Name | uotod JSON |
Version |
0.3.post2
JSON |
| download |
home_page | |
Summary | Unbalanced Optimal Transport for Object Detection |
upload_time | 2024-02-12 14:23:17 |
maintainer | |
docs_url | None |
author | |
requires_python | >=3.8 |
license | GNU LESSER GENERAL PUBLIC LICENSE Version 3, 29 June 2007 Copyright (C) 2007 Free Software Foundation, Inc. <https://fsf.org/> Everyone is permitted to copy and distribute verbatim copies of this license document, but changing it is not allowed. This version of the GNU Lesser General Public License incorporates the terms and conditions of version 3 of the GNU General Public License, supplemented by the additional permissions listed below. 0. Additional Definitions. As used herein, "this License" refers to version 3 of the GNU Lesser General Public License, and the "GNU GPL" refers to version 3 of the GNU General Public License. "The Library" refers to a covered work governed by this License, other than an Application or a Combined Work as defined below. An "Application" is any work that makes use of an interface provided by the Library, but which is not otherwise based on the Library. Defining a subclass of a class defined by the Library is deemed a mode of using an interface provided by the Library. A "Combined Work" is a work produced by combining or linking an Application with the Library. The particular version of the Library with which the Combined Work was made is also called the "Linked Version". The "Minimal Corresponding Source" for a Combined Work means the Corresponding Source for the Combined Work, excluding any source code for portions of the Combined Work that, considered in isolation, are based on the Application, and not on the Linked Version. The "Corresponding Application Code" for a Combined Work means the object code and/or source code for the Application, including any data and utility programs needed for reproducing the Combined Work from the Application, but excluding the System Libraries of the Combined Work. 1. Exception to Section 3 of the GNU GPL. You may convey a covered work under sections 3 and 4 of this License without being bound by section 3 of the GNU GPL. 2. Conveying Modified Versions. If you modify a copy of the Library, and, in your modifications, a facility refers to a function or data to be supplied by an Application that uses the facility (other than as an argument passed when the facility is invoked), then you may convey a copy of the modified version: a) under this License, provided that you make a good faith effort to ensure that, in the event an Application does not supply the function or data, the facility still operates, and performs whatever part of its purpose remains meaningful, or b) under the GNU GPL, with none of the additional permissions of this License applicable to that copy. 3. Object Code Incorporating Material from Library Header Files. The object code form of an Application may incorporate material from a header file that is part of the Library. You may convey such object code under terms of your choice, provided that, if the incorporated material is not limited to numerical parameters, data structure layouts and accessors, or small macros, inline functions and templates (ten or fewer lines in length), you do both of the following: a) Give prominent notice with each copy of the object code that the Library is used in it and that the Library and its use are covered by this License. b) Accompany the object code with a copy of the GNU GPL and this license document. 4. Combined Works. You may convey a Combined Work under terms of your choice that, taken together, effectively do not restrict modification of the portions of the Library contained in the Combined Work and reverse engineering for debugging such modifications, if you also do each of the following: a) Give prominent notice with each copy of the Combined Work that the Library is used in it and that the Library and its use are covered by this License. b) Accompany the Combined Work with a copy of the GNU GPL and this license document. c) For a Combined Work that displays copyright notices during execution, include the copyright notice for the Library among these notices, as well as a reference directing the user to the copies of the GNU GPL and this license document. d) Do one of the following: 0) Convey the Minimal Corresponding Source under the terms of this License, and the Corresponding Application Code in a form suitable for, and under terms that permit, the user to recombine or relink the Application with a modified version of the Linked Version to produce a modified Combined Work, in the manner specified by section 6 of the GNU GPL for conveying Corresponding Source. 