topologysegmentation 0.1.0__tar.gz

topologysegmentation-0.1.0/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 proninegor
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

topologysegmentation-0.1.0/MANIFEST.in

@@ -0,0 +1,14 @@
+include LICENSE NOTICE README.md QUICKSTART.md RELEASING.md pyproject.toml
+recursive-include src/topologysegmentation *.py
+recursive-include src/topologysegmentation/_native *.c
+include src/topologysegmentation/VERSION
+include src/topologysegmentation/losses/vendor/tloss/LICENSE
+prune experiments
+prune tests
+prune configs
+prune splits
+prune runs
+prune logs
+prune checkpoints
+prune src/topologysegmentation/losses/vendor/gd
+global-exclude *.pt *.pth *.so *.dylib *.log *.pyc

topologysegmentation-0.1.0/NOTICE

@@ -0,0 +1,39 @@
+topologysegmentation
+Copyright (c) 2026 proninegor
+
+This product bundles third-party losses. Each file keeps its own copyright
+header. The project license (MIT, see LICENSE) covers the original code.
+The bundled Apache-2.0 file remains under Apache-2.0.
+
+Apache-2.0
+  src/topologysegmentation/losses/vendor/tloss/tloss.py
+  Copyright 2023 University of Basel and Lucerne University of Applied
+  Sciences and Arts Authors
+  https://github.com/Digital-Dermatology/t-loss
+  The license text is in src/topologysegmentation/losses/vendor/tloss/LICENSE
+
+MIT
+  src/topologysegmentation/losses/vendor/cldice/
+  Suprosanna Shit, Johannes C. Paetzold et al.
+  https://github.com/jocpae/clDice
+  https://arxiv.org/abs/2003.07311
+
+  src/topologysegmentation/losses/vendor/bm/
+  Nico Stucki, Johannes C. Paetzold, Suprosanna Shit, Bjoern Menze, Ulrich Bauer
+  https://github.com/nstucki/Betti-matching
+  https://proceedings.mlr.press/v202/stucki23a
+
+  src/topologysegmentation/losses/vendor/sat/
+  Bo Wen, Haochen Zhang, Dirk-Uwe G. Bartsch, William R. Freeman,
+  Truong Q. Nguyen, Cheolhong An
+  https://github.com/JRC-VPLab/SATLoss
+  https://arxiv.org/abs/2412.02076
+  Cubical persistence in this folder is adapted from torch-topological
+  (https://github.com/aidos-lab/pytorch-topological) and calls GUDHI.
+
+Not included in the wheel
+  src/topologysegmentation/losses/vendor/gd/ has no license grant in the source
+  file, so it is not packaged. It is an unregistered 3D Generalized Dice
+  (Sudre et al., https://arxiv.org/abs/1707.03237).
+  The C++ extension for bm3d is not vendored. Build
+  https://github.com/nstucki/Betti-Matching-3D (MIT) separately.

