singlepipe 0.0.1__tar.gz
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- singlepipe-0.0.1/LICENSE +21 -0
- singlepipe-0.0.1/PKG-INFO +135 -0
- singlepipe-0.0.1/README.md +102 -0
- singlepipe-0.0.1/pyproject.toml +70 -0
- singlepipe-0.0.1/setup.cfg +4 -0
- singlepipe-0.0.1/singlepipe/__init__.py +8 -0
- singlepipe-0.0.1/singlepipe/adapters/__init__.py +0 -0
- singlepipe-0.0.1/singlepipe/adapters/hf_adapter.py +119 -0
- singlepipe-0.0.1/singlepipe/model_types/__init__.py +3 -0
- singlepipe-0.0.1/singlepipe/model_types/types.py +79 -0
- singlepipe-0.0.1/singlepipe/schemas/__init__.py +25 -0
- singlepipe-0.0.1/singlepipe/schemas/audio_schemas/__init__.py +0 -0
- singlepipe-0.0.1/singlepipe/schemas/audio_schemas/audio_classification.py +40 -0
- singlepipe-0.0.1/singlepipe/schemas/audio_schemas/speech_to_text.py +40 -0
- singlepipe-0.0.1/singlepipe/schemas/audio_schemas/text_to_audio.py +40 -0
- singlepipe-0.0.1/singlepipe/schemas/audio_schemas/zero_shot_audio_classification.py +40 -0
- singlepipe-0.0.1/singlepipe/schemas/image_schemas/__init__.py +0 -0
- singlepipe-0.0.1/singlepipe/schemas/image_schemas/depth_estimation.py +40 -0
- singlepipe-0.0.1/singlepipe/schemas/image_schemas/image_classification.py +40 -0
- singlepipe-0.0.1/singlepipe/schemas/image_schemas/image_embedding.py +40 -0
- singlepipe-0.0.1/singlepipe/schemas/image_schemas/image_segmentation.py +40 -0
- singlepipe-0.0.1/singlepipe/schemas/image_schemas/keypoint_matching.py +40 -0
- singlepipe-0.0.1/singlepipe/schemas/image_schemas/mask_generation.py +40 -0
- singlepipe-0.0.1/singlepipe/schemas/image_schemas/object_detection.py +290 -0
- singlepipe-0.0.1/singlepipe/schemas/image_schemas/zero_shot_image_classification.py +40 -0
- singlepipe-0.0.1/singlepipe/schemas/image_schemas/zero_shot_object_detection.py +40 -0
- singlepipe-0.0.1/singlepipe/schemas/multimodal_schemas/__init__.py +0 -0
- singlepipe-0.0.1/singlepipe/schemas/multimodal_schemas/document_question_answering.py +40 -0
- singlepipe-0.0.1/singlepipe/schemas/multimodal_schemas/image_text_to_text.py +40 -0
- singlepipe-0.0.1/singlepipe/schemas/multimodal_schemas/table_question_answering.py +40 -0
- singlepipe-0.0.1/singlepipe/schemas/ocr_schemas/__init__.py +0 -0
- singlepipe-0.0.1/singlepipe/schemas/ocr_schemas/ocr.py +40 -0
- singlepipe-0.0.1/singlepipe/schemas/registry.py +97 -0
- singlepipe-0.0.1/singlepipe/schemas/table_schemas/__init__.py +0 -0
- singlepipe-0.0.1/singlepipe/schemas/table_schemas/tabular_classification.py +40 -0
- singlepipe-0.0.1/singlepipe/schemas/table_schemas/tabular_regression.py +40 -0
- singlepipe-0.0.1/singlepipe/schemas/text_schemas/__init__.py +0 -0
- singlepipe-0.0.1/singlepipe/schemas/text_schemas/fill_mask.py +40 -0
- singlepipe-0.0.1/singlepipe/schemas/text_schemas/question_answering.py +40 -0
- singlepipe-0.0.1/singlepipe/schemas/text_schemas/summarization.py +40 -0
- singlepipe-0.0.1/singlepipe/schemas/text_schemas/text2text_generation.py +40 -0
- singlepipe-0.0.1/singlepipe/schemas/text_schemas/text_classification.py +40 -0
- singlepipe-0.0.1/singlepipe/schemas/text_schemas/text_embedding.py +40 -0
- singlepipe-0.0.1/singlepipe/schemas/text_schemas/text_generation.py +40 -0
- singlepipe-0.0.1/singlepipe/schemas/text_schemas/token_classification.py +40 -0
- singlepipe-0.0.1/singlepipe/schemas/text_schemas/translation.py +40 -0
- singlepipe-0.0.1/singlepipe/schemas/text_schemas/zero_shot_text_classification.py +40 -0
- singlepipe-0.0.1/singlepipe/schemas/video_schemas/__init__.py +0 -0
- singlepipe-0.0.1/singlepipe/schemas/video_schemas/video_classification.py +40 -0
- singlepipe-0.0.1/singlepipe.egg-info/PKG-INFO +135 -0
- singlepipe-0.0.1/singlepipe.egg-info/SOURCES.txt +52 -0
- singlepipe-0.0.1/singlepipe.egg-info/dependency_links.txt +1 -0
- singlepipe-0.0.1/singlepipe.egg-info/requires.txt +16 -0
- singlepipe-0.0.1/singlepipe.egg-info/top_level.txt +1 -0
singlepipe-0.0.1/LICENSE
ADDED
|
@@ -0,0 +1,21 @@
|
|
|
1
|
+
MIT License
|
|
2
|
+
|
|
3
|
+
Copyright (c) 2026 Ovcharov Gleb
|
|
4
|
+
|
|
5
|
+
Permission is hereby granted, free of charge, to any person obtaining a copy
|
|
6
|
+
of this software and associated documentation files (the "Software"), to deal
|
|
7
|
+
in the Software without restriction, including without limitation the rights
|
|
8
|
+
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
|
|
9
|
+
copies of the Software, and to permit persons to whom the Software is furnished
|
|
10
|
+
to do so, subject to the following conditions:
|
|
11
|
+
|
|
12
|
+
The above copyright notice and this permission notice shall be included in all
|
|
13
|
+
copies or substantial portions of the Software.