1) Use a suitable shared library mechanism for linking with the Library. A suitable mechanism is one that (a) uses at run time a copy of the Library already present on the user's computer system, and (b) will operate properly with a modified version of the Library that is interface-compatible with the Linked Version. e) Provide Installation Information, but only if you would otherwise be required to provide such information under section 6 of the GNU GPL, and only to the extent that such information is necessary to install and execute a modified version of the Combined Work produced by recombining or relinking the Application with a modified version of the Linked Version. (If you use option 4d0, the Installation Information must accompany the Minimal Corresponding Source and Corresponding Application Code. If you use option 4d1, you must provide the Installation Information in the manner specified by section 6 of the GNU GPL for conveying Corresponding Source.) 5. Combined Libraries. You may place library facilities that are a work based on the Library side by side in a single library together with other library facilities that are not Applications and are not covered by this License, and convey such a combined library under terms of your choice, if you do both of the following: a) Accompany the combined library with a copy of the same work based on the Library, uncombined with any other library facilities, conveyed under the terms of this License. b) Give prominent notice with the combined library that part of it is a work based on the Library, and explaining where to find the accompanying uncombined form of the same work. 6. Revised Versions of the GNU Lesser General Public License. The Free Software Foundation may publish revised and/or new versions of the GNU Lesser General Public License from time to time. Such new versions will be similar in spirit to the present version, but may differ in detail to address new problems or concerns. Each version is given a distinguishing version number. If the Library as you received it specifies that a certain numbered version of the GNU Lesser General Public License "or any later version" applies to it, you have the option of following the terms and conditions either of that published version or of any later version published by the Free Software Foundation. If the Library as you received it does not specify a version number of the GNU Lesser General Public License, you may choose any version of the GNU Lesser General Public License ever published by the Free Software Foundation. If the Library as you received it specifies that a proxy can decide whether future versions of the GNU Lesser General Public License shall apply, that proxy's public statement of acceptance of any version is permanent authorization for you to choose that version for the Library. |
keywords |
pytorch
machine learning
|
VCS |
|
bugtrack_url |
|
requirements |
No requirements were recorded.
|
Travis-CI |
No Travis.
|
coveralls test coverage |
No coveralls.
|
# Unbalanced Optimal Transport: A Unified Framework for Object Detection
<a href="https://hdeplaen.github.io/uotod/" target="_blank">Presentation</a>
<a href="https://openaccess.thecvf.com/content/CVPR2023/papers/De_Plaen_Unbalanced_Optimal_Transport_A_Unified_Framework_for_Object_Detection_CVPR_2023_paper.pdf" target="_blank">Paper</a>
<a href="https://openaccess.thecvf.com/content/CVPR2023/supplemental/De_Plaen_Unbalanced_Optimal_Transport_CVPR_2023_supplemental.pdf" target="_blank">Supplementary</a>
<a href="https://uotod.readthedocs.io/en/latest/" target="_blank">Documentation</a>
![GitHub License](https://img.shields.io/github/license/hdeplaen/uotod)
![PyPI - Downloads](https://img.shields.io/pypi/dm/uotod)
![PyPI - Version](https://img.shields.io/pypi/v/uotod)
[![Documentation Status](https://readthedocs.org/projects/uotod/badge/?version=latest)](https://uotod.readthedocs.io/en/latest/?badge=latest)
[![Test Status](https://github.com/hdeplaen/uotod/actions/workflows/test.yaml/badge.svg?branch=main)](https://github.com/hdeplaen/uotod/actions/workflows/test.yaml)
[![Build Status](https://github.com/hdeplaen/uotod/actions/workflows/build.yaml/badge.svg?branch=main)](https://github.com/hdeplaen/uotod/actions/workflows/build.yaml)
[//]: # (![GitHub all releases](https://img.shields.io/github/downloads/hdeplaen/uotod/total))
H. De Plaen, P.-F. De Plaen, J. A. K. Suykens, M. Proesmans, T. Tuytelaars, and L. Van Gool, “Unbalanced Optimal Transport: A Unified Framework for Object Detection,” in *Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)*, Jun. 2023, pp. 3198–3207.
This work has be presented at CVPR 2023 in Vancouver, Canada. The paper and additional resources can be found on the [following website](https://hdeplaen.github.io/uotod/). The paper is in open access and can also be found on the [CVF website](https://openaccess.thecvf.com/content/CVPR2023/html/De_Plaen_Unbalanced_Optimal_Transport_A_Unified_Framework_for_Object_Detection_CVPR_2023_paper.html) as well as on [IEEE Xplore](https://ieeexplore.ieee.org/document/10204500).
