num_samples)	Bootstraps tensor along the last dimention
`example_documented_function`(arg1, arg2)	Summary line: this function is not used anywhere, it's just an example.
`select_indices`(iterable[, indices])	Select elements form iterable

Module Contents

gflownet.utils.common.set_device(device)[source]

Get torch device from device.

Examples

>>> set_device("cuda")
device(type='cuda')

>>> set_device("cpu")
device(type='cpu')

>>> set_device(torch.device("cuda"))
device(type='cuda')

Parameters:: device (Union[str, torch.device]) – Device.
Returns:: torch.device – torch device.

gflownet.utils.common.set_float_precision(precision)[source]

Get torch float type from precision.

Examples

>>> set_float_precision(32)
torch.float32

>>> set_float_precision(torch.float32)
torch.float32

Parameters:: precision (Union[int, torch.dtype]) – Precision.
Returns:: torch.dtype – torch float type.
Raises:: ValueError – If precision is not one of [16, 32, 64].

gflownet.utils.common.set_int_precision(precision)[source]

Get torch integer type from int precision.

Examples

>>> set_int_precision(32)
torch.int32

>>> set_int_precision(torch.int32)
torch.int32

Parameters:: precision (Union[int, torch.dtype]) – Integer precision.
Returns:: torch.dtype – torch integer type.
Raises:: ValueError – If precision is not one of [16, 32, 64].

gflownet.utils.common.torch2np(x)[source]

Convert a torch tensor to a numpy array.

Parameters:: x (Union[torch.Tensor, np.ndarray, list]) – Data to be converted.
Returns:: np.ndarray – Converted data.

gflownet.utils.common.download_file_if_not_exists(path, url)[source]

Download a file from google drive if path doestn’t exist. url should be in the format: https://drive.google.com/uc?id=FILE_ID

Parameters:

path (str)
url (str)

gflownet.utils.common.resolve_path(path)[source]

Resolve a path by expanding environment variables, user home directory, and making it absolute.

Examples

>>> resolve_path("~/scratch/$SLURM_JOB_ID/data")
Path("/home/user/scratch/12345/data")

Parameters:: path (Union[str, Path]) – Path to be resolved.
Returns:: Path – Resolved path.
Return type:: pathlib.Path

gflownet.utils.common.find_latest_checkpoint(ckpt_dir)[source]

Find the latest checkpoint in the input directory.

If the directory contains a checkpoint file with the name “final”, that checkpoint is returned. Otherwise, the latest checkpoint is returned based on the iteration number set in the file names.

Parameters:: ckpt_dir (Union[str, Path]) – Directory in which to search for the checkpoints.
Returns:: Path – Path to the latest checkpoint.
Raises:: ValueError – If no checkpoint files are found in the input directory.

gflownet.utils.common.read_hydra_config(rundir=None, config_name='config')[source]

gflownet.utils.common.gflownet_from_config(config, env=None)[source]

Create GFlowNet from a Hydra OmegaConf config.

Parameters:

config (DictConfig) – Config.
env (GFlowNetEnv) – Optional environment instance to be used in the initialization.

Returns:

GFN – GFlowNet.

gflownet.utils.common.load_gflownet_from_rundir(rundir, no_wandb=True, print_config=False, device=None, load_last_checkpoint=True, is_resumed=False)[source]

Load GFlowNet from a run path (directory with a .hydra directory inside).

Parameters:

rundir (Union[str, Path]) – Path to the run directory. Must contain a .hydra directory.
no_wandb (bool, optional) – Whether to disable wandb in the GFN init, by default True.
print_config (bool, optional) – Whether to print the loaded config, by default False.
device (str, optional) – Device to which the models should be moved. If None (default), take the device from the loaded config.
load_last_checkpoint (bool, optional) – Whether to load the final models, by default True.
is_resumed (bool, optional) – Whether the GFlowNet is loaded to resume training.

Returns:

Tuple[GFN, DictConfig] – Loaded GFlowNet and the loaded config.

Raises:

ValueError – If no checkpoints are found in the directory.

gflownet.utils.common.batch_with_rest(start, stop, step, tensor=False)[source]

Yields batches of indices from start to stop with step size. The last batch may be smaller than step.

Parameters:

start (int) – Start index
stop (int) – End index (exclusive)
step (int) – Step size
tensor (bool, optional) – Whether to return a torch tensor of indices instead of a numpy array, by default False.

Yields:

Union[np.ndarray, torch.Tensor] – Batch of indices

gflownet.utils.common.tfloat(x, device, float_type)[source]

Convert input to a float tensor. If the input is a list of tensors, the tensors are stacked along the first dimension.

The resulting tensor is moved to the specified device.

Parameters:

x (Union[List[torch.Tensor], torch.Tensor, List[Union[int, float]], Union[int,)
float]] – Input to be converted to a float tensor.
device (torch.device) – Device to which the tensor should be moved.
float_type (torch.dtype) – Float type to which the tensor should be converted.