|
|
14
|
+
|
|
15
|
+
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
|
16
|
+
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
|
17
|
+
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
|
|
18
|
+
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
|
|
19
|
+
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
|
|
20
|
+
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
|
|
21
|
+
SOFTWARE.
|
|
@@ -0,0 +1,135 @@
|
|
|
1
|
+
Metadata-Version: 2.4
|
|
2
|
+
Name: singlepipe
|
|
3
|
+
Version: 0.0.1
|
|
4
|
+
Summary: Typed interface for ML model inference across HuggingFace, PyTorch, and other backends
|
|
5
|
+
Author-email: Gleb <glebov274@gmail.com>
|
|
6
|
+
License-Expression: MIT
|
|
7
|
+
Keywords: ml,inference,huggingface,pytorch,pydantic,typed
|
|
8
|
+
Classifier: Development Status :: 2 - Pre-Alpha
|
|
9
|
+
Classifier: Intended Audience :: Developers
|
|
10
|
+
Classifier: Programming Language :: Python :: 3
|
|
11
|
+
Classifier: Programming Language :: Python :: 3.11
|
|
12
|
+
Classifier: Programming Language :: Python :: 3.12
|
|
13
|
+
Classifier: Programming Language :: Python :: 3.13
|
|
14
|
+
Classifier: Programming Language :: Python :: 3.14
|
|
15
|
+
Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
|
|
16
|
+
Requires-Python: >=3.11
|
|
17
|
+
Description-Content-Type: text/markdown
|
|
18
|
+
License-File: LICENSE
|
|
19
|
+
Requires-Dist: httpx>=0.28.1
|
|
20
|
+
Requires-Dist: pillow>=12.3.0
|
|
21
|
+
Requires-Dist: pydantic>=2.13.4
|
|
22
|
+
Requires-Dist: numpy>=2.0
|
|
23
|
+
Requires-Dist: transformers>=5.14.1
|
|
24
|
+
Provides-Extra: torch
|
|
25
|
+
Requires-Dist: torch>=2.13.0; extra == "torch"
|
|
26
|
+
Provides-Extra: detr
|
|
27
|
+
Requires-Dist: torch>=2.13.0; extra == "detr"
|
|
28
|
+
Requires-Dist: timm>=1.0.28; extra == "detr"
|
|
29
|
+
Provides-Extra: all
|
|
30
|
+
Requires-Dist: torch>=2.13.0; extra == "all"
|
|
31
|
+
Requires-Dist: timm>=1.0.28; extra == "all"
|
|
32
|
+
Dynamic: license-file
|
|
33
|
+
|
|
34
|
+
# SinglePipe
|
|
35
|
+
|
|
36
|
+
> Typed interface for ML model inference: one API for HuggingFace, PyTorch, and (eventually) other backends.
|
|
37
|
+
|
|
38
|
+
⚠️ **Status: pre-alpha.** This PyPI release reserves the package name. The library is under active development, the API is unstable, and only one task — object detection via HuggingFace — is currently supported. Not recommended for production use.
|
|
39
|
+
|
|
40
|
+
## Why
|
|
41
|
+
|
|
42
|
+
Working with ML models means drowning in mismatched formats: HuggingFace returns `list[dict]`, a custom `nn.Module` returns tensors, a remote API returns JSON with its own schema. SinglePipe gives you **one entry point** and a **typed output** regardless of what's under the hood.
|
|
43
|
+
|
|
44
|
+
## Features
|
|
45
|
+
|
|
46
|
+
- **Typed input/output schemas** built on pydantic — validation out of the box, IDE autocomplete, JSON serialization ready.
|
|
47
|
+
- **Flexible input** — an image can arrive as `PIL.Image`, `numpy.ndarray`, `torch.Tensor`, a file path, a URL, or raw bytes; SinglePipe normalizes it to a canonical format.
|
|
48
|
+
- **Unified interface across backends** — HuggingFace today, PyTorch, ONNX, and remote APIs next.
|
|
49
|
+
- **Pipeline composition** *(in progress)* — combine multiple models into one typed flow (detector → OCR → classifier) with automatic type-based wiring.
|
|
50
|
+
- **Async and parallel execution of independent nodes** *(on the roadmap)*.