![Different matching strategies. All are particular cases of Unbalanced Optimal Transport](img/illustration.png)
## Abstract
*TL;DR: We introduce a much more versatile new class of matching strategies unifying many existing ones, as well as being well suited for GPUs.*
During training, supervised object detection tries to correctly match the predicted bounding boxes and associated classification scores to the ground truth. This is essential to determine which predictions are to be pushed towards which solutions, or to be discarded. Popular matching strategies include matching to the closest ground truth box (mostly used in combination with anchors), or matching via the Hungarian algorithm (mostly used in anchor-free methods). Each of these strategies comes with its own properties, underlying losses, and heuristics. We show how Unbalanced Optimal Transport unifies these different approaches and opens a whole continuum of methods in between. This allows for a finer selection of the desired properties. Experimentally, we show that training an object detection model with Unbalanced Optimal Transport is able to reach the state-of-the-art both in terms of Average Precision and Average Recall as well as to provide a faster initial convergence. The approach is well suited for GPU implementation, which proves to be an advantage for large-scale models.
## Install
### PyPI
Using PyPI, it suffices to run `pip install uotod`. Just rerun this command to update the package to its newest version.
### Build From Source
You can also download it directly from the GitHub repository, then build and install it.
```bash
git clone --recursive https://github.com/hdeplaen/uotod
cd uotod
python3 -m pip install -r requirements.txt
python3 -m setup build
python3 -m pip install
```
### Compiled Acceleration
The package is **available on all dsitributions** and runs well. However, only the combinations marked with a green ✅ can
take advantage of the compiled version of Sinkhorn's algorithm directly from PyPI. On a not support combination, you may always build it
from the source to also have access to Sinkhorn's compiled version of the algorithm. Nevertheless, the PyTorch implementation
of **Sinkhorn's algorithm is always available** (used by default), this only refers to an additional compiled version.
| **OS** | **Linux** | **MacOS** | **Windows** |
|----------------- |:---------: |:-----------:|:-----------: |
| **Python 3.8** | ✅ | ✅ | ☑️ |
| **Python 3.9** | ✅ | ✅ | ☑️ |
| **Python 3.10** | ✅ | ✅ | ☑️ |
| **Python 3.11** | ✅ | ✅ | ☑️ |
| **Python 3.12** | ✅ | ☑️ | ☑️ |
- ✅: Python implementation + compiled acceleration, _both directly from PyPI_
- ☑️: Python implementation _directly from PyPI_ (+ possible compiled acceleration if building from source)
## Examples
### OT matching with GIoU loss:
```python
from uotod.match import BalancedSinkhorn
from uotod.loss import GIoULoss
ot = BalancedSinkhorn(
loc_match_module=GIoULoss(reduction="none"),
background_cost=0.,
)
```
### Hungarian matching (bipartite) with GIoU loss:
```python
from uotod.match import Hungarian
from uotod.loss import GIoULoss
hungarian = Hungarian(
loc_match_module=GIoULoss(reduction="none"),
background_cost=0.,
)
```
### Loss from SSD solved with Unbalanced Optimal Transport:
```python
from torch.nn import L1Loss, CrossEntropyLoss
from uotod.match import UnbalancedSinkhorn
from uotod.loss import DetectionLoss, IoULoss
matching_method = UnbalancedSinkhorn(
cls_match_module=None, # No classification cost
loc_match_module=IoULoss(reduction="none"),
background_cost=0.5, # Threshold for matching to background
is_anchor_based=True, # Use anchor-based matching
reg_target=1e-3, # Relax the constraint that each ground truth is matched to exactly one prediction
)
loss = DetectionLoss(
matching_method=matching_method,
cls_loss_module=CrossEntropyLoss(reduction="none"),
loc_loss_module=L1Loss(reduction="none"),
)
preds = ...
targets = ...
anchors = ...
loss_value = loss(preds, targets, anchors)
```
### Loss from DETR solved with Optimal Transport (num_classes=3):
```python
import torch
from torch.nn import L1Loss, CrossEntropyLoss
from uotod.match import BalancedSinkhorn
from uotod.loss import DetectionLoss
from uotod.loss import MultipleObjectiveLoss, GIoULoss, NegativeProbLoss
matching_method = BalancedSinkhorn(
cls_match_module=NegativeProbLoss(reduction="none"),
loc_match_module=MultipleObjectiveLoss(
losses=[GIoULoss(reduction="none"), L1Loss(reduction="none")],
weights=[1., 5.],
),
background_cost=0., # Does not influence the matching when using balanced OT
)
loss = DetectionLoss(
matching_method=matching_method,
cls_loss_module=CrossEntropyLoss(
reduction="none",
weight=torch.tensor([0.1, 1., 1.]) # down-weight the loss for the no-object class
),
loc_loss_module=MultipleObjectiveLoss(
losses=[GIoULoss(reduction="none"), L1Loss(reduction="none")],
weights=[1., 5.],
),
)
preds = ...
