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[numpy.ndarray[Any, numpy.dtype[numpy.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, numpy.ndarray[Any, numpy.dtype[numpy.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: numpy.ndarray[Any, numpy.dtype[numpy.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
Transform 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.
- fit(X: numpy.ndarray)
Fits PCA to data in X.
- Parameters
X – Datapoints stored in numpy array of size n x d.
- Returns
PCA object to transform datapoints.
- transform(X: numpy.ndarray)
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.
- lightly.utils.embeddings_2d.fit_pca(embeddings: numpy.ndarray, n_components: int = 2, fraction: Optional[float] = None)
Fits PCA to 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
ValueError – If fraction < 0 or fraction > 1.
.benchmarking
.debug
- lightly.utils.debug.apply_transform_without_normalize(image: PIL.Image.Image, transform)
Applies the transform to the image but skips ToTensor and Normalize.
- lightly.utils.debug.generate_grid_of_augmented_images(input_images: List[PIL.Image.Image], collate_function: Union[lightly.data.collate.BaseCollateFunction, lightly.data.collate.MultiViewCollateFunction])
Returns a grid of augmented images. Images in a column belong together.
This function ignores the transforms ToTensor and Normalize 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[PIL.Image.Image], collate_function: Union[lightly.data.collate.BaseCollateFunction, lightly.data.collate.MultiViewCollateFunction])
Returns a figure showing original images in the left column and augmented images to their right.
This function ignores the transforms ToTensor and Normalize 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: torch.Tensor) torch.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 number 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 number 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.
This code was taken and adapted from here: https://github.com/Spijkervet/SimCLR
- static backward(ctx, *grads: torch.Tensor) torch.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 theforward()
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 toforward()
. 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 havectx.needs_input_grad[0] = True
if the first input toforward()
needs gradient computated w.r.t. the output.
- static forward(ctx, input: torch.Tensor) Tuple[torch.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 inbackward
(equivalently,vjp
) orctx.save_for_forward()
if they are intended to be used for injvp
.
- lightly.utils.dist.eye_rank(n: int, device: Optional[torch.device] = None) torch.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.
- lightly.utils.dist.gather(input: torch.Tensor) Tuple[torch.Tensor]
Gathers this tensor from all processes. Supports backprop.
- lightly.utils.dist.rank() int
Returns the rank of the current process.
- lightly.utils.dist.rank_zero_only(fn)
Decorator that only runs the function 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[lightly.utils.reordering._K], items: Sequence[lightly.utils.reordering._V], sorted_keys: Sequence[lightly.utils.reordering._K]) List[lightly.utils.reordering._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