Ecosyste.ms: Awesome
An open API service indexing awesome lists of open source software.
https://github.com/cgarciae/simple-detr
https://github.com/cgarciae/simple-detr
Last synced: 7 days ago
JSON representation
- Host: GitHub
- URL: https://github.com/cgarciae/simple-detr
- Owner: cgarciae
- License: mit
- Created: 2020-05-31T19:10:07.000Z (over 4 years ago)
- Default Branch: master
- Last Pushed: 2022-11-22T06:10:25.000Z (almost 2 years ago)
- Last Synced: 2024-05-02T01:24:39.309Z (7 months ago)
- Language: Python
- Size: 441 KB
- Stars: 8
- Watchers: 3
- Forks: 6
- Open Issues: 10
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# Simple DETR
This codebase simplifies the access to the [DETR](https://github.com/facebookresearch/detr) model, specifically it provides:
* A configurable `DETR` class which gives you access to different backbones.
* It inherits from `nn.Module` if you want to mix it with other pytorch code.
* A `predict` method which performs lightweight postprocessing and gives you back bounding boxes and scores.
This codebase takes the code from the [standalone notebook](https://colab.research.google.com/github/facebookresearch/detr/blob/colab/notebooks/detr_demo.ipynb) found on the official repo and structures it more like what you would use for a simple deployment or CLI tool.
### Installation
Using pip:
```bash
pip install -r requirements.txt
```
Using poetry:
```bash
poetry install
```
### Evaluating Samples
If you want to play around just with the model just put images in the `test-images` folder and run:
```bash
python main.py
```
This will pop-up matplotlib plots of the images with the bounding box information overlaid:
Options:
```
Usage: main.py [OPTIONS]
Options:
--images-path PATH (default = test-images)
--backbone TEXT (default = detr_resnet50)
--threshold FLOAT (default = 0.7)
--device TEXT (default = cpu)
```
### API Reference
You can use the `DETR` class on your own Python / PyTorch code. For example:
```python
from detr import DETR
from PIL import Image
model = DETR(backbone="detr_resnet50", threshold=0.7, device="cpu")
image = Image.open("path/to/image.png")
scores, boxes = model.predict(image)
```
#### class DETR
```python
def __init__(
self,
threshold: float = 0.7,
backbone: str = "detr_resnet50",
device: str = "cpu",
pretrained: bool = True,
**kwargs,
):
"""
Arguments:
threshold: (float = 0.7) probability threshold required to keep a box.
backbone: (str = detr_resnet50) one of:
* detr_resnet50
* detr_resnet50_dc5
* detr_resnet101
* detr_resnet101_dc5
device: (str = cpu) device to use e.g. "cpu", "cuda", etc
pretrained: (bool = True)
**kwargs: backbone options, for more information see https://github.com/facebookresearch/detr/blob/master/hubconf.py
"""
```
#### DETR.forward
```python
def forward(self, image: torch.Tensor) -> tp.Dict[str, torch.Tensor]:
```
#### DETR.predict
```python
def predict(self, image: Image) -> tp.Tuple[torch.Tensor, torch.Tensor]:
"""
Calculates the scores and bounding boxes for an image.
Arguments:
image: (PIL.Image) image sample.
Returns:
(scores, boxes): (Tuple[Tensor, Tensor]) tensors of the scores and boxes.
"""
```