targets = ...
loss_value = loss(preds, targets)
```
## Color Boxes Dataset
![Examples from the Color Boxes Dataset](img/colorboxes.png)
## Citation
If you make any use of our work, please refer to us as:
```bibtex
@InProceedings{De_Plaen_2023_CVPR,
author = {De Plaen, Henri and De Plaen, Pierre-Fran\c{c}ois and Suykens, Johan A. K. and Proesmans, Marc and Tuytelaars, Tinne and Van Gool, Luc},
title = {Unbalanced Optimal Transport: A Unified Framework for Object Detection},
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2023},
pages = {3198-3207}
}
```
## Acknowledgements
EU: The research leading to these results has received funding from the European Research Council under the European Union’s Horizon 2020 research and innovation program / ERC Advanced Grant E-DUALITY (787960). This paper reflects only the authors’ views and the Union is not liable for any use that may be made of the contained information. Research Council KUL: Optimization frameworks for deep kernel machines C14/18/068. Flemish Government: FWO: projects: GOA4917N (Deep Restricted Kernel Machines: Methods and Foundations), PhD/Postdoc grant; This research received funding from the Flemish Government (AI Research Program). All the authors are also affiliated to Leuven.AI - KU Leuven institute for AI, B-3000, Leuven, Belgium.
<p style="text-align: center;">
<img src="https://hdeplaen.github.io/uotod/img/eu.png" alt="European Union" style="height:80px;"/>
<img src="https://hdeplaen.github.io/uotod/img/erc.png" alt="European Research Council" style="height:80px;"/>
<img src="https://hdeplaen.github.io/uotod/img/fwo.png" alt="Fonds voor Wetenschappelijk Onderzoek" style="height:80px;"/>
<img src="https://hdeplaen.github.io/uotod/img/vl.png" alt="Flanders AI" style="height:80px;"/>
<img src="https://hdeplaen.github.io/uotod/img/kuleuven.png" alt="KU Leuven" style="height:80px;"/>
</p>
Raw data
{
"_id": null,
"home_page": "",
"name": "uotod",
"maintainer": "",
"docs_url": null,
"requires_python": ">=3.8",
"maintainer_email": "",
"keywords": "pytorch,machine learning",
"author": "",
"author_email": "Henri DE PLAEN <henri.deplaen@gmail.com>, Pierre-Fran\u00e7ois DE PLAEN <pierre-francois.deplaen@esat.kuleuven.be>",
"download_url": "https://files.pythonhosted.org/packages/d3/fe/733285392e909725a2227c42bf65221e33f657d0f1efbe0526feb72626e7/uotod-0.3.post2.tar.gz",
"platform": null,
"description": "# Unbalanced Optimal Transport: A Unified Framework for Object Detection\n<a href=\"https://hdeplaen.github.io/uotod/\" target=\"_blank\">Presentation</a> \n<a href=\"https://openaccess.thecvf.com/content/CVPR2023/papers/De_Plaen_Unbalanced_Optimal_Transport_A_Unified_Framework_for_Object_Detection_CVPR_2023_paper.pdf\" target=\"_blank\">Paper</a> \n<a href=\"https://openaccess.thecvf.com/content/CVPR2023/supplemental/De_Plaen_Unbalanced_Optimal_Transport_CVPR_2023_supplemental.pdf\" target=\"_blank\">Supplementary</a> \n<a href=\"https://uotod.readthedocs.io/en/latest/\" target=\"_blank\">Documentation</a> \n\n\n![GitHub License](https://img.shields.io/github/license/hdeplaen/uotod)\n![PyPI - Downloads](https://img.shields.io/pypi/dm/uotod)\n![PyPI - Version](https://img.shields.io/pypi/v/uotod)\n[![Documentation Status](https://readthedocs.org/projects/uotod/badge/?version=latest)](https://uotod.readthedocs.io/en/latest/?badge=latest)\n[![Test Status](https://github.com/hdeplaen/uotod/actions/workflows/test.yaml/badge.svg?branch=main)](https://github.com/hdeplaen/uotod/actions/workflows/test.yaml)\n[![Build Status](https://github.com/hdeplaen/uotod/actions/workflows/build.yaml/badge.svg?branch=main)](https://github.com/hdeplaen/uotod/actions/workflows/build.yaml)\n\n[//]: # (![GitHub all releases](https://img.shields.io/github/downloads/hdeplaen/uotod/total))\n\nH. De Plaen, P.-F. De Plaen, J. A. K. Suykens, M. Proesmans, T. Tuytelaars, and L. Van Gool, \u201cUnbalanced Optimal Transport: A Unified Framework for Object Detection,\u201d in *Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)*, Jun. 2023, pp. 3198\u20133207.\n\nThis work has be presented at CVPR 2023 in Vancouver, Canada. The paper and additional resources can be found on the [following website](https://hdeplaen.github.io/uotod/). The paper is in open access and can also be found on the [CVF website](https://openaccess.thecvf.com/content/CVPR2023/html/De_Plaen_Unbalanced_Optimal_Transport_A_Unified_Framework_for_Object_Detection_CVPR_2023_paper.html) as well as on [IEEE Xplore](https://ieeexplore.ieee.org/document/10204500).