Detectron2 wrapper for DETR
=======
We provide a Detectron2 wrapper for DETR, thus providing a way to better integrate it in the existing detection ecosystem. It can be used for example to easily leverage datasets or backbones provided in Detectron2.
This wrapper currently supports only box detection, and is intended to be as close as possible to the original implementation, and we checked that it indeed match the results. Some notable facts and caveats:
- The data augmentation matches DETR's original data augmentation. This required patching the RandomCrop augmentation from Detectron2, so you'll need a version from the master branch from June 24th 2020 or more recent.
- To match DETR's original backbone initialization, we use the weights of a ResNet50 trained on imagenet using torchvision. This network uses a different pixel mean and std than most of the backbones available in Detectron2 by default, so extra care must be taken when switching to another one. Note that no other torchvision models are available in Detectron2 as of now, though it may change in the future.
- The gradient clipping mode is "full_model", which is not the default in Detectron2.
# Usage
To install Detectron2, please follow the [official installation instructions](https://github.com/facebookresearch/detectron2/blob/master/INSTALL.md).
## Evaluating a model
For convenience, we provide a conversion script to convert models trained by the main DETR training loop into the format of this wrapper. To download and convert the main Resnet50 model, simply do:
```
python converter.py --source_model https://dl.fbaipublicfiles.com/detr/detr-r50-e632da11.pth --output_model converted_model.pth
```
You can then evaluate it using:
```
python train_net.py --eval-only --config configs/detr_256_6_6_torchvision.yaml MODEL.WEIGHTS "converted_model.pth"
```
## Training
To train DETR on a single node with 8 gpus, simply use:
```
python train_net.py --config configs/detr_256_6_6_torchvision.yaml --num-gpus 8
```
To fine-tune DETR for instance segmentation on a single node with 8 gpus, simply use:
```
python train_net.py --config configs/detr_segm_256_6_6_torchvision.yaml --num-gpus 8 MODEL.DETR.FROZEN_WEIGHTS <model_path>
```
目标检测源码解读1111
需积分: 5 186 浏览量
2022-12-26
20:01:54
上传
评论
收藏 108KB ZIP 举报
detr目标检测源码解读.zip （50个子文件） 
detr-master 
test_all.py 9KB .circleci config.yml 703B main.py 11KB hubconf.py 6KB tox.ini 65B datasets __init__.py 897B coco.py 5KB panoptic_eval.py 1KB transforms.py 8KB coco_panoptic.py 4KB __pycache__ panoptic_eval.cpython-36.pyc 2KB coco.cpython-36.pyc 5KB transforms.cpython-36.pyc 9KB __init__.cpython-36.pyc 849B coco_eval.cpython-36.pyc 7KB coco_eval.py 9KB models __init__.py 143B segmentation.py 15KB position_encoding.py 4KB matcher.py 4KB backbone.py 4KB detr.py 17KB transformer.py 13KB __pycache__ transformer.cpython-36.pyc 9KB position_encoding.cpython-36.pyc 4KB segmentation.cpython-36.pyc 13KB matcher.cpython-36.pyc 4KB detr.cpython-36.pyc 15KB __init__.cpython-36.pyc 288B backbone.cpython-36.pyc 5KB .gitignore 189B engine.py 6KB __pycache__ engine.cpython-36.pyc 6KB util __init__.py 71B box_ops.py 3KB misc.py 15KB plot_utils.py 4KB __pycache__ box_ops.cpython-36.pyc 3KB misc.cpython-36.pyc 14KB __init__.cpython-36.pyc 141B d2 detr __init__.py 176B dataset_mapper.py 4KB detr.py 11KB config.py 888B configs detr_256_6_6_torchvision.yaml 1012B detr_segm_256_6_6_torchvision.yaml 1KB converter.py 3KB README.md 2KB train_net.py 5KB run_with_submitit.py 3KB资源评论