|
|
51
|
+
|
|
52
|
+
## Installation
|
|
53
|
+
|
|
54
|
+
Not available yet. This release reserves the package name on PyPI.
|
|
55
|
+
|
|
56
|
+
Requires Python 3.11+.
|
|
57
|
+
|
|
58
|
+
## Quick start
|
|
59
|
+
|
|
60
|
+
```python
|
|
61
|
+
from singlepipe.adapters.hf_adapter import HFAdapter
|
|
62
|
+
from singlepipe.model_types import TaskType
|
|
63
|
+
|
|
64
|
+
adapter = HFAdapter(
|
|
65
|
+
model_name="facebook/detr-resnet-50",
|
|
66
|
+
task=TaskType.OBJECT_DETECTION,
|
|
67
|
+
)
|
|
68
|
+
|
|
69
|
+
result = adapter("http://images.cocodataset.org/val2017/000000039769.jpg")
|
|
70
|
+
|
|
71
|
+
print(f"Found {len(result)} objects")
|
|
72
|
+
for det in result:
|
|
73
|
+
print(f" {det.label}: score={det.score:.2f}, box={det.box.as_tuple()}")
|
|
74
|
+
```
|
|
75
|
+
|
|
76
|
+
Accepted inputs: `PIL.Image`, `numpy.ndarray` (HWC), `torch.Tensor` (CHW), a file path (`str` or `Path`), an `http(s)://` URL, or raw `bytes`.
|
|
77
|
+
|
|
78
|
+
## Working with results
|
|
79
|
+
|
|
80
|
+
The result is a typed pydantic object with a rich API:
|
|
81
|
+
|
|
82
|
+
```python
|
|
83
|
+
# Filtering
|
|
84
|
+
confident = result.filter_by_score(0.7)
|
|
85
|
+
cats_only = result.filter_by_label("cat")
|
|
86
|
+
some = result.filter_by_label(["cat", "dog", "person"])
|
|
87
|
+
|
|
88
|
+
# Iteration
|
|
89
|
+
for det in result:
|
|
90
|
+
print(det.label, det.score)
|
|
91
|
+
print(f"Total: {len(result)}")
|
|
92
|
+
|
|
93
|
+
# Geometry
|
|
94
|
+
box = result.predictions[0].box
|
|
95
|
+
box.width, box.height, box.area, box.center
|
|
96
|
+
box.clip(image_width=800, image_height=600)
|
|
97
|
+
box.iou(other_box)
|
|
98
|
+
|
|
99
|
+
# Serialization
|
|
100
|
+
result.model_dump()
|
|
101
|
+
result.model_dump_json()
|
|
102
|
+
```
|
|
103
|
+
|
|
104
|
+
## Supported tasks
|
|
105
|
+
|
|
106
|
+
| Task | Backend | Status |
|
|
107
|
+
|---|---|---|
|
|
108
|
+
| Object detection | HuggingFace | ✅ |
|
|
109
|
+
| Image classification | HuggingFace | 🚧 planned |
|
|
110
|
+
| Text classification | HuggingFace | 🚧 planned |
|
|
111
|
+
| Text generation | HuggingFace | 🚧 planned |
|
|
112
|
+
|
|
113
|
+
## Roadmap
|
|
114
|
+
|
|
115
|
+
- [x] HF adapter for single-model inference
|
|
116
|
+
- [x] Typed schemas for object detection
|
|
117
|
+
- [ ] Second task (image classification) — verifies the pattern generalizes
|
|
118
|
+
- [ ] `Node` interface for unified pipeline nodes
|
|
119
|
+
- [ ] Linear pipeline composition (`Sequential`)
|
|
120
|
+
- [ ] PyTorch adapter (for custom `nn.Module`)
|
|
121
|
+
- [ ] Full DAG with branching
|
|
122
|
+
- [ ] Async and parallel execution of independent nodes
|
|
123
|
+
- [ ] ONNX Runtime backend
|
|
124
|
+
- [ ] Remote-API backend (OpenAI-compatible)
|
|
125
|
+
|
|
126
|
+
## Development
|
|
127
|
+
|
|
128
|
+
```bash
|
|
129
|
+
uv sync --group dev
|
|
130
|
+
uv run pytest
|
|
131
|
+
```
|
|
132
|
+
|
|
133
|
+
## License
|
|
134
|
+
|
|
135
|
+
MIT — see [LICENSE](LICENSE).
|
|
@@ -0,0 +1,102 @@
|
|
|
1
|
+
# SinglePipe
|
|
2
|
+
|
|
3
|
+
> Typed interface for ML model inference: one API for HuggingFace, PyTorch, and (eventually) other backends.
|
|
4
|
+
|
|
5
|
+
⚠️ **Status: pre-alpha.** This PyPI release reserves the package name. The library is under active development, the API is unstable, and only one task — object detection via HuggingFace — is currently supported. Not recommended for production use.
|
|
6
|
+
|
|
7
|
+
## Why
|
|
8
|
+
|
|
9
|
+
Working with ML models means drowning in mismatched formats: HuggingFace returns `list[dict]`, a custom `nn.Module` returns tensors, a remote API returns JSON with its own schema. SinglePipe gives you **one entry point** and a **typed output** regardless of what's under the hood.
