fiftyone.utils.torch

module documentation

(source)

PyTorch utilities.

Class	`ClassifierOutputProcessor`	Output processor for single label classifiers.
Class	`DetectorOutputProcessor`	Output processor for object detectors.
Class	`InstanceSegmenterOutputProcessor`	Output processor for instance segementers.
Class	`KeypointDetectorOutputProcessor`	Output processor for keypoint detection models.
Class	`MaxResize`	Transform that resizes the PIL image or torch Tensor, if necessary, so that its maximum dimensions are at most the specified size.
Class	`MinResize`	Transform that resizes the PIL image or torch Tensor, if necessary, so that its minimum dimensions are at least the specified size.
Class	`OutputProcessor`	Interface for processing the outputs of Torch models.
Class	`PatchSize`	Transform that center crops the PIL image or torch Tensor, if necessary, so that its dimensions are multiples of the specified patch size.
Class	`SaveLayerTensor`	Callback that saves the input/output tensor of the specified layer of a Torch model during each `forward()` call.
Class	`SemanticSegmenterOutputProcessor`	Output processor for semantic segementers.
Class	`ToPILImage`	Transform that converts a tensor or ndarray to a PIL image, while also allowing PIL images to passthrough.
Class	`TorchEmbeddingsMixin`	Mixin for Torch models that can generate embeddings.
Class	`TorchImageClassificationDataset`	A `torch:torch.utils.data.Dataset` for image classification.
Class	`TorchImageDataset`	A `torch:torch.utils.data.Dataset` of images.
Class	`TorchImageModel`	Wrapper for evaluating a Torch model on images.
Class	`TorchImageModelConfig`	Configuration for running a `TorchImageModel`.
Class	`TorchImagePatchesDataset`	A `torch:torch.utils.data.Dataset` of image patch tensors extracted from a list of images.
Class	`TorchSamplesMixin`	Undocumented
Function	`ensure_torch_hub_requirements`	Verifies that the package requirements from the `requirements.txt` file on disk associated with a downloaded PyTorch Hub model are installed.
Function	`find_torch_hub_requirements`	Locates the `requirements.txt` file on disk associated with a downloaded PyTorch Hub model.
Function	`from_image_classification_dir_tree`	Creates a `torch:torch.utils.data.Dataset` for the given image classification dataset directory tree.
Function	`install_torch_hub_requirements`	Installs the package requirements from the `requirements.txt` file on disk associated with a downloaded PyTorch Hub model.
Function	`load_torch_hub_image_model`	Loads an image model from PyTorch Hub as a `TorchImageModel`.
Function	`load_torch_hub_raw_model`	Loads a raw model from PyTorch Hub as a `torch:torch.nn.Module`.
Function	`load_torch_hub_requirements`	Loads the package requirements from the `requirements.txt` file on disk associated with a downloaded PyTorch Hub model.
Function	`recommend_num_workers`	Recommend a number of workers for running a `torch:torch.utils.data.DataLoader`.
Variable	`logger`	Undocumented
Function	`_is_string_array`	Undocumented
Function	`_load_image`	Undocumented
Function	`_polyline_to_bbox`	Undocumented
Function	`_polylines_to_bboxes`	Undocumented
Function	`_strip_comments`	Undocumented
Function	`_to_bytes_array`	Undocumented
Function	`_to_eta_bbox`	Undocumented

def ensure_torch_hub_requirements(repo_or_dir, source='github', error_level=None, log_success=False): (source) ¶

Verifies that the package requirements from the requirements.txt file on disk associated with a downloaded PyTorch Hub model are installed.

Example usage:

import fiftyone.utils.torch as fout

fout.ensure_torch_hub_requirements("facebookresearch/dinov2")

Parameters
repo_or_dir	see `torch:torch.hub.load`
source:"github"	see `torch:torch.hub.load`
error_level:`None`	the error level to use, defined as: 0: raise error if requirement is not satisfied 1: log warning if requirement is not satisfied 2: ignore unsatisifed requirements By default, `fiftyone.config.requirement_error_level` is used
log_success:`False`	whether to generate a log message if a requirement is satisfied

def find_torch_hub_requirements(repo_or_dir, source='github'): (source) ¶

Locates the requirements.txt file on disk associated with a downloaded PyTorch Hub model.