topologysegmentation-0.1.0/PKG-INFO

@@ -0,0 +1,139 @@
+Metadata-Version: 2.4
+Name: topologysegmentation
+Version: 0.1.0
+Summary: Losses and metrics for topology-aware image segmentation
+Author-email: proninegor <pr0n1n3@yandex.ru>
+License-Expression: MIT AND Apache-2.0
+Project-URL: Homepage, https://github.com/ProninEgorDm/TopologyLosses
+Project-URL: Repository, https://github.com/ProninEgorDm/TopologyLosses
+Project-URL: Issues, https://github.com/ProninEgorDm/TopologyLosses/issues
+Keywords: segmentation,topology,persistent-homology,pytorch,loss
+Classifier: Development Status :: 3 - Alpha
+Classifier: Intended Audience :: Science/Research
+Classifier: Operating System :: OS Independent
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Topic :: Scientific/Engineering :: Image Processing
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Requires-Python: <3.13,>=3.12
+Description-Content-Type: text/markdown
+License-File: LICENSE
+License-File: NOTICE
+License-File: src/topologysegmentation/losses/vendor/tloss/LICENSE
+Requires-Dist: numpy>=2.2
+Requires-Dist: torch>=2.2
+Provides-Extra: ph
+Requires-Dist: gudhi>=3.13.0; extra == "ph"
+Requires-Dist: pot>=0.9.7.post1; extra == "ph"
+Requires-Dist: torch-topological>=0.1.9; extra == "ph"
+Provides-Extra: plot
+Requires-Dist: panel>=1.8.10; extra == "plot"
+Provides-Extra: all
+Requires-Dist: gudhi>=3.13.0; extra == "all"
+Requires-Dist: pot>=0.9.7.post1; extra == "all"
+Requires-Dist: torch-topological>=0.1.9; extra == "all"
+Requires-Dist: panel>=1.8.10; extra == "all"
+Dynamic: license-file
+
+# topologysegmentation
+
+Losses and metrics for topology-aware image segmentation.
+
+```bash
+pip install topologysegmentation          # Dice, clDice, pooling, critical points, T-Loss
+pip install "topologysegmentation[ph]"    # also Hu, Betti matching, and SAT
+```
+
+```python
+import torch
+from topologysegmentation import DiceLoss, create_loss, loss_shapes, topology_errors
+
+logits = torch.randn(2, 1, 64, 64)   # frame (N, 1, H, W) or volume (N, 1, D, H, W)
+target = torch.zeros(2, 1, 64, 64)
+target[:, :, 16:48, 16:48] = 1
+
+dice = DiceLoss()
+cldice = create_loss("cldice")
+value = dice(logits, target) + 0.1 * cldice(logits, target)
+report = topology_errors(logits.sigmoid(), target)
+loss_shapes()   # {"dice": (2, 3), "sat": (2,), ...}
+```
+
+Every loss applies sigmoid itself. Pass logits. `topology_errors` is eval-only and expects probabilities. Python 3.12 is required. `bm3d` needs a local build of [Betti-Matching-3D](https://github.com/nstucki/Betti-Matching-3D): put its `betti_matching` module on `PYTHONPATH`.
+
+Third-party licenses are listed in [NOTICE](https://github.com/ProninEgorDm/TopologyLosses/blob/main/NOTICE). Training this repository is described in [QUICKSTART.md](https://github.com/ProninEgorDm/TopologyLosses/blob/main/QUICKSTART.md).
+
+## Installation
+
+The base package depends only on `torch` and `numpy`. Homology lives in the `ph` extra.
+
+| Command | What you get |
+| --- | --- |
+| `pip install topologysegmentation` | `dice`, `bce`, `focal`, `combined`, `cldice`, `tloss`, `pooling`, `critical`, Dice/IoU, centerline Dice |
+| `pip install "topologysegmentation[ph]"` | also `hu`, `bm`, `sat`, and Betti numbers |
+| `pip install "topologysegmentation[plot]"` | plots for Betti matching |
+| `pip install "topologysegmentation[all]"` | both extras |
+
+## Shapes
+
+A frame is `(N, 1, H, W)`. A volume is `(N, 1, D, H, W)`. One frame may be `(H, W)` or a channel-free batch `(N, H, W)`. A volume keeps the channel axis: `(D, H, W)` is read as a batch of frames. A side of length 1 is still that axis: a `1×N` strip is a frame, and a slab of depth 1 is a volume.
+
+`loss_shapes()` maps each trainable name to the spatial ranks it accepts. `2` is a frame, `3` is a volume.
+
+| Ranks | Losses |
+| --- | --- |
+| 2 and 3 | `dice`, `bce`, `focal`, `combined`, `cldice`, `tloss`, `pooling`, `critical`, `hu`, `bm3d` |
+| 2 | `sat`, `bm` |
+