Returns:

Union[torch.Tensor, List[torch.Tensor]] – Float tensor.

gflownet.utils.common.tlong(x, device)[source]

Convert input to a long tensor. If the input is a list of tensors, the tensors are stacked along the first dimension.

The resulting tensor is moved to the specified device.

Parameters:

x (Union[List[torch.Tensor], torch.Tensor, List[Union[int, float]], Union[int,)
float]] – Input to be converted to a long tensor.
device (torch.device) – Device to which the tensor should be moved.

Returns:

Union[torch.Tensor, List[torch.Tensor]] – Long tensor.

gflownet.utils.common.tint(x, device, int_type)[source]

Convert input to an integer tensor. If the input is a list of tensors, the tensors are stacked along the first dimension.

The resulting tensor is moved to the specified device.

Parameters:

x (Union[List[torch.Tensor], torch.Tensor, List[Union[int, float]], Union[int,)
float]] – Input to be converted to an integer tensor.
device (torch.device) – Device to which the tensor should be moved.
int_type (torch.dtype) – Integer type to which the tensor should be converted.

Returns:

Union[torch.Tensor, List[torch.Tensor]] – Integer tensor.

gflownet.utils.common.tbool(x, device)[source]

Convert input to a boolean tensor. If the input is a list of tensors, the tensors are stacked along the first dimension.

The resulting tensor is moved to the specified device.

Parameters:

x (Union[List[torch.Tensor], torch.Tensor, List[Union[int, float]], Union[int,)
float]] – Input to be converted to a boolean tensor.
device (torch.device) – Device to which the tensor should be moved.

Returns:

Union[torch.Tensor, List[torch.Tensor]] – Boolean tensor.

gflownet.utils.common.concat_items(list_of_items, indices=None)[source]

Concatenates a list of items into a single tensor or array.

Parameters:

list_of_items – List of items to be concatenated, i.e. list of arrays or list of tensors.
indices (Union[List[np.ndarray], List[torch.Tensor]], optional) – Indices to select in the resulting concatenated tensor or array, by default None.

Returns:

Union[np.ndarray, torch.Tensor] – Concatenated tensor or array, with optional selection of indices.

Raises:

NotImplementedError – If the input type is not supported, i.e., not a list of arrays or a list of tensors.

gflownet.utils.common.extend(orig, new)[source]

Extends the original list or tensor with the new list or tensor.

Returns:

Union[List, TensorType[”…”]] – Extended list or tensor.

Raises:

NotImplementedError – If the input type is not supported, i.e., not a list or a tensor.

Parameters:

orig (Union[List, torchtyping.TensorType[...]])
new (Union[List, torchtyping.TensorType[...]])

Return type:

Union[List, torchtyping.TensorType[…]]

gflownet.utils.common.copy(x)[source]

Makes copy of the input tensor or list.

A tensor is cloned and detached from the computational graph.

Parameters:: x (Union[List, TensorType["..."]]) – Input tensor or list to be copied.
Returns:: Union[List, TensorType[”…”]] – Copy of the input tensor or list.

gflownet.utils.common.bootstrap_samples(tensor, num_samples)[source]: Bootstraps tensor along the last dimention returns tensor of the shape [initial_shape, num_samples]

gflownet.utils.common.example_documented_function(arg1, arg2)[source]

Summary line: this function is not used anywhere, it’s just an example.

Extended description of function from the docstrings tutorial Full Example.

Refer to

functions with gflownet.utils.common.set_device()
classes with gflownet.gflownet.GFlowNetAgent
methods with gflownet.envs.base.GFlowNetEnv.get_action_space()
constants with gflownet.envs.base.CMAP

Prepenend with ~ to refer to the name of the object only instead of the full path -> set_device() will display as set_device instead of the full path.

Great maths:

\[\int_0^1 x^2 dx = \frac{1}{3}\]

Important

A docstring with math MUST be a raw Python string (a string prepended with an r: r"raw") to avoid backslashes being treated as escape characters.

Alternatively, you can use double backslashes.

Warning

Display a warning. See Learn by example. (<– this is a cross reference, learn about it here)

Examples

>>> function(1, 'a')
True
>>> function(1, 2)
True

>>> function(1, 1)
Traceback (most recent call last):
    ...

Notes

This block uses $ ... $ for inline maths -> $e^{\frac{x}{2}}$.

Or $$ ... $$ for block math instead of the .. math: directive above.

\[\int_0^1 x^2 dx = \frac{1}{3}\]

Parameters:

arg1 (int) – Description of arg1
arg2 (str) – Description of arg2

Returns:

bool – Description of return value

gflownet.utils.common.select_indices(iterable, indices=None)[source]

Select elements form iterable

Parameters:

iterable (list, tuple, tensor or np.ndarray) – An iterable to select elements from. It can have multiple dimensions and selection is always preformed over the first dimension
indices (list, tuple, tensor or np.ndarray) – 1-dimentional sequence of indecies for selecting elements, optional. If None, the iterable will be returned as is. Default None

Returns:

list, tuple, tensor or np.ndarray – A sequence of selected elements. The type of the returned sequence is the same as the type of the input iterable