Example usage:

import fiftyone.utils.torch as fout

req_path = fout.find_torch_hub_requirements("facebookresearch/dinov2")

print(req_path)
# '~/.cache/torch/hub/facebookresearch_dinov2_main/requirements.txt'

Parameters
repo_or_dir	see `torch:torch.hub.load`
source:"github"	see `torch:torch.hub.load`
Returns
the path to the requirements file on disk

def from_image_classification_dir_tree(dataset_dir): (source) ¶

Creates a torch:torch.utils.data.Dataset for the given image classification dataset directory tree.

The directory should have the following format:

<dataset_dir>/
    <classA>/
        <image1>.<ext>
        <image2>.<ext>
        ...
    <classB>/
        <image1>.<ext>
        <image2>.<ext>
        ...

Parameters
dataset_dir	the dataset directory
Returns
a `torchvision:torchvision.datasets.ImageFolder`

def install_torch_hub_requirements(repo_or_dir, source='github', error_level=None): (source) ¶

Installs the package requirements from the requirements.txt file on disk associated with a downloaded PyTorch Hub model.

Example usage:

import fiftyone.utils.torch as fout

fout.install_torch_hub_requirements("facebookresearch/dinov2")

Parameters
repo_or_dir	see `torch:torch.hub.load`
source:"github"	see `torch:torch.hub.load`
error_level:`None`	the error level to use, defined as: 0: raise error if the install fails 1: log warning if the install fails 2: ignore install fails By default, `fiftyone.config.requirement_error_level` is used

def load_torch_hub_image_model(repo_or_dir, model, hub_kwargs=None, **kwargs): (source) ¶

Loads an image model from PyTorch Hub as a TorchImageModel.

Example usage:

import fiftyone.utils.torch as fout

model = fout.load_torch_hub_image_model(
    "facebookresearch/dinov2",
    "dinov2_vits14",
    image_patch_size=14,
    embeddings_layer="head",
)

assert model.has_embeddings is True

Parameters
repo_or_dir	see `torch:torch.hub.load`
model	see `torch:torch.hub.load`
hub_kwargs	Undocumented
**kwargs	additional parameters for `TorchImageModelConfig`
Returns
a `TorchImageModel`

def load_torch_hub_raw_model(*args, **kwargs): (source) ¶

Loads a raw model from PyTorch Hub as a torch:torch.nn.Module.

Example usage:

import fiftyone.utils.torch as fout

model = fout.load_torch_hub_raw_model(
    "facebookresearch/dinov2",
    "dinov2_vits14",
)

print(type(model))
# <class 'dinov2.models.vision_transformer.DinoVisionTransformer'>

Parameters
*args	positional arguments for `torch:torch.hub.load`
**kwargs	keyword arguments for `torch:torch.hub.load`
Returns
a `torch:torch.nn.Module`

def load_torch_hub_requirements(repo_or_dir, source='github'): (source) ¶

Loads the package requirements from the requirements.txt file on disk associated with a downloaded PyTorch Hub model.

Example usage:

import fiftyone.utils.torch as fout

requirements = fout.load_torch_hub_requirements("facebookresearch/dinov2")

print(requirements)
# ['torch==2.0.0', 'torchvision==0.15.0', ...]

Parameters
repo_or_dir	see `torch:torch.hub.load`
source:"github"	see `torch:torch.hub.load`
Returns
a list of requirement strings

def recommend_num_workers(): (source) ¶

Recommend a number of workers for running a torch:torch.utils.data.DataLoader.

Returns
the recommended number of workers

logger = (source) ¶

Undocumented

def _is_string_array(targets): (source) ¶

Undocumented

def _load_image(image_path, use_numpy, force_rgb): (source) ¶

Undocumented

def _polyline_to_bbox(points): (source) ¶

Undocumented

def _polylines_to_bboxes(points): (source) ¶

Undocumented

def _strip_comments(requirement_str): (source) ¶

Undocumented

def _to_bytes_array(strs): (source) ¶

Undocumented

def _to_eta_bbox(bounding_box): (source) ¶

Undocumented