+`sat` and `bm` raise `ValueError` on a volume. For a volume, Betti matching is `bm3d`.
+
+## Losses
+
+Every loss is called the same way: `create_loss(name)(logits, target)`.
+
+| Name | What it does | Dependency | Weight next to Dice |
+| --- | --- | --- | --- |
+| `dice`, `bce`, `focal`, `combined` | Overlap and per-pixel losses. Any spatial rank | package | 1 |
+| `cldice` | Soft skeleton, Shit et al., [arXiv:2003.07311](https://arxiv.org/abs/2003.07311). Ten pooling iterations, frames and volumes | package | 0.1–0.3 |
+| `tloss` | Robust Student-t likelihood for a noisy mask, per voxel. `nu` is learned. Any shape | package | 0.1 |
+| `pooling` | Multi-scale occupancy after max-pooling. A frame adds two axis projections, a volume adds three | package | 1 |
+| `critical` | Endpoints, junctions, and isolated voxels. Eight neighbors on a frame, twenty-six on a volume | package | 1 |
+| `hu` | Diagram matching toward birth 1 and death 0, extra bars toward the diagonal. Hu et al., NeurIPS 2019. Frames and volumes | `ph` | 0.1 |
+| `bm` | Betti matching, Stucki et al., ICML 2023. Frames | `ph` | 0.1 |
+| `bm3d` | The same objective through the C++ extension. Frames and volumes | compiled `betti_matching` | 0.1 |
+| `sat` | Spatially aware diagram matching, divided by the image side length. Frames | `ph` | 0.1 |
+
+Foreground voxels that touch at a corner are one component. `betti_numbers` on a frame is `(β0, β1)`. On a volume it is `(β0, β1, β2)`. An empty mask is all zeros. A single foreground voxel is `(1, 0)` on a frame and `(1, 0, 0)` on a volume.
+
+## Speed
+
+One CPU (Apple M1), forward and backward together. Frames are VessMAP labels (`annotator1`, native 256×256), resized with nearest-neighbor interpolation. The prediction is that same mask dilated by one pixel, with 12% of the vessel pixels removed, passed as logits ±4. A leading number is the batch: `3×1024²` is three such frames. `sat` and `bm` are frames only, so this table has no volumes.
+
+| Loss | 256² | 512² | 1024² | 3×1024² | 4×256² |
+| --- | ---: | ---: | ---: | ---: | ---: |
+| `dice` | 0.4 ms | 0.6 ms | 3.1 ms | 7.8 ms | 0.7 ms |
+| `bce` | 0.5 ms | 1.5 ms | 4.7 ms | 13 ms | 1.0 ms |
+| `focal` | 1.2 ms | 2.8 ms | 12 ms | 35 ms | 3.1 ms |
+| `combined` | 1.1 ms | 1.9 ms | 7.2 ms | 25 ms | 2.0 ms |
+| `tloss` | 0.6 ms | 1.0 ms | 3.2 ms | 8.7 ms | 0.8 ms |
+| `pooling` | 6.7 ms | 25 ms | 0.11 s | 0.15 s | 9.3 ms |
+| `critical` | 3.3 ms | 9.3 ms | 47 ms | 0.12 s | 7.7 ms |
+| `cldice` | 99 ms | 0.41 s | 1.9 s | 5.3 s | 0.32 s |
+| `hu` | 0.26 s | 1.1 s | 4.5 s | 14 s | 1.0 s |
+| `sat` | 0.24 s | 1.0 s | 4.2 s | 13 s | 0.98 s |
+| `bm` | 0.64 s | 4.4 s | 73 s | 4.8 min | 2.8 s |
+
+These masks have a handful of components and loops (for example `(β0, β1)` of `(5, 0)`, `(4, 7)`, `(19, 4)`). Times up to 512² are the median of three passes. At 1024² and for `3×1024²`, `cldice`, `hu`, `sat`, and `bm` are one pass after a warmup.
+
+## Metrics
+
+`dice_iou(probabilities, target)` returns per-batch Dice and IoU after a 0.5 threshold. Frames and volumes both work. On a mask compared with itself the score is 1, including a single voxel.
+
+`centerline_dice(probabilities, target)` is the Dice score between morphological skeletons.
+
+`betti_numbers(mask)` is `(β0, β1)` or `(β0, β1, β2)`. The binary count does not import `gudhi`. `hu` does, through the `ph` extra, because it needs the critical voxels of the diagram.
+
+`topology_errors(probabilities, target)` reports the mean absolute Betti error (including `betti_b2_mae` on a volume), the centerline Dice, and a per-image breakdown. This is a validation report, not a loss. A frame does not grow a `β2` term when one side is 1; a one-slice volume does.
+
+## License
+
+Original code is MIT, see [LICENSE](https://github.com/ProninEgorDm/TopologyLosses/blob/main/LICENSE). T-Loss stays under Apache-2.0, so the package license expression is `MIT AND Apache-2.0`. The file table is in [NOTICE](https://github.com/ProninEgorDm/TopologyLosses/blob/main/NOTICE).

