lightly.utils

.io

I/O operations to save and load embeddings.

lightly.utils.io.check_embeddings(path: str, remove_additional_columns: bool = False) → None

Raises an error if the embeddings csv file has not the correct format

Use this check whenever you want to upload an embedding to the Lightly Platform. This method only checks whether the header row matches the specs: https://docs.lightly.ai/self-supervised-learning/getting_started/command_line_tool.html#id1

Parameters

  • path – Path to the embedding csv file

  • remove_additional_columns – If True, all additional columns which are not in {filenames, embeddings_x, labels} are removed. If false, they are kept unchanged.

Raises

RuntimeError –

lightly.utils.io.load_embeddings(path: str) → Tuple[NDArray[np.float64], List[int], List[str]]

Loads embeddings from a csv file in a Lightly compatible format.

Parameters

path – Path to the csv file.

Returns

The embeddings as a numpy array, labels as a list of integers, and filenames as a list of strings in the order they were saved.

The embeddings will always be of the Float32 datatype.

Examples

>>> import lightly.utils.io as io
>>> embeddings, labels, filenames = io.load_embeddings(
>>>     'path/to/my/embeddings.csv')
lightly.utils.io.load_embeddings_as_dict(path: str, embedding_name: str = 'default', return_all: bool = False) → Union[Any, Tuple[Any, NDArray[np.float64], List[int], List[str]]]

Loads embeddings from csv and store it in a dictionary for transfer.

Loads embeddings to a dictionary which can be serialized and sent to the Lightly servers. It is recommended that the embedding_name is always specified because the Lightly web-app does not allow two embeddings with the same name.

Parameters

  • path – Path to the csv file.

  • embedding_name – Name of the embedding for the platform.

  • return_all – If true, return embeddings, labels, and filenames, too.

Returns

A dictionary containing the embedding information (see load_embeddings)

Examples

>>> import lightly.utils.io as io
>>> embedding_dict = io.load_embeddings_as_dict(
>>>     'path/to/my/embeddings.csv',
>>>     embedding_name='MyEmbeddings')
>>>
>>> result = io.load_embeddings_as_dict(
>>>     'path/to/my/embeddings.csv',
>>>     embedding_name='MyEmbeddings',
>>>     return_all=True)
>>> embedding_dict, embeddings, labels, filenames = result
lightly.utils.io.save_embeddings(path: str, embeddings: NDArray[np.float64], labels: List[int], filenames: List[str]) → None

Saves embeddings in a csv file in a Lightly compatible format.

Creates a csv file at the location specified by path and saves embeddings, labels, and filenames.

Parameters

  • path – Path to the csv file.

  • embeddings – Embeddings of the images as a numpy array (n x d).

  • labels – List of integer labels.

  • filenames – List of filenames.

Raises

ValueError – If embeddings, labels, and filenames have different lengths.

Examples

>>> import lightly.utils.io as io
>>> io.save_embeddings(
>>>     'path/to/my/embeddings.csv',
>>>     embeddings,
>>>     labels,
>>>     filenames)
lightly.utils.io.save_schema(path: str, task_type: str, ids: List[int], names: List[str]) → None

Saves a prediction schema in the right format.

Parameters

  • path – Where to store the schema.

  • task_type – Task type (e.g. classification, object-detection).

  • ids – List of category ids.

  • names – List of category names.

lightly.utils.io.save_tasks(path: str, tasks: List[str]) → None

Saves a list of prediction task names in the right format.

Parameters

  • path – Where to store the task names.

  • tasks – List of task names.

.embeddings_2d

Transforms embeddings to two-dimensional space for visualization.

class lightly.utils.embeddings_2d.PCA(n_components: int = 2, eps: float = 1e-10)

Handmade PCA to bypass sklearn dependency.

n_components

Number of principal components to keep.

eps

Epsilon for numerical stability.

mean

Mean of the data.

w

Eigenvectors of the covariance matrix.

fit(X: NDArray[np.float32]) → PCA

Fits PCA to data in X.

Parameters

X – Datapoints stored in numpy array of size n x d.

Returns

PCA – The fitted PCA object to transform data points.

transform(X: NDArray[np.float32]) → NDArray[np.float32]

Uses PCA to transform data in X.

Parameters

X – Datapoints stored in numpy array of size n x d.

Returns

Numpy array of n x p datapoints where p <= d.

Raises

ValueError – If PCA is not fitted before calling this method.

lightly.utils.embeddings_2d.fit_pca(embeddings: NDArray[np.float32], n_components: int = 2, fraction: Optional[float] = None) → PCA

Fits PCA to a randomly selected subset of embeddings.

For large datasets, it can be unfeasible to perform PCA on the whole data. This method can fit a PCA on a fraction of the embeddings in order to save computational resources.

Parameters

  • embeddings – Datapoints stored in numpy array of size n x d.

  • n_components – Number of principal components to keep.

  • fraction – Fraction of the dataset to fit PCA on.

Returns

A transformer which can be used to transform embeddings to lower dimensions.

Raises

If fraction ≤ 0 or fraction > 1. –

.benchmarking

.debug

lightly.utils.debug.apply_transform_without_normalize(image: Image, transform) → Image

Applies the transform to the image but skips ToTensor and Normalize.

Parameters

  • image – The input PIL image.

  • transform – The transformation to apply, excluding ToTensor and Normalize.

Returns

The transformed image.

lightly.utils.debug.generate_grid_of_augmented_images(input_images: List[Image], collate_function: Union[BaseCollateFunction, MultiViewCollateFunction]) → List[List[Image]]

Returns a grid of augmented images. Images in a column belong together.

This function ignores the ToTensor and Normalize transforms for visualization purposes.

Parameters

  • input_images – List of PIL images for which the augmentations should be plotted.