|
|
10
|
+
|
|
11
|
+
## Features
|
|
12
|
+
|
|
13
|
+
- **Typed input/output schemas** built on pydantic — validation out of the box, IDE autocomplete, JSON serialization ready.
|
|
14
|
+
- **Flexible input** — an image can arrive as `PIL.Image`, `numpy.ndarray`, `torch.Tensor`, a file path, a URL, or raw bytes; SinglePipe normalizes it to a canonical format.
|
|
15
|
+
- **Unified interface across backends** — HuggingFace today, PyTorch, ONNX, and remote APIs next.
|
|
16
|
+
- **Pipeline composition** *(in progress)* — combine multiple models into one typed flow (detector → OCR → classifier) with automatic type-based wiring.
|
|
17
|
+
- **Async and parallel execution of independent nodes** *(on the roadmap)*.
|
|
18
|
+
|
|
19
|
+
## Installation
|
|
20
|
+
|
|
21
|
+
Not available yet. This release reserves the package name on PyPI.
|
|
22
|
+
|
|
23
|
+
Requires Python 3.11+.
|
|
24
|
+
|
|
25
|
+
## Quick start
|
|
26
|
+
|
|
27
|
+
```python
|
|
28
|
+
from singlepipe.adapters.hf_adapter import HFAdapter
|
|
29
|
+
from singlepipe.model_types import TaskType
|
|
30
|
+
|
|
31
|
+
adapter = HFAdapter(
|
|
32
|
+
model_name="facebook/detr-resnet-50",
|
|
33
|
+
task=TaskType.OBJECT_DETECTION,
|
|
34
|
+
)
|
|
35
|
+
|
|
36
|
+
result = adapter("http://images.cocodataset.org/val2017/000000039769.jpg")
|
|
37
|
+
|
|
38
|
+
print(f"Found {len(result)} objects")
|
|
39
|
+
for det in result:
|
|
40
|
+
print(f" {det.label}: score={det.score:.2f}, box={det.box.as_tuple()}")
|
|
41
|
+
```
|
|
42
|
+
|
|
43
|
+
Accepted inputs: `PIL.Image`, `numpy.ndarray` (HWC), `torch.Tensor` (CHW), a file path (`str` or `Path`), an `http(s)://` URL, or raw `bytes`.
|
|
44
|
+
|
|
45
|
+
## Working with results
|
|
46
|
+
|
|
47
|
+
The result is a typed pydantic object with a rich API:
|
|
48
|
+
|
|
49
|
+
```python
|
|
50
|
+
# Filtering
|
|
51
|
+
confident = result.filter_by_score(0.7)
|
|
52
|
+
cats_only = result.filter_by_label("cat")
|
|
53
|
+
some = result.filter_by_label(["cat", "dog", "person"])
|
|
54
|
+
|
|
55
|
+
# Iteration
|
|
56
|
+
for det in result:
|
|
57
|
+
print(det.label, det.score)
|
|
58
|
+
print(f"Total: {len(result)}")
|
|
59
|
+
|
|
60
|
+
# Geometry
|
|
61
|
+
box = result.predictions[0].box
|
|
62
|
+
box.width, box.height, box.area, box.center
|
|
63
|
+
box.clip(image_width=800, image_height=600)
|
|
64
|
+
box.iou(other_box)
|
|
65
|
+
|
|
66
|
+
# Serialization
|
|
67
|
+
result.model_dump()
|
|
68
|
+
result.model_dump_json()
|
|
69
|
+
```
|
|
70
|
+
|
|
71
|
+
## Supported tasks
|
|
72
|
+
|
|
73
|
+
| Task | Backend | Status |
|
|
74
|
+
|---|---|---|
|
|
75
|
+
| Object detection | HuggingFace | ✅ |
|
|
76
|
+
| Image classification | HuggingFace | 🚧 planned |
|
|
77
|
+
| Text classification | HuggingFace | 🚧 planned |
|
|
78
|
+
| Text generation | HuggingFace | 🚧 planned |
|
|
79
|
+
|
|
80
|
+
## Roadmap
|
|
81
|
+
|
|
82
|
+
- [x] HF adapter for single-model inference
|
|
83
|
+
- [x] Typed schemas for object detection
|
|
84
|
+
- [ ] Second task (image classification) — verifies the pattern generalizes
|
|
85
|
+
- [ ] `Node` interface for unified pipeline nodes
|
|
86
|
+
- [ ] Linear pipeline composition (`Sequential`)
|
|
87
|
+
- [ ] PyTorch adapter (for custom `nn.Module`)
|
|
88
|
+
- [ ] Full DAG with branching
|
|
89
|
+
- [ ] Async and parallel execution of independent nodes
|
|
90
|
+
- [ ] ONNX Runtime backend
|
|
91
|
+
- [ ] Remote-API backend (OpenAI-compatible)
|
|
92
|
+
|
|
93
|
+
## Development
|
|
94
|
+
|
|
95
|
+
```bash
|
|
96
|
+
uv sync --group dev
|
|
97
|
+
uv run pytest
|
|
98
|
+
```
|
|
99
|
+
|
|
100
|
+
## License
|
|
101
|
+
|
|
102
|
+
MIT — see [LICENSE](LICENSE).