topologysegmentation-0.1.0/QUICKSTART.md

@@ -0,0 +1,68 @@
+# Быстрый старт
+
+Две разные установки. Колесо с PyPI — это лоссы. Команды `train.py` есть только в клоне репозитория.
+
+## Пакет
+
+Нужен Python 3.12.
+
+```bash
+pip install topologysegmentation
+pip install "topologysegmentation[ph]"
+```
+
+Первая команда ставит Dice, clDice, T-Loss, `pooling` и `critical`. Вторая добавляет `gudhi`, `pot` и `torch-topological` для `hu`, `bm`, `sat` и чисел Бетти.
+
+```python
+from topologysegmentation import DiceLoss, create_loss, loss_shapes
+
+# Кадр (N, 1, H, W) или объём (N, 1, D, H, W). Сигмоида считается внутри лосса.
+loss = DiceLoss()(logits, target) + 0.1 * create_loss("cldice")(logits, target)
+loss_shapes()  # у каждого имени набор размерностей: 2 — кадр, 3 — объём
+```
+
+`dice`, `bce`, `focal`, `combined`, `cldice`, `tloss`, `pooling`, `critical`, `hu` и `bm3d` принимают кадр и объём. `sat` и `bm` только кадр: на объёме они бросают `ValueError`, для объёма есть `bm3d`. Тензор `(D, H, W)` без канала читается как батч плоских кадров.
+
+`betti_numbers` на кадре возвращает `(β0, β1)`, на объёме `(β0, β1, β2)`. При первом вызове, если найден `clang`, `cc` или `gcc`, собирается `cubical_betti.c` из колеса. Без компилятора считается запасной путь на NumPy. `TOPOLOGYSEGMENTATION_NATIVE=0` оставляет только его. `bm` так же собирает `cubical_matching.c`; `TOPOLOGYSEGMENTATION_MATCHING=0` оставляет для него питоновские циклы.
+
+Скорость на текущей машине, вместе с backward:
+
+```bash
+python tests/bench_losses.py
+```
+
+`bm3d` этой командой не ставится. Расширение `betti_matching` собирается из [Betti-Matching-3D](https://github.com/nstucki/Betti-Matching-3D), каталог `build` добавляется в `PYTHONPATH`.
+
+Полный список лоссов, весов и лицензий — в [README.md](README.md).
+
+## Клон репозитория
+
+`uv sync` ставит пакет в окружение `.venv` и группу `dev` из `pyproject.toml`: трансформер, чтение кадров, тензорборд. Эти пакеты в колесо не входят.
+
+```bash
+uv sync
+python tests/test_contract.py
+python train.py --config configs/experiments/dice.yaml --max-steps 1 --image-size 64 --batch-size 2 --epochs 1
+```
+
+Полный прогон:
+
+```bash
+python train.py --config configs/experiments/dice.yaml
+```
+
+Сравнение конфигов из списка `DEFAULT_CONFIGS` в `experiments/sweep.py`:
+
+```bash
+python train.py sweep --max-steps 1 --image-size 64 --batch-size 2
+```
+
+Согласие двух аннотаторов на общих кадрах:
+
+```bash
+python train.py agreement
+```
+
+Кадры VessMAP лежат в `src/VessMAP` и в git не входят. Сплит с сидом 42 — `splits/vessmap_seed42.json`: 80 обучающих, 12 валидационных, 8 тестовых. Каталог прогона `runs/<время>_<имя>/` содержит копию конфига, сплит, `history.json` и `checkpoints/best.pt`. Поле `parts` в истории — уже взвешенный вклад каждого лосса.
+
+Конфиги новых лоссов: `configs/experiments/dice_hu.yaml`, `dice_pooling.yaml`, `dice_critical.yaml`. Веса там уже подобраны под масштаб лосса, см. таблицу в README.

