v0.17.0

See https://github.com/quic/ai-hub-models/releases/v0.17.0 for changelog.

Signed-off-by: QAIHM Team <[email protected]>

.pre-commit-config.yaml

@@ -58,6 +58,11 @@ repos:
     hooks:
     -   id: shellcheck
         exclude: '\.yml$'
+-   repo: https://github.com/asottile/pyupgrade
+    rev: v3.16.0
+    hooks:
+    -   id: pyupgrade
+        args: ['--py39-plus', '--keep-runtime-typing']
 -   repo: https://github.com/pycqa/isort
     rev: 5.12.0
     hooks:

README.md

@@ -51,11 +51,13 @@ and many more.
 
 ## Installation
 
-We currently support **Python >=3.8 and <= 3.10.** We recommend using a Python
+We currently support **Python 3.9, 3.10 (recommended), 3.11, and 3.12.** We recommend using a Python
 virtual environment
 ([miniconda](https://docs.anaconda.com/free/miniconda/miniconda-install/) or
 [virtualenv](https://virtualenv.pypa.io/en/latest/)).
 
+*NOTE: Many quantized models are supported only with python 3.10*.
+
 You can setup a virtualenv using:
 ```
 python -m venv qai_hub_models_env && source qai_hub_models_env/bin/activate
@@ -181,7 +183,6 @@ Here is a simplified example of code that can be used to run the entire model
 on a cloud hosted device:
 
 ```python
-from typing import Tuple
 import torch
 import qai_hub as hub
 from qai_hub_models.models.yolov7 import Model as YOLOv7Model
@@ -359,10 +360,14 @@ Qualcomm® AI Hub Models is licensed under BSD-3. See the [LICENSE file](../LICE
 | [DETR-ResNet101-DC5](https://aihub.qualcomm.com/models/detr_resnet101_dc5) | [qai_hub_models.models.detr_resnet101_dc5](qai_hub_models/models/detr_resnet101_dc5/README.md) | ✔️ | ✔️ | ✔️
 | [DETR-ResNet50](https://aihub.qualcomm.com/models/detr_resnet50) | [qai_hub_models.models.detr_resnet50](qai_hub_models/models/detr_resnet50/README.md) | ✔️ | ✔️ | ✔️
 | [DETR-ResNet50-DC5](https://aihub.qualcomm.com/models/detr_resnet50_dc5) | [qai_hub_models.models.detr_resnet50_dc5](qai_hub_models/models/detr_resnet50_dc5/README.md) | ✔️ | ✔️ | ✔️
+| [FaceAttribNet](https://aihub.qualcomm.com/models/face_attrib_net) | [qai_hub_models.models.face_attrib_net](qai_hub_models/models/face_attrib_net/README.md) | ✔️ | ✔️ | ✔️
+| [FootTrackNet_Quantized](https://aihub.qualcomm.com/models/foot_track_net_quantized) | [qai_hub_models.models.foot_track_net_quantized](qai_hub_models/models/foot_track_net_quantized/README.md) | ✔️ | ✔️ | ✔️
+| [Lightweight-Face-Detection](https://aihub.qualcomm.com/models/face_det_lite) | [qai_hub_models.models.face_det_lite](qai_hub_models/models/face_det_lite/README.md) | ✔️ | ✔️ | ✔️
 | [MediaPipe-Face-Detection](https://aihub.qualcomm.com/models/mediapipe_face) | [qai_hub_models.models.mediapipe_face](qai_hub_models/models/mediapipe_face/README.md) | ✔️ | ✔️ | ✔️
 | [MediaPipe-Face-Detection-Quantized](https://aihub.qualcomm.com/models/mediapipe_face_quantized) | [qai_hub_models.models.mediapipe_face_quantized](qai_hub_models/models/mediapipe_face_quantized/README.md) | ✔️ | ✔️ | ✔️
 | [MediaPipe-Hand-Detection](https://aihub.qualcomm.com/models/mediapipe_hand) | [qai_hub_models.models.mediapipe_hand](qai_hub_models/models/mediapipe_hand/README.md) | ✔️ | ✔️ | ✔️
 | [PPE-Detection](https://aihub.qualcomm.com/models/gear_guard_net) | [qai_hub_models.models.gear_guard_net](qai_hub_models/models/gear_guard_net/README.md) | ✔️ | ✔️ | ✔️
+| [PPE-Detection-Quantized](https://aihub.qualcomm.com/models/gear_guard_net_quantized) | [qai_hub_models.models.gear_guard_net_quantized](qai_hub_models/models/gear_guard_net_quantized/README.md) | ✔️ | ✔️ | ✔️
 | [Person-Foot-Detection](https://aihub.qualcomm.com/models/foot_track_net) | [qai_hub_models.models.foot_track_net](qai_hub_models/models/foot_track_net/README.md) | ✔️ | ✔️ | ✔️
 | [YOLOv11-Detection](https://aihub.qualcomm.com/models/yolov11_det) | [qai_hub_models.models.yolov11_det](qai_hub_models/models/yolov11_det/README.md) | ✔️ | ✔️ | ✔️
 | [YOLOv8-Detection](https://aihub.qualcomm.com/models/yolov8_det) | [qai_hub_models.models.yolov8_det](qai_hub_models/models/yolov8_det/README.md) | ✔️ | ✔️ | ✔️