|
|
@@ -0,0 +1,70 @@
|
|
|
1
|
+
[project]
|
|
2
|
+
name = "singlepipe"
|
|
3
|
+
version = "0.0.1"
|
|
4
|
+
description = "Typed interface for ML model inference across HuggingFace, PyTorch, and other backends"
|
|
5
|
+
readme = "README.md"
|
|
6
|
+
requires-python = ">=3.11"
|
|
7
|
+
license = "MIT"
|
|
8
|
+
license-files = ["LICENSE"]
|
|
9
|
+
authors = [
|
|
10
|
+
{ name = "Gleb", email = "glebov274@gmail.com" },
|
|
11
|
+
]
|
|
12
|
+
keywords = ["ml", "inference", "huggingface", "pytorch", "pydantic", "typed"]
|
|
13
|
+
classifiers = [
|
|
14
|
+
"Development Status :: 2 - Pre-Alpha",
|
|
15
|
+
"Intended Audience :: Developers",
|
|
16
|
+
"Programming Language :: Python :: 3",
|
|
17
|
+
"Programming Language :: Python :: 3.11",
|
|
18
|
+
"Programming Language :: Python :: 3.12",
|
|
19
|
+
"Programming Language :: Python :: 3.13",
|
|
20
|
+
"Programming Language :: Python :: 3.14",
|
|
21
|
+
"Topic :: Scientific/Engineering :: Artificial Intelligence",
|
|
22
|
+
]
|
|
23
|
+
dependencies = [
|
|
24
|
+
"httpx>=0.28.1",
|
|
25
|
+
"pillow>=12.3.0",
|
|
26
|
+
"pydantic>=2.13.4",
|
|
27
|
+
"numpy>=2.0",
|
|
28
|
+
"transformers>=5.14.1",
|
|
29
|
+
]
|
|
30
|
+
|
|
31
|
+
[project.optional-dependencies]
|
|
32
|
+
torch = [
|
|
33
|
+
"torch>=2.13.0",
|
|
34
|
+
]
|
|
35
|
+
detr = [
|
|
36
|
+
"torch>=2.13.0",
|
|
37
|
+
"timm>=1.0.28",
|
|
38
|
+
]
|
|
39
|
+
all = [
|
|
40
|
+
"torch>=2.13.0",
|
|
41
|
+
"timm>=1.0.28",
|
|
42
|
+
]
|
|
43
|
+
|
|
44
|
+
[dependency-groups]
|
|
45
|
+
dev = [
|
|
46
|
+
"pytest>=8",
|
|
47
|
+
"ruff>=0.16.1",
|
|
48
|
+
"mypy>=2.3.0",
|
|
49
|
+
]
|
|
50
|
+
|
|
51
|
+
[build-system]
|
|
52
|
+
requires = ["setuptools>=77"]
|
|
53
|
+
build-backend = "setuptools.build_meta"
|
|
54
|
+
|
|
55
|
+
[tool.setuptools.packages.find]
|
|
56
|
+
where = ["."]
|
|
57
|
+
include = ["singlepipe*"]
|
|
58
|
+
|
|
59
|
+
[tool.ruff]
|
|
60
|
+
line-length = 100
|
|
61
|
+
target-version = "py311"
|
|
62
|
+
|
|
63
|
+
[tool.ruff.lint]
|
|
64
|
+
select = ["E", "F", "I", "UP", "B", "SIM"]
|
|
65
|
+
|
|
66
|
+
[tool.mypy]
|
|
67
|
+
python_version = "3.11"
|
|
68
|
+
strict = false
|
|
69
|
+
warn_return_any = true
|
|
70
|
+
warn_unused_ignores = true
|
|
File without changes
|
|
@@ -0,0 +1,119 @@
|
|
|
1
|
+
from typing import TYPE_CHECKING, Any
|
|
2
|
+
|
|
3
|
+
from transformers import pipeline as hf_pipeline
|
|
4
|
+
|
|
5
|
+
from singlepipe.model_types import TaskType
|
|
6
|
+
from singlepipe.schemas import get_schemas
|
|
7
|
+
|
|
8
|
+
if TYPE_CHECKING:
|
|
9
|
+
pass
|
|
10
|
+
|
|
11
|
+
|
|
12
|
+
class HFAdapter:
|
|
13
|
+
"""Typed adapter around ``transformers.pipeline`` for a specific task.
|
|
14
|
+
|
|
15
|
+
Wraps a HuggingFace pipeline with SinglePipe's input/output schemas:
|
|
16
|
+
the raw input is validated and normalized into the schema's canonical
|
|
17
|
+
form, the pipeline is invoked, and its raw output is validated and
|
|
18
|
+
materialized as a typed SinglePipe output object.
|
|
19
|
+
|
|
20
|
+
The heavy work (loading the model, tokenizer/processor, and running
|
|
21
|
+
pre- and post-processing) is delegated to ``transformers.pipeline``.
|
|
22
|
+
This class only enforces the typed boundary on both sides.
|
|
23
|
+
|
|
24
|
+
Parameters
|
|
25
|
+
----------
|
|
26
|
+
model_name:
|
|
27
|
+
HuggingFace model id (e.g. ``"facebook/detr-resnet-50"``) or a
|
|
28
|
+
local path to model weights.