topologysegmentation-0.1.0/README.md

@@ -0,0 +1,103 @@
+# topologysegmentation
+
+Losses and metrics for topology-aware image segmentation.
+
+```bash
+pip install topologysegmentation          # Dice, clDice, pooling, critical points, T-Loss
+pip install "topologysegmentation[ph]"    # also Hu, Betti matching, and SAT
+```
+
+```python
+import torch
+from topologysegmentation import DiceLoss, create_loss, loss_shapes, topology_errors
+
+logits = torch.randn(2, 1, 64, 64)   # frame (N, 1, H, W) or volume (N, 1, D, H, W)
+target = torch.zeros(2, 1, 64, 64)
+target[:, :, 16:48, 16:48] = 1
+
+dice = DiceLoss()
+cldice = create_loss("cldice")
+value = dice(logits, target) + 0.1 * cldice(logits, target)
+report = topology_errors(logits.sigmoid(), target)
+loss_shapes()   # {"dice": (2, 3), "sat": (2,), ...}
+```
+
+Every loss applies sigmoid itself. Pass logits. `topology_errors` is eval-only and expects probabilities. Python 3.12 is required. `bm3d` needs a local build of [Betti-Matching-3D](https://github.com/nstucki/Betti-Matching-3D): put its `betti_matching` module on `PYTHONPATH`.
+
+Third-party licenses are listed in [NOTICE](https://github.com/ProninEgorDm/TopologyLosses/blob/main/NOTICE). Training this repository is described in [QUICKSTART.md](https://github.com/ProninEgorDm/TopologyLosses/blob/main/QUICKSTART.md).
+
+## Installation
+
+The base package depends only on `torch` and `numpy`. Homology lives in the `ph` extra.
+
+| Command | What you get |
+| --- | --- |
+| `pip install topologysegmentation` | `dice`, `bce`, `focal`, `combined`, `cldice`, `tloss`, `pooling`, `critical`, Dice/IoU, centerline Dice |
+| `pip install "topologysegmentation[ph]"` | also `hu`, `bm`, `sat`, and Betti numbers |
+| `pip install "topologysegmentation[plot]"` | plots for Betti matching |
+| `pip install "topologysegmentation[all]"` | both extras |
+
+## Shapes
+
+A frame is `(N, 1, H, W)`. A volume is `(N, 1, D, H, W)`. One frame may be `(H, W)` or a channel-free batch `(N, H, W)`. A volume keeps the channel axis: `(D, H, W)` is read as a batch of frames. A side of length 1 is still that axis: a `1×N` strip is a frame, and a slab of depth 1 is a volume.
+
+`loss_shapes()` maps each trainable name to the spatial ranks it accepts. `2` is a frame, `3` is a volume.
+
+| Ranks | Losses |
+| --- | --- |
+| 2 and 3 | `dice`, `bce`, `focal`, `combined`, `cldice`, `tloss`, `pooling`, `critical`, `hu`, `bm3d` |
+| 2 | `sat`, `bm` |
+
+`sat` and `bm` raise `ValueError` on a volume. For a volume, Betti matching is `bm3d`.
+
+## Losses
+
+Every loss is called the same way: `create_loss(name)(logits, target)`.
+
+| Name | What it does | Dependency | Weight next to Dice |
+| --- | --- | --- | --- |
+| `dice`, `bce`, `focal`, `combined` | Overlap and per-pixel losses. Any spatial rank | package | 1 |
+| `cldice` | Soft skeleton, Shit et al., [arXiv:2003.07311](https://arxiv.org/abs/2003.07311). Ten pooling iterations, frames and volumes | package | 0.1–0.3 |
+| `tloss` | Robust Student-t likelihood for a noisy mask, per voxel. `nu` is learned. Any shape | package | 0.1 |
+| `pooling` | Multi-scale occupancy after max-pooling. A frame adds two axis projections, a volume adds three | package | 1 |