\n\n![Different matching strategies. All are particular cases of Unbalanced Optimal Transport](img/illustration.png)\n\n## Abstract\n*TL;DR: We introduce a much more versatile new class of matching strategies unifying many existing ones, as well as being well suited for GPUs.*\n\nDuring training, supervised object detection tries to correctly match the predicted bounding boxes and associated classification scores to the ground truth. This is essential to determine which predictions are to be pushed towards which solutions, or to be discarded. Popular matching strategies include matching to the closest ground truth box (mostly used in combination with anchors), or matching via the Hungarian algorithm (mostly used in anchor-free methods). Each of these strategies comes with its own properties, underlying losses, and heuristics. We show how Unbalanced Optimal Transport unifies these different approaches and opens a whole continuum of methods in between. This allows for a finer selection of the desired properties. Experimentally, we show that training an object detection model with Unbalanced Optimal Transport is able to reach the state-of-the-art both in terms of Average Precision and Average Recall as well as to provide a faster initial convergence. The approach is well suited for GPU implementation, which proves to be an advantage for large-scale models.\n\n## Install\n### PyPI\n\nUsing PyPI, it suffices to run `pip install uotod`. Just rerun this command to update the package to its newest version.\n\n### Build From Source\n\nYou can also download it directly from the GitHub repository, then build and install it.\n\n```bash\ngit clone --recursive https://github.com/hdeplaen/uotod\ncd uotod\npython3 -m pip install -r requirements.txt\npython3 -m setup build\npython3 -m pip install\n ```\n\n### Compiled Acceleration\n\nThe package is **available on all dsitributions** and runs well. However, only the combinations marked with a green \u2705 can \ntake advantage of the compiled version of Sinkhorn's algorithm directly from PyPI. On a not support combination, you may always build it \nfrom the source to also have access to Sinkhorn's compiled version of the algorithm. Nevertheless, the PyTorch implementation \nof **Sinkhorn's algorithm is always available** (used by default), this only refers to an additional compiled version. \n\n| **OS** \t| **Linux** \t| **MacOS** \t | **Windows** \t|\n|-----------------\t|:---------:\t|:-----------:|:-----------:\t|\n| **Python 3.8** \t| \u2705 \t| \u2705 \t | \u2611\ufe0f \t|\n| **Python 3.9** \t| \u2705 \t| \u2705 \t | \u2611\ufe0f \t|\n| **Python 3.10** \t| \u2705 \t| \u2705 \t | \u2611\ufe0f \t|\n| **Python 3.11** \t| \u2705 \t| \u2705 \t | \u2611\ufe0f \t|\n| **Python 3.12** \t| \u2705 \t| \u2611\ufe0f | \u2611\ufe0f \t|\n\n- \u2705: Python implementation + compiled acceleration, _both directly from PyPI_\n- \u2611\ufe0f: Python implementation _directly from PyPI_ (+ possible compiled acceleration if building from source)\n\n## Examples\n\n### OT matching with GIoU loss:\n\n```python\nfrom uotod.match import BalancedSinkhorn\nfrom uotod.loss import GIoULoss\n\not = BalancedSinkhorn(\n loc_match_module=GIoULoss(reduction=\"none\"),\n background_cost=0.,\n)\n```\n\n### Hungarian matching (bipartite) with GIoU loss:\n\n```python\nfrom uotod.match import Hungarian\nfrom uotod.loss import GIoULoss\n\nhungarian = Hungarian(\n loc_match_module=GIoULoss(reduction=\"none\"),\n background_cost=0.,\n)\n```\n\n### Loss from SSD solved with Unbalanced Optimal Transport:\n\n```python\nfrom torch.nn import L1Loss, CrossEntropyLoss\n\nfrom uotod.match import UnbalancedSinkhorn\nfrom uotod.loss import DetectionLoss, IoULoss\n\nmatching_method = UnbalancedSinkhorn(\n cls_match_module=None, # No classification cost\n loc_match_module=IoULoss(reduction=\"none\"),\n background_cost=0.5, # Threshold for matching to background\n is_anchor_based=True, # Use anchor-based matching\n reg_target=1e-3, # Relax the constraint that each ground truth is matched to exactly one prediction\n)\n\nloss = DetectionLoss(\n matching_method=matching_method,\n cls_loss_module=CrossEntropyLoss(reduction=\"none\"),\n loc_loss_module=L1Loss(reduction=\"none\"),\n)\n\npreds = ...