qai_hub_models/_version.py

@@ -2,4 +2,4 @@
 # Copyright (c) 2024 Qualcomm Innovation Center, Inc. All rights reserved.
 # SPDX-License-Identifier: BSD-3-Clause
 # ---------------------------------------------------------------------
-__version__ = "0.16.2"
+__version__ = "0.17.0"

qai_hub_models/datasets/__init__.py

@@ -3,7 +3,6 @@
 # SPDX-License-Identifier: BSD-3-Clause
 # ---------------------------------------------------------------------
 
-from typing import Dict, List, Type
 
 from .bsd300 import BSD300Dataset
 from .coco import CocoDataset
@@ -12,15 +11,15 @@
 from .imagenette import ImagenetteDataset
 from .pascal_voc import VOCSegmentationDataset
 
-ALL_DATASETS: List[Type[BaseDataset]] = [
+ALL_DATASETS: list[type[BaseDataset]] = [
     CocoDataset,
     VOCSegmentationDataset,
     BSD300Dataset,
     ImagenetDataset,
     ImagenetteDataset,
 ]
 
-DATASET_NAME_MAP: Dict[str, Type[BaseDataset]] = {
+DATASET_NAME_MAP: dict[str, type[BaseDataset]] = {
     dataset_cls.dataset_name(): dataset_cls for dataset_cls in ALL_DATASETS
 }
 

qai_hub_models/datasets/bsd300.py

@@ -5,7 +5,6 @@
 from __future__ import annotations
 
 import os
-from typing import Tuple
 
 import numpy as np
 import torch
@@ -66,7 +65,7 @@ def _prepare_data(self):
     def __len__(self):
         return DATASET_LENGTH
 
-    def __getitem__(self, item) -> Tuple[torch.Tensor, torch.Tensor]:
+    def __getitem__(self, item) -> tuple[torch.Tensor, torch.Tensor]:
         # We use the super resolution GT-and-test image preparation from AIMET zoo:
         # https://github.com/quic/aimet-model-zoo/blob/d09d2b0404d10f71a7640a87e9d5e5257b028802/aimet_zoo_torch/quicksrnet/dataloader/utils.py#L51
 

qai_hub_models/datasets/coco.py

@@ -3,7 +3,7 @@
 # SPDX-License-Identifier: BSD-3-Clause
 # ---------------------------------------------------------------------
 import os
-from typing import Tuple, Union
+from typing import Union
 
 import torch
 from torch.utils.data.dataloader import default_collate
@@ -56,7 +56,7 @@ class CocoDataset(BaseDataset, CocoDetection):
     Contains ~5k images spanning 80 classes.
     """
 
-    def __init__(self, target_image_size: Union[int, Tuple[int, int]] = 640):
+    def __init__(self, target_image_size: Union[int, tuple[int, int]] = 640):
         BaseDataset.__init__(self, str(COCO_DATASET.path(extracted=True)))
         CocoDetection.__init__(
             self,
@@ -78,7 +78,7 @@ def __init__(self, target_image_size: Union[int, Tuple[int, int]] = 640):
         )
 
     def __getitem__(self, item):
-        image, target = super(CocoDataset, self).__getitem__(item)
+        image, target = super().__getitem__(item)
         width, height = image.size
         boxes = []
         labels = []

qai_hub_models/datasets/pascal_voc.py

@@ -3,7 +3,6 @@
 # SPDX-License-Identifier: BSD-3-Clause
 # ---------------------------------------------------------------------
 
-from typing import Tuple
 
 import numpy as np
 import torch
@@ -30,7 +29,7 @@ class VOCSegmentationDataset(BaseDataset):
         https://host.robots.ox.ac.uk/pascal/VOC/voc2012/
     """
 