+| `critical` | Endpoints, junctions, and isolated voxels. Eight neighbors on a frame, twenty-six on a volume | package | 1 |
+| `hu` | Diagram matching toward birth 1 and death 0, extra bars toward the diagonal. Hu et al., NeurIPS 2019. Frames and volumes | `ph` | 0.1 |
+| `bm` | Betti matching, Stucki et al., ICML 2023. Frames | `ph` | 0.1 |
+| `bm3d` | The same objective through the C++ extension. Frames and volumes | compiled `betti_matching` | 0.1 |
+| `sat` | Spatially aware diagram matching, divided by the image side length. Frames | `ph` | 0.1 |
+
+Foreground voxels that touch at a corner are one component. `betti_numbers` on a frame is `(β0, β1)`. On a volume it is `(β0, β1, β2)`. An empty mask is all zeros. A single foreground voxel is `(1, 0)` on a frame and `(1, 0, 0)` on a volume.
+
+## Speed
+
+One CPU (Apple M1), forward and backward together. Frames are VessMAP labels (`annotator1`, native 256×256), resized with nearest-neighbor interpolation. The prediction is that same mask dilated by one pixel, with 12% of the vessel pixels removed, passed as logits ±4. A leading number is the batch: `3×1024²` is three such frames. `sat` and `bm` are frames only, so this table has no volumes.
+
+| Loss | 256² | 512² | 1024² | 3×1024² | 4×256² |
+| --- | ---: | ---: | ---: | ---: | ---: |
+| `dice` | 0.4 ms | 0.6 ms | 3.1 ms | 7.8 ms | 0.7 ms |
+| `bce` | 0.5 ms | 1.5 ms | 4.7 ms | 13 ms | 1.0 ms |
+| `focal` | 1.2 ms | 2.8 ms | 12 ms | 35 ms | 3.1 ms |
+| `combined` | 1.1 ms | 1.9 ms | 7.2 ms | 25 ms | 2.0 ms |
+| `tloss` | 0.6 ms | 1.0 ms | 3.2 ms | 8.7 ms | 0.8 ms |
+| `pooling` | 6.7 ms | 25 ms | 0.11 s | 0.15 s | 9.3 ms |
+| `critical` | 3.3 ms | 9.3 ms | 47 ms | 0.12 s | 7.7 ms |
+| `cldice` | 99 ms | 0.41 s | 1.9 s | 5.3 s | 0.32 s |
+| `hu` | 0.26 s | 1.1 s | 4.5 s | 14 s | 1.0 s |
+| `sat` | 0.24 s | 1.0 s | 4.2 s | 13 s | 0.98 s |
+| `bm` | 0.64 s | 4.4 s | 73 s | 4.8 min | 2.8 s |
+
+These masks have a handful of components and loops (for example `(β0, β1)` of `(5, 0)`, `(4, 7)`, `(19, 4)`). Times up to 512² are the median of three passes. At 1024² and for `3×1024²`, `cldice`, `hu`, `sat`, and `bm` are one pass after a warmup.
+
+## Metrics
+
+`dice_iou(probabilities, target)` returns per-batch Dice and IoU after a 0.5 threshold. Frames and volumes both work. On a mask compared with itself the score is 1, including a single voxel.
+
+`centerline_dice(probabilities, target)` is the Dice score between morphological skeletons.
+
+`betti_numbers(mask)` is `(β0, β1)` or `(β0, β1, β2)`. The binary count does not import `gudhi`. `hu` does, through the `ph` extra, because it needs the critical voxels of the diagram.
+
+`topology_errors(probabilities, target)` reports the mean absolute Betti error (including `betti_b2_mae` on a volume), the centerline Dice, and a per-image breakdown. This is a validation report, not a loss. A frame does not grow a `β2` term when one side is 1; a one-slice volume does.
+
+## License
+
+Original code is MIT, see [LICENSE](https://github.com/ProninEgorDm/TopologyLosses/blob/main/LICENSE). T-Loss stays under Apache-2.0, so the package license expression is `MIT AND Apache-2.0`. The file table is in [NOTICE](https://github.com/ProninEgorDm/TopologyLosses/blob/main/NOTICE).

