Vayuvahana Technologies Private Limited VajraV1 is a state-of-the-art (SOTA) real time object detection model inspired by the YOLO model architectures. VajraV1 is a family of fast, lightweight models that can be used for a variety of tasks like object detection and tracking, instance segmentation, oriented object detection, pose detection, and image classification.
To request for an Enterprise License please get in touch via Email
| Model | Size (pixels) | mAPval 50-95 |
Speed RTX 4090 TensorRT10 Latency (ms) |
Params (M) | FLOPs (B) |
|---|---|---|---|---|---|
| VajraV1-nano-det | 640 | 44.3 | 1.0 | 3.78 | 13.7 |
| VajraV1-small-det | 640 | 50.4 | 1.1 | 11.58 | 47.9 |
| VajraV1-medium-det | 640 | 52.7 | 1.5 | 20.29 | 94.5 |
| VajraV1-large-det | 640 | 53.7 | 1.8 | 24.63 | 115.2 |
| VajraV1-xlarge-det | 640 | 56.2 | 3.2 | 72.7 | 208.3 |
| Model | Size (pixels) | Box mAPval 50-95 |
Mask mAPval 50-95 |
Speed RTX 4090 TensorRT10 Latency (ms) |
Params (M) | FLOPs (B) |
|---|---|---|---|---|---|---|
| VajraV1-nano-seg | 640 | 43.6 | 35.8 | 1.2 | 4.03 | 17.6 |
| VajraV1-small-seg | 640 | 50.2 | 40.5 | 1.2 | 12.23 | 61.9 |
| VajraV1-medium-seg | 640 | 52.6 | 42.3 | 1.7 | 22.6 | 149.9 |
| VajraV1-large-seg | 640 | 53.6 | 43.1 | 2.0 | 26.93 | 170.6 |
| VajraV1-xlarge-seg | 640 | 55.7 | 44.5 | 3.4 | 75 | 278.1 |
| Model | Size (pixels) | Pose mAPval 50-95 |
Pose mAPval 50 |
Speed RTX 4090 TensorRT10 Latency (ms) |
Params (M) | FLOPs (B) |
|---|---|---|---|---|---|---|
| VajraV1-nano-pose | 640 | 56.4 | 84.7 | 1.2 | 4.07 | 14.8 |
| VajraV1-small-pose | 640 | 65 | 88.9 | 1.4 | 12.07 | 49.6 |
| VajraV1-medium-pose | 640 | 68.5 | 89.9 | 1.8 | 21.15 | 98.2 |
| VajraV1-large-pose | 640 | 69.5 | 90.6 | 2.1 | 25.49 | 118.9 |
| VajraV1-xlarge-pose | 640 | 71.5 | 91.4 | 3.7 | 73.56 | 26.5 |
Install
Git clone the VayuAI SDK including all requirements in a Python>=3.8 environment.
git clone https://github.com/NamanMakkar/VayuAI.git
cd VayuAI
pip install .Usage
Vajra can be used in the Command Line Interface with a vajra or vayuvahana or vayuai
command:
vajra predict model=vajra-v1-nano-det img_size=640 source="path/to/source.jpg"Vajra can also be used directly in a Python environment, and accepts the same arguments as in the CLI example above:
from vajra import Vajra
model = Vajra("vajra-v1-nano-det")
train_results = model.train(
data="coco8.yaml",
epochs=100,
img_size=640,
device="cpu",
weight_decay=0.,
)
metrics = model.val()
results = model("path/to/img.jpg")
results[0].show()
path = model.export(format="onnx")Pretrained COCO weights can also be used for model inference
from vajra import Vajra
model = Vajra("vajra-v1-xlarge-det.pt")
results = model("path/to/img.jpg")
results[0].show()
path = model.export(format="engine", device=0, half=True)✅ detect
✅ pose
✅ obb
✅ segment
✅ small_obj_detect
✅ classify
✅ multilabel_classify
✅ Training
✅ Validation
✅ Prediction
✅ Multi-object Tracking
✅ Model Export -> (TensorRT, ONNX, OpenVINO, Torchscript, CoreML, TensorFlow.js, NCNN, Keras, PaddlePaddle)
✅ Benchmarking
✅ VajraV1-det
✅ VajraV1-cls
✅ VajraV1-pose
✅ VajraV1-seg
✅ VajraV1-obb
✅ SAM
✅ SAM2
✅ FastSAM
✅ MobileSAM
✅ EfficientNetV1
✅ EfficientNetV2
✅ VajraEffNetV1
✅ VajraEffNetV2
✅ ConvNeXtV1
✅ ConvNeXtV2
✅ ResNet
✅ ResNeSt
❌ ResNeXt (Coming Soon!)
❌ ResNetV2 (Coming Soon!)
✅ EdgeNeXt
✅ ME-NeSt
✅ VajraME-NeSt
✅ MixConvNeXt
❌ ViT (Coming Soon!)
❌ Swin (Coming Soon!)
❌ SwinV2 (Coming Soon!)
To be published
- https://github.com/ultralytics/ultralytics
- https://github.com/ultralytics/yolov5
- https://github.com/ouyanghaodong/DEYOv1.5
- https://github.com/WongKinYiu/yolov9
- https://github.com/meituan/YOLOv6
- https://github.com/huggingface/pytorch-image-models
- https://github.com/pytorch/vision
Vayuvahana Technologies Private Limited offers two licensing options:
-
AGPL-3.0 License: This is an OSI-approved open-source license for researchers for the purpose of promoting collaboration. See the LICENSE file for details.
-
Enterprise License: This license is designed for commercial use and enables integration of VayuAI software and AI models into commercial goods and services, bypassing the open-source requirements of AGPL-3.0. If your product requires embedding the software for commercial purposes or require access to more capable enterprise AI models in the future, reach out via Email.