\ntargets = ...\nanchors = ...\n\nloss_value = loss(preds, targets, anchors)\n```\n\n### Loss from DETR solved with Optimal Transport (num_classes=3):\n\n```python\nimport torch\nfrom torch.nn import L1Loss, CrossEntropyLoss\n\nfrom uotod.match import BalancedSinkhorn\nfrom uotod.loss import DetectionLoss\nfrom uotod.loss import MultipleObjectiveLoss, GIoULoss, NegativeProbLoss\n\nmatching_method = BalancedSinkhorn(\n cls_match_module=NegativeProbLoss(reduction=\"none\"),\n loc_match_module=MultipleObjectiveLoss(\n losses=[GIoULoss(reduction=\"none\"), L1Loss(reduction=\"none\")],\n weights=[1., 5.],\n ),\n background_cost=0., # Does not influence the matching when using balanced OT\n)\n\nloss = DetectionLoss(\n matching_method=matching_method,\n cls_loss_module=CrossEntropyLoss(\n reduction=\"none\",\n weight=torch.tensor([0.1, 1., 1.]) # down-weight the loss for the no-object class\n ),\n loc_loss_module=MultipleObjectiveLoss(\n losses=[GIoULoss(reduction=\"none\"), L1Loss(reduction=\"none\")],\n weights=[1., 5.],\n ),\n)\n\npreds = ...\ntargets = ...\nloss_value = loss(preds, targets)\n```\n\n\n## Color Boxes Dataset\n![Examples from the Color Boxes Dataset](img/colorboxes.png)\n\n## Citation\nIf you make any use of our work, please refer to us as:\n```bibtex\n@InProceedings{De_Plaen_2023_CVPR,\n author = {De Plaen, Henri and De Plaen, Pierre-Fran\\c{c}ois and Suykens, Johan A. K. and Proesmans, Marc and Tuytelaars, Tinne and Van Gool, Luc},\n title = {Unbalanced Optimal Transport: A Unified Framework for Object Detection},\n booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},\n month = {June},\n year = {2023},\n pages = {3198-3207}\n}\n```\n\n## Acknowledgements\nEU: The research leading to these results has received funding from the European Research Council under the European Union\u2019s Horizon 2020 research and innovation program / ERC Advanced Grant E-DUALITY (787960). This paper reflects only the authors\u2019 views and the Union is not liable for any use that may be made of the contained information. Research Council KUL: Optimization frameworks for deep kernel machines C14/18/068. Flemish Government: FWO: projects: GOA4917N (Deep Restricted Kernel Machines: Methods and Foundations), PhD/Postdoc grant; This research received funding from the Flemish Government (AI Research Program). All the authors are also affiliated to Leuven.AI - KU Leuven institute for AI, B-3000, Leuven, Belgium.\n<p style=\"text-align: center;\">\n<img src=\"https://hdeplaen.github.io/uotod/img/eu.png\" alt=\"European Union\" style=\"height:80px;\"/>\n<img src=\"https://hdeplaen.github.io/uotod/img/erc.png\" alt=\"European Research Council\" style=\"height:80px;\"/>\n<img src=\"https://hdeplaen.github.io/uotod/img/fwo.png\" alt=\"Fonds voor Wetenschappelijk Onderzoek\" style=\"height:80px;\"/>\n<img src=\"https://hdeplaen.github.io/uotod/img/vl.png\" alt=\"Flanders AI\" style=\"height:80px;\"/>\n<img src=\"https://hdeplaen.github.io/uotod/img/kuleuven.png\" alt=\"KU Leuven\" style=\"height:80px;\"/>\n</p>\n",
"bugtrack_url": null,
"license": "GNU LESSER GENERAL PUBLIC LICENSE Version 3, 29 June 2007 Copyright (C) 2007 Free Software Foundation, Inc. <https://fsf.org/> Everyone is permitted to copy and distribute verbatim copies of this license document, but changing it is not allowed. This version of the GNU Lesser General Public License incorporates the terms and conditions of version 3 of the GNU General Public License, supplemented by the additional permissions listed below. 