topologysegmentation-0.1.0/RELEASING.md

@@ -0,0 +1,49 @@
+# Выкладка на PyPI
+
+Колесо содержит только пакет `topologysegmentation`. Обучение VessMAP, тесты, датасет и сборка `betti_matching` остаются в git.
+
+Имя на PyPI — `topologysegmentation`, импорт тот же: `import topologysegmentation`. Имя `topologylosses` не использовать: на PyPI уже есть проект `topolosses`.
+
+## Версия
+
+Номер записан одной строкой в `src/topologysegmentation/VERSION`. Его читают и пакет, и сборка. Первая публикация — `0.1.0`. Повторно залить ту же версию нельзя: перед следующей сборкой строку нужно увеличить.
+
+## Сборка
+
+Из корня репозитория, Python 3.12:
+
+```bash
+uv build
+uvx twine check dist/*
+```
+
+Появляются `dist/topologysegmentation-0.1.0.tar.gz` и `dist/topologysegmentation-0.1.0-py3-none-any.whl`. Каталог `dist/` в git не входит.
+
+В архиве исходников и в колесе должны быть `LICENSE`, `NOTICE`, `src/topologysegmentation/losses/vendor/tloss/LICENSE`, `src/topologysegmentation/_native/cubical_betti.c` и `src/topologysegmentation/_native/cubical_matching.c`. Не должно быть `experiments/`, `configs/`, `tests/`, `runs/`, `VessMAP`, `Betti-Matching-3D`, собранных библиотек `cubical_betti_*.dylib` и `cubical_matching_*.dylib` и `losses/vendor/gd`.
+
+Колесо остаётся `py3-none-any`: компилятор для `pip install` не нужен. `cubical_betti.c` собирается при первом вызове `betti_numbers`, `cubical_matching.c` — при первом вызове `bm`, если в системе есть `clang`, `cc` или `gcc`. Без компилятора работают запасные пути на Python.
+
+```bash
+tar -tzf dist/topologysegmentation-0.1.0.tar.gz
+unzip -l dist/topologysegmentation-0.1.0-py3-none-any.whl
+```
+
+## Публикация
+
+Нужен токен со страницы https://pypi.org/manage/account/token/ с областью на весь аккаунт или на проект `topologysegmentation`. Токен TestPyPI на основной индекс не действует.
+
+```bash
+uvx twine upload dist/*
+```
+
+Проверка во временном окружении:
+
+```bash
+uv venv /tmp/tl-smoke --python 3.12
+uv pip install --python /tmp/tl-smoke/bin/python topologysegmentation
+/tmp/tl-smoke/bin/python -c "import topologysegmentation; print(topologysegmentation.__version__, topologysegmentation.available_losses())"
+```
+
+`uv venv` не кладёт в окружение исполняемый `pip`, поэтому ставить пакеты нужно через `uv pip --python`.
+
+Страница проекта берёт текст из `README.md`. `pip install topologysegmentation` не ставит `bm3d`: расширение собирается отдельно, лицензия MIT указана в `NOTICE`.