|
|
29
|
+
task:
|
|
30
|
+
SinglePipe task identifier. Its ``.value`` must match a valid
|
|
31
|
+
``transformers.pipeline`` task string (e.g. ``"object-detection"``).
|
|
32
|
+
Determines which input/output schemas are pulled from the registry.
|
|
33
|
+
device:
|
|
34
|
+
Device to place the model on. ``None`` = CPU, an ``int`` = CUDA
|
|
35
|
+
device index, or a device string such as ``"cpu"`` or ``"cuda:0"``.
|
|
36
|
+
**pipeline_kwargs:
|
|
37
|
+
Extra keyword arguments passed straight to ``transformers.pipeline``
|
|
38
|
+
(e.g. ``torch_dtype``, ``trust_remote_code``, ``revision``).
|
|
39
|
+
|
|
40
|
+
Attributes
|
|
41
|
+
----------
|
|
42
|
+
model_name : str
|
|
43
|
+
The model id or path this adapter was constructed with.
|
|
44
|
+
task : TaskType
|
|
45
|
+
The task this adapter is bound to.
|
|
46
|
+
pipeline : transformers.Pipeline
|
|
47
|
+
The underlying HuggingFace pipeline instance.
|
|
48
|
+
|
|
49
|
+
Examples
|
|
50
|
+
--------
|
|
51
|
+
>>> adapter = HFAdapter("facebook/detr-resnet-50", TaskType.OBJECT_DETECTION)
|
|
52
|
+
>>> result = adapter("http://images.cocodataset.org/val2017/000000039769.jpg")
|
|
53
|
+
>>> for det in result:
|
|
54
|
+
... print(det.label, det.score, det.box.as_tuple())
|
|
55
|
+
|
|
56
|
+
Notes
|
|
57
|
+
-----
|
|
58
|
+
- The model is loaded once at construction; subsequent calls reuse it.
|
|
59
|
+
- Batching, streaming, and async execution are not supported at this
|
|
60
|
+
layer — one input, one output per call.
|
|
61
|
+
- Task↔model compatibility is not pre-validated: mismatches surface as
|
|
62
|
+
errors from ``transformers.pipeline`` at construction time.
|
|
63
|
+
"""
|
|
64
|
+
|
|
65
|
+
|
|
66
|
+
|
|
67
|
+
|
|
68
|
+
def __init__(
|
|
69
|
+
self,
|
|
70
|
+
model_name: str,
|
|
71
|
+
task: TaskType,
|
|
72
|
+
device: int | str | None = None,
|
|
73
|
+
**pipeline_kwargs: Any,
|
|
74
|
+
) -> None:
|
|
75
|
+
self.model_name = model_name
|
|
76
|
+
self.task = task
|
|
77
|
+
|
|
78
|
+
self.pipeline = hf_pipeline(
|
|
79
|
+
task=task.value, # type: ignore[arg-type]
|
|
80
|
+
model=model_name,
|
|
81
|
+
device=device, # type: ignore[arg-type]
|
|
82
|
+
**pipeline_kwargs,
|
|
83
|
+
)
|
|
84
|
+
|
|
85
|
+
self._input_schema, self._output_schema = get_schemas(task)
|
|
86
|
+
|
|
87
|
+
|
|
88
|
+
|
|
89
|
+
|
|
90
|
+
|
|
91
|
+
def __call__(self, raw_input):
|
|
92
|
+
|
|
93
|
+
if isinstance(raw_input, self._input_schema):
|
|
94
|
+
validated_in = raw_input
|
|
95
|
+
elif isinstance(raw_input, dict):
|
|
96
|
+
validated_in = self._input_schema.model_validate(raw_input)
|
|
97
|
+
else:
|
|
98
|
+
validated_in = self._input_schema.from_raw(raw_input)
|
|
99
|
+
|
|
100
|
+
|
|
101
|
+
raw_output = self.pipeline(validated_in.to_pipeline_input())
|
|
102
|
+
|
|
103
|
+
return self._output_schema.from_pipeline_output(
|
|
104
|
+
raw_output,
|
|
105
|
+
**validated_in.output_context
|
|
106
|
+
)
|
|
107
|
+
|
|
108
|
+
def __repr__(self) -> str:
|
|
109
|
+
return f"HFAdapter(model={self.model_name!r}, task={self.task.value!r})"
|
|
110
|
+
|
|
111
|
+
|
|
112
|
+
if __name__ == "__main__":
|
|
113
|
+
adapter = HFAdapter("facebook/detr-resnet-50", TaskType.OBJECT_DETECTION)
|
|
114
|
+
print(adapter)
|
|
115
|
+
|
|
116
|
+
result = adapter("http://images.cocodataset.org/val2017/000000039769.jpg")
|
|
117
|
+
print(f"Found {len(result)} objects")
|
|
118
|
+
for det in result:
|
|
119
|
+
print(f" {det.label}: {det.score:.3f}")
|
|
@@ -0,0 +1,79 @@
|
|
|
1
|
+
from enum import StrEnum
|
|
2
|
+
|
|
3
|
+
|
|
4
|
+
class TaskType(StrEnum):
|
|
5
|
+
"""Semantic identifier of an ML task in SinglePipe.