0. Additional Definitions. As used herein, \"this License\" refers to version 3 of the GNU Lesser General Public License, and the \"GNU GPL\" refers to version 3 of the GNU General Public License. \"The Library\" refers to a covered work governed by this License, other than an Application or a Combined Work as defined below. An \"Application\" is any work that makes use of an interface provided by the Library, but which is not otherwise based on the Library. Defining a subclass of a class defined by the Library is deemed a mode of using an interface provided by the Library. A \"Combined Work\" is a work produced by combining or linking an Application with the Library. The particular version of the Library with which the Combined Work was made is also called the \"Linked Version\". The \"Minimal Corresponding Source\" for a Combined Work means the Corresponding Source for the Combined Work, excluding any source code for portions of the Combined Work that, considered in isolation, are based on the Application, and not on the Linked Version. The \"Corresponding Application Code\" for a Combined Work means the object code and/or source code for the Application, including any data and utility programs needed for reproducing the Combined Work from the Application, but excluding the System Libraries of the Combined Work. 1. Exception to Section 3 of the GNU GPL. You may convey a covered work under sections 3 and 4 of this License without being bound by section 3 of the GNU GPL. 2. Conveying Modified Versions. If you modify a copy of the Library, and, in your modifications, a facility refers to a function or data to be supplied by an Application that uses the facility (other than as an argument passed when the facility is invoked), then you may convey a copy of the modified version: a) under this License, provided that you make a good faith effort to ensure that, in the event an Application does not supply the function or data, the facility still operates, and performs whatever part of its purpose remains meaningful, or b) under the GNU GPL, with none of the additional permissions of this License applicable to that copy. 3. Object Code Incorporating Material from Library Header Files. The object code form of an Application may incorporate material from a header file that is part of the Library. You may convey such object code under terms of your choice, provided that, if the incorporated material is not limited to numerical parameters, data structure layouts and accessors, or small macros, inline functions and templates (ten or fewer lines in length), you do both of the following: a) Give prominent notice with each copy of the object code that the Library is used in it and that the Library and its use are covered by this License. b) Accompany the object code with a copy of the GNU GPL and this license document. 4. Combined Works. You may convey a Combined Work under terms of your choice that, taken together, effectively do not restrict modification of the portions of the Library contained in the Combined Work and reverse engineering for debugging such modifications, if you also do each of the following: a) Give prominent notice with each copy of the Combined Work that the Library is used in it and that the Library and its use are covered by this License. b) Accompany the Combined Work with a copy of the GNU GPL and this license document. c) For a Combined Work that displays copyright notices during execution, include the copyright notice for the Library among these notices, as well as a reference directing the user to the copies of the GNU GPL and this license document. d) Do one of the following: 0) Convey the Minimal Corresponding Source under the terms of this License, and the Corresponding Application Code in a form suitable for, and under terms that permit, the user to recombine or relink the Application with a modified version of the Linked Version to produce a modified Combined Work, in the manner specified by section 6 of the GNU GPL for conveying Corresponding Source. 1) Use a suitable shared library mechanism for linking with the Library. A suitable mechanism is one that (a) uses at run time a copy of the Library already present on the user's computer system, and (b) will operate properly with a modified version of the Library that is interface-compatible with the Linked Version. e) Provide Installation Information, but only if you would otherwise be required to provide such information under section 6 of the GNU GPL, and only to the extent that such information is necessary to install and execute a modified version of the Combined Work produced by recombining or relinking the Application with a modified version of the Linked Version. (If you use option 4d0, the Installation Information must accompany the Minimal Corresponding Source and Corresponding Application Code. If you use option 4d1, you must provide the Installation Information in the manner specified by section 6 of the GNU GPL for conveying Corresponding Source.) 