-    def __init__(self, split: str = "train", image_size: Tuple[int, int] = (224, 224)):
+    def __init__(self, split: str = "train", image_size: tuple[int, int] = (224, 224)):
         BaseDataset.__init__(self, str(VOC_ASSET.path().parent / DEVKIT_FOLDER_NAME))
         assert split in ["train", "val", "trainval"]
         self.split = split
@@ -44,7 +43,7 @@ def __init__(self, split: str = "train", image_size: Tuple[int, int] = (224, 224
         self.images = []
         self.categories = []
 
-        with open(splits_dir / (split + ".txt"), "r") as f:
+        with open(splits_dir / (split + ".txt")) as f:
             lines = f.read().splitlines()
 
         for line in lines:

qai_hub_models/evaluators/base_evaluators.py

@@ -5,7 +5,8 @@
 from __future__ import annotations
 
 from abc import ABC, abstractmethod
-from typing import Callable, Collection, Tuple, Union
+from collections.abc import Callable, Collection
+from typing import Union
 
 import torch
 from torch.utils.data.dataloader import DataLoader
@@ -15,7 +16,7 @@
 _ModelIO: TypeAlias = Union[Collection[torch.Tensor], torch.Tensor]
 # Typically is a torch DataLoader, but anything with the collection signature is acceptable.
 _DataLoader: TypeAlias = Union[
-    DataLoader, Collection[Union[_ModelIO, Tuple[_ModelIO, _ModelIO]]]
+    DataLoader, Collection[Union[_ModelIO, tuple[_ModelIO, _ModelIO]]]
 ]
 
 

qai_hub_models/evaluators/detection_evaluator.py

@@ -4,7 +4,7 @@
 # ---------------------------------------------------------------------
 from __future__ import annotations
 
-from typing import Collection
+from collections.abc import Collection
 
 import torch
 from podm.metrics import (  # type: ignore
@@ -33,7 +33,11 @@ def __init__(
         self.scale_x = 1 / image_height
         self.scale_y = 1 / image_width
 
-    def add_batch(self, output: Collection[torch.Tensor], gt: Collection[torch.Tensor]):
+    def add_batch(
+        self,
+        output: Collection[torch.Tensor],
+        gt: Collection[torch.Tensor],
+    ):
         # This evaluator supports 1 output tensor at a time.
         image_id, _, _, bboxes, classes = gt
         pred_boxes, pred_scores, pred_class_idx = output

qai_hub_models/global_requirements.txt

@@ -6,61 +6,86 @@
 # THIS FILE WAS AUTO-GENERATED. DO NOT EDIT MANUALLY.
 
 Deprecated==1.2.11
-aimet-torch==1.32.1.post1; sys_platform == "linux"
+Pillow>10,<12
+aimet-torch==1.32.1.post1; sys_platform == "linux" and python_version == "3.10"
 albumentations==0.5.2
 audio2numpy==0.1.2
 basicsr==1.4.2
-boto3==1.34.119
-botocore==1.34.119
-coverage==5.3.1
+boto3>=1.34,<1.36
+botocore>=1.34,<1.36
 data-gradients==0.3.1
 datasets==2.14.5
 diffusers[torch]==0.21.4
-easydict==1.10
+easydict==1.13
 einops==0.3.2
 ftfy==6.1.1
+gdown==4.7.1
+gitpython==3.1.42
+huggingface-hub>=0.23.1,<0.24
+huggingface_hub>=0.23.1,<0.24
 hydra-core==1.3.0
 imageio[ffmpeg]==2.31.5
 imagesize==1.4.1
-jinja2==3.0.3
-keyrings.envvars==1.1.0; python_version >= '3.9' # used only by CI
+ipython==8.12.3
+jinja2<3.2
+jsonschema>4,<5
+keyrings.envvars==1.1.0
 kornia==0.5.0
 librosa==0.10.1
+matplotlib==3.7.5
 mmcv==2.1.0
 mmdet==3.2.0
 mmpose==1.2.0
-mypy==0.991
+mypy==1.13.0
+numpy>=1.23.5,< 2 # 1.23.5 required by AIMET
 object-detection-metrics==0.4.post1
+onnx>=1.14.1,<1.17 # ONNX must be at least 1.14.1. AIMET-torch and AIMET-ONNX use different ONNX versions.
+onnxsim<=0.4.36
 openai-whisper==20231117
-pre-commit==3.5.0
+opencv-python>4,<5
+packaging>23,<24
+pandas>=1.4.3,<2.3 # 1.4 required by AIMET
+pre-commit==4.0.1
+prettytable==3.11.0
+psutil>6,<7
 pycocotools==2.0.7
-pytest-cov==4.1.0
-pytest-xdist==3.3.1
-pytorch-lightning==1.6.0
+pytest>7,<9
+pytest-cov>=5,<5.2
+pytest-xdist>3,<4
+pytorch-lightning>2,<3
+pyyaml==6.0.2
+qai_hub>=0.18.1
 rapidfuzz==3.8.1
 regex==2023.10.3
+requests_toolbelt==1.0.0
 ruamel-yaml==0.18.6
 samplerate==0.2.1
 schema==0.7.5
-scikit-image==0.21.0
-scikit-learn==1.1.3
-scipy==1.8.1
+scikit-image>0.21.0,<0.25
+scikit-learn>1.1,<1.6
+scipy>=1.8.1,<2 # 1.8.1 is for AIMET
 seaborn==0.11.0
 sentencepiece==0.2.0
 shapely==2.0.3
 soundfile==0.12.1
 stringcase==1.2.0
+tabulate==0.9.0
 tensorboard==2.13.0
+termcolor<=2.5.0
 tflite==2.10.0
 thop==0.1.1.post2209072238
 timm==1.0.3
+torch>=2.1.2,<2.5.0 # 2.1.2 is for AIMET. 2.5 won't work with torchvision yet.
 torchmetrics==1.4.0.post0
+torchvision>=0.16.2,<0.21
+tqdm>=4.66
 transformers==4.41.1
 treelib==1.6.1
 types-PyYAML==6.0.12.12
 types-pillow==10.2.0.20240213
 types-requests==2.31.0.6
 types-tabulate==0.9.0.20240106
+typing-extensions>=4.12.2
 ultralytics==8.0.193
 webdataset==0.2.86
 wheel==0.44.0