|
|
6
|
+
|
|
7
|
+
``TaskType`` is the vocabulary the schema registry keys on:
|
|
8
|
+
each member corresponds to a pair of input/output schemas
|
|
9
|
+
describing the contract of that task.
|
|
10
|
+
|
|
11
|
+
Enum values are intentionally aligned with HuggingFace
|
|
12
|
+
``transformers.pipeline`` task strings whenever a matching HF
|
|
13
|
+
pipeline exists — this lets ``HFAdapter`` pass the value straight
|
|
14
|
+
through without a lookup table. Tasks with no HF equivalent
|
|
15
|
+
(``TABULAR_CLASSIFICATION``, ``TABULAR_REGRESSION``, ``OCR``)
|
|
16
|
+
keep values chosen for readability; they are served by other
|
|
17
|
+
adapters.
|
|
18
|
+
|
|
19
|
+
Because ``TaskType`` inherits from ``StrEnum``, members compare
|
|
20
|
+
equal to their string values and can be used interchangeably
|
|
21
|
+
with strings — in dict keys, HF pipeline arguments, JSON, etc.
|
|
22
|
+
|
|
23
|
+
The enum grows as new tasks gain first-class support. A member
|
|
24
|
+
listed here is *known* to SinglePipe; whether schemas are actually
|
|
25
|
+
registered for it is a separate question answered by the registry
|
|
26
|
+
(``get_schemas`` raises a descriptive error for known-but-
|
|
27
|
+
unregistered tasks).
|
|
28
|
+
|
|
29
|
+
Examples
|
|
30
|
+
--------
|
|
31
|
+
>>> TaskType.OBJECT_DETECTION == "object-detection"
|
|
32
|
+
True
|
|
33
|
+
>>> TaskType.OBJECT_DETECTION.value
|
|
34
|
+
'object-detection'
|
|
35
|
+
"""
|
|
36
|
+
|
|
37
|
+
# Image
|
|
38
|
+
IMAGE_CLASSIFICATION = "image-classification"
|
|
39
|
+
IMAGE_SEGMENTATION = "image-segmentation"
|
|
40
|
+
OBJECT_DETECTION = "object-detection"
|
|
41
|
+
DEPTH_ESTIMATION = "depth-estimation"
|
|
42
|
+
IMAGE_EMBEDDING = "image-feature-extraction"
|
|
43
|
+
ZERO_SHOT_IMAGE_CLASSIFICATION = "zero-shot-image-classification"
|
|
44
|
+
ZERO_SHOT_OBJECT_DETECTION = "zero-shot-object-detection"
|
|
45
|
+
MASK_GENERATION = "mask-generation"
|
|
46
|
+
KEYPOINT_MATCHING = "keypoint-matching"
|
|
47
|
+
|
|
48
|
+
# Text
|
|
49
|
+
TEXT_CLASSIFICATION = "text-classification"
|
|
50
|
+
TOKEN_CLASSIFICATION = "token-classification"
|
|
51
|
+
TEXT_GENERATION = "text-generation"
|
|
52
|
+
TEXT2TEXT_GENERATION = "text2text-generation"
|
|
53
|
+
FILL_MASK = "fill-mask"
|
|
54
|
+
SUMMARIZATION = "summarization"
|
|
55
|
+
TRANSLATION = "translation"
|
|
56
|
+
QUESTION_ANSWERING = "question-answering"
|
|
57
|
+
TEXT_EMBEDDING = "feature-extraction"
|
|
58
|
+
ZERO_SHOT_TEXT_CLASSIFICATION = "zero-shot-classification"
|
|
59
|
+
|
|
60
|
+
# Audio
|
|
61
|
+
AUDIO_CLASSIFICATION = "audio-classification"
|
|
62
|
+
SPEECH_TO_TEXT = "automatic-speech-recognition"
|
|
63
|
+
TEXT_TO_AUDIO = "text-to-audio"
|
|
64
|
+
ZERO_SHOT_AUDIO_CLASSIFICATION = "zero-shot-audio-classification"
|
|
65
|
+
|
|
66
|
+
# Vidio
|
|
67
|
+
VIDEO_CLASSIFICATION = "video-classification"
|
|
68
|
+
|
|
69
|
+
# Multimodel
|
|
70
|
+
IMAGE_TEXT_TO_TEXT = "image-text-to-text"
|
|
71
|
+
DOCUMENT_QUESTION_ANSWERING = "document-question-answering"
|
|
72
|
+
TABLE_QUESTION_ANSWERING = "table-question-answering"
|
|
73
|
+
|
|
74
|
+
# Table
|
|
75
|
+
TABULAR_CLASSIFICATION = "tabular-classification"
|
|
76
|
+
TABULAR_REGRESSION = "tabular-regression"
|
|
77
|
+
|
|
78
|
+
# OCR
|
|
79
|
+
OCR = "ocr"
|
|
@@ -0,0 +1,25 @@
|
|
|
1
|
+
from singlepipe.model_types import TaskType
|
|
2
|
+
|
|
3
|
+
from .image_schemas.object_detection import (
|
|
4
|
+
BBox,
|
|
5
|
+
Detection,
|
|
6
|
+
ObjectDetectionInput,
|
|
7
|
+
ObjectDetectionOutput,
|
|
8
|
+
)
|
|
9
|
+
from .registry import InputOutputSchema, get_schemas, register
|
|
10
|
+
|
|
11
|
+
register(
|
|
12
|
+
TaskType.OBJECT_DETECTION,
|
|
13
|
+
ObjectDetectionInput,
|
|
14
|
+
ObjectDetectionOutput,
|
|
15
|
+
)
|
|
16
|
+
|
|
17
|
+
__all__ = [
|
|
18
|
+
"BBox",
|
|
19
|
+
"Detection",
|
|
20
|
+
"InputOutputSchema",
|
|
21
|
+
"ObjectDetectionInput",
|
|
22
|
+
"ObjectDetectionOutput",
|
|
23
|
+
"get_schemas",
|
|
24
|
+
"register",
|
|
25
|
+
]
|
|
File without changes
|
|
@@ -0,0 +1,40 @@
|
|
|
1
|
+
"""Placeholder schemas for TaskType.AUDIO_CLASSIFICATION.