topologysegmentation-0.1.0/pyproject.toml

@@ -0,0 +1,94 @@
+[build-system]
+requires = ["setuptools>=77"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "topologysegmentation"
+dynamic = ["version"]
+description = "Losses and metrics for topology-aware image segmentation"
+readme = "README.md"
+requires-python = ">=3.12,<3.13"
+license = "MIT AND Apache-2.0"
+license-files = [
+    "LICENSE",
+    "NOTICE",
+    "src/topologysegmentation/losses/vendor/tloss/LICENSE",
+]
+authors = [{ name = "proninegor", email = "pr0n1n3@yandex.ru" }]
+keywords = ["segmentation", "topology", "persistent-homology", "pytorch", "loss"]
+classifiers = [
+    "Development Status :: 3 - Alpha",
+    "Intended Audience :: Science/Research",
+    "Operating System :: OS Independent",
+    "Programming Language :: Python :: 3.12",
+    "Topic :: Scientific/Engineering :: Image Processing",
+    "Topic :: Scientific/Engineering :: Artificial Intelligence",
+]
+dependencies = [
+    "numpy>=2.2",
+    "torch>=2.2",
+]
+
+# ph: лоссы и метрики на кубической гомологии (hu, bm, sat, числа Бетти).
+# plot: графики Betti matching. Сам лосс bm их не импортирует.
+[project.optional-dependencies]
+ph = [
+    "gudhi>=3.13.0",
+    "pot>=0.9.7.post1",
+    "torch-topological>=0.1.9",
+]
+plot = [
+    "panel>=1.8.10",
+]
+all = [
+    "gudhi>=3.13.0",
+    "pot>=0.9.7.post1",
+    "torch-topological>=0.1.9",
+    "panel>=1.8.10",
+]
+
+[project.urls]
+Homepage = "https://github.com/ProninEgorDm/TopologyLosses"
+Repository = "https://github.com/ProninEgorDm/TopologyLosses"
+Issues = "https://github.com/ProninEgorDm/TopologyLosses/issues"
+
+# Версия пакета — одна строка в src/topologysegmentation/VERSION.
+# Файл читается при сборке без импорта torch.
+[tool.setuptools.dynamic]
+version = { file = "src/topologysegmentation/VERSION" }
+
+[tool.setuptools.packages.find]
+where = ["src"]
+include = ["topologysegmentation*"]
+# У GDLoss в файле нет лицензии, в колесо он не входит.
+exclude = ["topologysegmentation.losses.vendor.gd*"]
+
+[tool.setuptools.package-data]
+topologysegmentation = ["VERSION", "losses/vendor/tloss/LICENSE", "_native/*.c"]
+
+# Локальный репозиторий: uv sync ставит эту группу вместе с пакетом.
+# В колесо PyPI она не попадает.
+[dependency-groups]
+dev = [
+    "einops>=0.8.2",
+    "gudhi>=3.13.0",
+    "ipykernel>=7.2.0",
+    "monai>=1.5.2",
+    "opencv-python>=4.13.0.92",
+    "panel>=1.8.10",
+    "pixi>=1.0.1",
+    "pot>=0.9.7.post1",
+    "pyyaml>=6.0.3",
+    "scikit-image>=0.25.2",
+    "scikit-learn>=1.7.2",
+    "seaborn>=0.13.2",
+    "tensorboard>=2.20.0",
+    "thop>=0.1.1",
+    "tifffile>=2025.5.10",
+    "torch-topological>=0.1.9",
+    "torchvision>=0.26.0",
+    "tqdm>=4.67.3",
+]
+
+[tool.uv]
+default-groups = ["dev"]

topologysegmentation-0.1.0/setup.cfg

@@ -0,0 +1,4 @@
+[egg_info]
+tag_build = 
+tag_date = 0
+

topologysegmentation-0.1.0/src/topologysegmentation/VERSION

@@ -0,0 +1 @@
+0.1.0

topologysegmentation-0.1.0/src/topologysegmentation/__init__.py

@@ -0,0 +1,27 @@
+"""Losses and metrics for topology-aware segmentation.
+
+The VessMAP training pipeline lives in ``experiments/`` and is not installed
+with this package. The release number is the single line in ``VERSION``.
+"""
+
+from pathlib import Path
+
+__version__ = Path(__file__).with_name("VERSION").read_text(encoding="utf-8").strip()
+
+from topologysegmentation.losses.registry import available_losses, create_loss, loss_shapes
+from topologysegmentation.losses.segmentation import BCELoss, CombinedLoss, DiceLoss, FocalLoss
+from topologysegmentation.metrics import betti_numbers, centerline_dice, dice_iou, topology_errors
+
+__all__ = [
+    "BCELoss",
+    "CombinedLoss",
+    "DiceLoss",
+    "FocalLoss",
+    "available_losses",
+    "betti_numbers",
+    "centerline_dice",
+    "create_loss",
+    "dice_iou",
+    "loss_shapes",
+    "topology_errors",
+]