5. Combined Libraries. You may place library facilities that are a work based on the Library side by side in a single library together with other library facilities that are not Applications and are not covered by this License, and convey such a combined library under terms of your choice, if you do both of the following: a) Accompany the combined library with a copy of the same work based on the Library, uncombined with any other library facilities, conveyed under the terms of this License. b) Give prominent notice with the combined library that part of it is a work based on the Library, and explaining where to find the accompanying uncombined form of the same work. 6. Revised Versions of the GNU Lesser General Public License. The Free Software Foundation may publish revised and/or new versions of the GNU Lesser General Public License from time to time. Such new versions will be similar in spirit to the present version, but may differ in detail to address new problems or concerns. Each version is given a distinguishing version number. If the Library as you received it specifies that a certain numbered version of the GNU Lesser General Public License \"or any later version\" applies to it, you have the option of following the terms and conditions either of that published version or of any later version published by the Free Software Foundation. If the Library as you received it does not specify a version number of the GNU Lesser General Public License, you may choose any version of the GNU Lesser General Public License ever published by the Free Software Foundation. If the Library as you received it specifies that a proxy can decide whether future versions of the GNU Lesser General Public License shall apply, that proxy's public statement of acceptance of any version is permanent authorization for you to choose that version for the Library. ",
"summary": "Unbalanced Optimal Transport for Object Detection",
"version": "0.3.post2",
"project_urls": {
"Documentation": "https://uotod.readthedocs.io/en/latest/",
"E-DUALITY": "https://www.esat.kuleuven.be/stadius/E/",
"ESAT-PSI": "https://www.esat.kuleuven.be/psi/",
"ESAT-STADIUS": "https://www.esat.kuleuven.be/stadius/",
"GitHub": "https://github.com/hdeplaen/uotod",
"Homepage": "https://github.com/hdeplaen/uotod",
"Issues": "https://github.com/hdeplaen/uotod/issues"
},
"split_keywords": [
"pytorch",
"machine learning"
],
"urls": [
{
"comment_text": "",
"digests": {
"blake2b_256": "9e4ca2b2ffbfeab9d498df00fe3797399524cae57f80e821b90726a360b4e30b",
"md5": "616013dc2b8c01e95be13f7995e17445",
"sha256": "66059564ea27141d301a43e422d070dbaaf010629faa5955f864951553bf5f54"
},
"downloads": -1,
"filename": "uotod-0.3.post2-py3-none-any.whl",
"has_sig": false,
"md5_digest": "616013dc2b8c01e95be13f7995e17445",
"packagetype": "bdist_wheel",
"python_version": "py3",
"requires_python": ">=3.8",
"size": 52929,
"upload_time": "2024-02-12T14:23:04",
"upload_time_iso_8601": "2024-02-12T14:23:04.966044Z",
"url": "https://files.pythonhosted.org/packages/9e/4c/a2b2ffbfeab9d498df00fe3797399524cae57f80e821b90726a360b4e30b/uotod-0.3.post2-py3-none-any.whl",
"yanked": false,
"yanked_reason": null
},
{
"comment_text": "",
"digests": {
"blake2b_256": "d3fe733285392e909725a2227c42bf65221e33f657d0f1efbe0526feb72626e7",
"md5": "a0f7f0c506c109fc696930dcb948f821",
"sha256": "d3bdfd3f887d5280e22e89efbe71d80483d59911162565e3ee282dd08643bc29"
},
"downloads": -1,
"filename": "uotod-0.3.post2.tar.gz",
"has_sig": false,
"md5_digest": "a0f7f0c506c109fc696930dcb948f821",
"packagetype": "sdist",
"python_version": "source",
"requires_python": ">=3.8",
"size": 1780420,
"upload_time": "2024-02-12T14:23:17",
"upload_time_iso_8601": "2024-02-12T14:23:17.617749Z",
"url": "https://files.pythonhosted.org/packages/d3/fe/733285392e909725a2227c42bf65221e33f657d0f1efbe0526feb72626e7/uotod-0.3.post2.tar.gz",
"yanked": false,
"yanked_reason": null
}
],
"upload_time": "2024-02-12 14:23:17",
"github": true,
"gitlab": false,
"bitbucket": false,
"codeberg": false,
"github_user": "hdeplaen",
"github_project": "uotod",
"travis_ci": false,
"coveralls": false,
"github_actions": true,
"requirements": [],
"lcname": "uotod"
}