  • collate_function – The collate function of the self-supervised learning algorithm. Must be of type BaseCollateFunction or MultiViewCollateFunction.

Returns

A grid of augmented images. Images in a column belong together.

lightly.utils.debug.plot_augmented_images(input_images: List[Image], collate_function: Union[BaseCollateFunction, MultiViewCollateFunction])

Plots original images and augmented images in a figure.

This function ignores the ToTensor and Normalize transforms for visualization purposes.

Parameters

  • input_images – List of PIL images for which the augmentations should be plotted.

  • collate_function – The collate function of the self-supervised learning algorithm. Must be of type BaseCollateFunction or MultiViewCollateFunction.

Returns

A figure showing the original images in the left column and the augmented images to their right. If the collate_function is an instance of the BaseCollateFunction, two example augmentations are shown. For MultiViewCollateFunctions all the generated views are shown.

lightly.utils.debug.std_of_l2_normalized(z: Tensor) → Tensor

Calculates the mean of the standard deviation of z along each dimension.

This measure was used by [0] to determine the level of collapse of the learned representations. If the returned value is 0., the outputs z have collapsed to a constant vector. If the output z has a zero-mean isotropic Gaussian distribution [0], the returned value should be close to 1/sqrt(d), where d is the dimensionality of the output.

[0]: https://arxiv.org/abs/2011.10566

Parameters

z – A torch tensor of shape batch_size x dimension.

Returns

The mean of the standard deviation of the l2 normalized tensor z along each dimension.

.dist

class lightly.utils.dist.GatherLayer(*args, **kwargs)

Gather tensors from all processes, supporting backward propagation.

Adapted from the Solo-Learn project: https://github.com/vturrisi/solo-learn/blob/b69b4bd27472593919956d9ac58902a301537a4d/solo/utils/misc.py#L187

static backward(ctx: FunctionCtx, *grads: Tensor) → Tensor

Defines a formula for differentiating the operation with backward mode automatic differentiation (alias to the vjp function).

This function is to be overridden by all subclasses.

It must accept a context ctx as the first argument, followed by as many outputs as the forward() returned (None will be passed in for non tensor outputs of the forward function), and it should return as many tensors, as there were inputs to forward(). Each argument is the gradient w.r.t the given output, and each returned value should be the gradient w.r.t. the corresponding input. If an input is not a Tensor or is a Tensor not requiring grads, you can just pass None as a gradient for that input.

The context can be used to retrieve tensors saved during the forward pass. It also has an attribute ctx.needs_input_grad as a tuple of booleans representing whether each input needs gradient. E.g., backward() will have ctx.needs_input_grad[0] = True if the first input to forward() needs gradient computated w.r.t. the output.

static forward(ctx: FunctionCtx, input: Tensor) → Tuple[Tensor, ...]

Performs the operation.

This function is to be overridden by all subclasses.

It must accept a context ctx as the first argument, followed by any number of arguments (tensors or other types).

The context can be used to store arbitrary data that can be then retrieved during the backward pass. Tensors should not be stored directly on ctx (though this is not currently enforced for backward compatibility). Instead, tensors should be saved either with ctx.save_for_backward() if they are intended to be used in backward (equivalently, vjp) or ctx.save_for_forward() if they are intended to be used for in jvp.

lightly.utils.dist.eye_rank(n: int, device: Optional[device] = None) → Tensor

Returns an (n, n * world_size) zero matrix with the diagonal for the rank of this process set to 1.

Example output where n=3, the current process has rank 1, and there are 4 processes in total:

rank0 rank1 rank2 rank3 0 0 0 | 1 0 0 | 0 0 0 | 0 0 0 0 0 0 | 0 1 0 | 0 0 0 | 0 0 0 0 0 0 | 0 0 1 | 0 0 0 | 0 0 0

Equivalent to torch.eye for undistributed settings or if world size == 1.

Parameters

  • n – Size of the square matrix on a single process.

  • device – Device on which the matrix should be created.

Returns

A tensor with the appropriate diagonal filled for this rank.

lightly.utils.dist.gather(input: Tensor) → Tuple[Tensor]

Gathers a tensor from all processes and supports backpropagation.

lightly.utils.dist.rank() → int

Returns the rank of the current process.

lightly.utils.dist.rank_zero_only(fn: Callable[[...], R]) → Callable[[...], Optional[R]]

Decorator to ensure the function only runs on the process with rank 0.

Example

>>> @rank_zero_only
>>> def print_rank_zero(message: str):
>>>     print(message)
>>>
>>> print_rank_zero("Hello from rank 0!")
lightly.utils.dist.world_size() → int

Returns the current world size (number of distributed processes).

.reordering

lightly.utils.reordering.sort_items_by_keys(keys: Sequence[_K], items: Sequence[_V], sorted_keys: Sequence[_K]) → List[_V]

Sorts the items in the same order as the sorted keys.

Parameters

  • keys – Keys by which items can be identified.

  • items – Items to sort.

  • sorted_keys – Keys in sorted order.

Returns

The list of sorted items.

Examples

>>> keys = [3, 2, 1]
>>> items = ['!', 'world', 'hello']
>>> sorted_keys = [1, 2, 3]
>>> sorted_items = sort_items_by_keys(
>>>     keys,
>>>     items,
>>>     sorted_keys,
>>> )
>>> print(sorted_items)
>>> > ['hello', 'world', '!']

.version_compare

Utility method for comparing versions of libraries

lightly.utils.version_compare.version_compare(v0: str, v1: str) → int

Returns 1 if version of v0 is larger than v1 and -1 otherwise

Use this method to compare Python package versions and see which one is newer.

Examples

>>> # compare two versions
>>> version_compare('1.2.0', '1.1.2')
>>> 1