|
|
2
|
+
|
|
3
|
+
Not yet implemented. Exists to unlock ``HFAdapter`` type overloads.
|
|
4
|
+
Replace with real schemas when this task gains first-class support.
|
|
5
|
+
"""
|
|
6
|
+
|
|
7
|
+
from typing import Any, Self
|
|
8
|
+
|
|
9
|
+
from pydantic import BaseModel, ConfigDict
|
|
10
|
+
|
|
11
|
+
|
|
12
|
+
class AudioClassificationInput(BaseModel):
|
|
13
|
+
"""Placeholder input schema. Fields TBD."""
|
|
14
|
+
|
|
15
|
+
model_config = ConfigDict(arbitrary_types_allowed=True)
|
|
16
|
+
|
|
17
|
+
@classmethod
|
|
18
|
+
def from_raw(cls, raw: Any) -> Self:
|
|
19
|
+
raise NotImplementedError(
|
|
20
|
+
"AudioClassificationInput is not implemented yet."
|
|
21
|
+
)
|
|
22
|
+
|
|
23
|
+
def to_pipeline_input(self) -> Any:
|
|
24
|
+
raise NotImplementedError(
|
|
25
|
+
"AudioClassificationInput is not implemented yet."
|
|
26
|
+
)
|
|
27
|
+
|
|
28
|
+
@property
|
|
29
|
+
def output_context(self) -> dict[str, Any]:
|
|
30
|
+
return {}
|
|
31
|
+
|
|
32
|
+
|
|
33
|
+
class AudioClassificationOutput(BaseModel):
|
|
34
|
+
"""Placeholder output schema. Fields TBD."""
|
|
35
|
+
|
|
36
|
+
@classmethod
|
|
37
|
+
def from_pipeline_output(cls, raw_output: Any, **context: Any) -> Self:
|
|
38
|
+
raise NotImplementedError(
|
|
39
|
+
"AudioClassificationOutput is not implemented yet."
|
|
40
|
+
)
|
|
@@ -0,0 +1,40 @@
|
|
|
1
|
+
"""Placeholder schemas for TaskType.SPEECH_TO_TEXT.
|
|
2
|
+
|
|
3
|
+
Not yet implemented. Exists to unlock ``HFAdapter`` type overloads.
|
|
4
|
+
Replace with real schemas when this task gains first-class support.
|
|
5
|
+
"""
|
|
6
|
+
|
|
7
|
+
from typing import Any, Self
|
|
8
|
+
|
|
9
|
+
from pydantic import BaseModel, ConfigDict
|
|
10
|
+
|
|
11
|
+
|
|
12
|
+
class SpeechToTextInput(BaseModel):
|
|
13
|
+
"""Placeholder input schema. Fields TBD."""
|
|
14
|
+
|
|
15
|
+
model_config = ConfigDict(arbitrary_types_allowed=True)
|
|
16
|
+
|
|
17
|
+
@classmethod
|
|
18
|
+
def from_raw(cls, raw: Any) -> Self:
|
|
19
|
+
raise NotImplementedError(
|
|
20
|
+
"SpeechToTextInput is not implemented yet."
|
|
21
|
+
)
|
|
22
|
+
|
|
23
|
+
def to_pipeline_input(self) -> Any:
|
|
24
|
+
raise NotImplementedError(
|
|
25
|
+
"SpeechToTextInput is not implemented yet."
|
|
26
|
+
)
|
|
27
|
+
|
|
28
|
+
@property
|
|
29
|
+
def output_context(self) -> dict[str, Any]:
|
|
30
|
+
return {}
|
|
31
|
+
|
|
32
|
+
|
|
33
|
+
class SpeechToTextOutput(BaseModel):
|
|
34
|
+
"""Placeholder output schema. Fields TBD."""
|
|
35
|
+
|
|
36
|
+
@classmethod
|
|
37
|
+
def from_pipeline_output(cls, raw_output: Any, **context: Any) -> Self:
|
|
38
|
+
raise NotImplementedError(
|
|
39
|
+
"SpeechToTextOutput is not implemented yet."
|
|
40
|
+
)
|