qai_hub_models/models/_shared/body_detection/app.py

@@ -2,7 +2,8 @@
 # Copyright (c) 2024 Qualcomm Innovation Center, Inc. All rights reserved.
 # SPDX-License-Identifier: BSD-3-Clause
 # ---------------------------------------------------------------------
-from typing import Callable, List
+
+from collections.abc import Callable
 
 import numpy as np
 import torch
@@ -12,36 +13,12 @@
 from qai_hub_models.utils.image_processing import resize_pad
 
 
-def preprocess(img: np.ndarray, height: int, width: int):
-    """
-    Preprocess model input.
-
-    Inputs:
-        img: np.ndarray
-            Input image of shape [H, W, C]
-        height: int
-            Model input height.
-        width: int
-            Model input width
-    Outputs:
-        input: torch.Tensor
-            Preprocessed model input. Shape is (1, C, H, W)
-        scale: float
-            Scaling factor of input image and network input image.
-        pad: List[float]
-            Top and left padding size.
-    """
-    img = torch.from_numpy(img).permute(2, 0, 1).unsqueeze_(0) / 255.0
-    input, scale, pad = resize_pad(img, (height, width))
-    return input, scale, pad
-
-
-def decode(output: List[torch.Tensor], thr: float) -> np.ndarray:
+def decode(output: list[torch.Tensor], thr: float) -> np.ndarray:
     """
     Decode model output to bounding boxes, class indices and scores.
 
     Inputs:
-        output: List[torch.Tensor]
+        output: list[torch.Tensor]
             Model output.
         thr: float
             Detection threshold. Predictions lower than the thresholds will be discarded.
@@ -87,20 +64,20 @@ def decode(output: List[torch.Tensor], thr: float) -> np.ndarray:
 
 
 def postprocess(
-    output: List[torch.Tensor],
+    output: list[torch.Tensor],
     scale: float,
-    pad: List[int],
+    pad: list[int],
     conf_thr: float,
     iou_thr: float,
 ) -> np.ndarray:
     """
     Post process model output.
     Inputs:
-        output: List[torch.Tensor]
+        output: list[torch.Tensor]
             Multi-scale model output.
         scale: float
             Scaling factor from input image and model input.
-        pad: List[int]
+        pad: list[int]
             Padding sizes from input image and model input.
         conf_thr: float
             Confidence threshold of detections.
@@ -163,7 +140,8 @@ def detect(self, imgfile: str, height: int, width: int, conf: float) -> np.ndarr
             (cls_id, x1, y1, x2, y2, score)
         """
         img = np.array(load_image(imgfile))
-        input, scale, pad = preprocess(img, height, width)
+        img = torch.from_numpy(img).permute(2, 0, 1).unsqueeze_(0)
+        input, scale, pad = resize_pad(img, (height, width))
         output = self.model(input)
         for t, o in enumerate(output):
             output[t] = o.permute(0, 2, 3, 1).detach()