Hao-Gong/pytorch_detectron

Pytorch detectron including efficientDetD0-D7, fpn, cascade rcnn, faster_rcnn,yoloV3,yoloV4,yoloV4-tiny

pytorch > 1.0

setup：

python setup.py install

support efficientDetD0-D7, fpn, cascade rcnn, faster_rcnn,yoloV3,yoloV4,yoloV4-tiny

name	built	test
efficientdetD0	Yes	Yes
efficientdetD1	Yes	Yes
efficientdetD2	Yes	Yes
efficientdetD3	Yes	Yes
efficientdetD4	Yes	Yes
efficientdetD5	Yes	Yes
efficientdetD6	Yes	Yes
efficientdetD7	Yes	Yes
fpn resnet18	Yes	Yes
fpn resnet34	Yes	Yes
fpn resnet50	Yes	Yes
fpn resnet101	Yes	Yes
fpn resnet152	Yes	Yes
cascade_fpn resnet18	Yes	Yes
cascade_fpn resnet34	Yes	Yes
cascade_fpn resnet50	Yes	Yes
cascade_fpn resnet101	Yes	Yes
cascade_fpn resnet152	Yes	Yes
faster_rcnn resnet18	Yes	Yes
faster_rcnn resnet34	Yes	Yes
faster_rcnn resnet50	Yes	Yes
faster_rcnn resnet101	Yes	Yes
faster_rcnn resnet152	Yes	Yes
yolov3	Yes	Yes
yolov4	Yes	Yes
yolov4-tiny	Yes	Yes

dataset augmentation：

train_dataset_transforms	Function
BrightJitter	-
HueSaturationJitter	-
RandomCroper	-
Mosaic	-
Normalizer	-
HorizontalFliper	-
VerticalFlipe	-
RandomBlur	-
NoiseAdder	-
RandomAffiner	-
TransformAnnotXYWH	bbox XminYminXmaxYmax -> XYWH inside 0-1

because of lack of training resource, some test only done on VOC2007 trainval/Test

model	#GPUs	batch size	lr	lr_decay	max_epoch	time/epoch	mem/GPU	mAP
[Res-101-Faster-Rcnn-800]	1	4	1e-3	-	-	- hr	- MB	71.1%
[Res-101-FPN-800]	1	8	1e-3	-	-	- hr	- MB	74.3%
[Res-50-FPN-800]	1	8	1e-3	-	-	- hr	- MB	71.4%
[Res-101-Cascader-Rcnn-800]	1	4	1e-3	-	-	- hr	- MB	72.7%
[EfficientDet-D0]	1	4	1e-3	-	-	- hr	- MB	48.9%
[EfficientDet-D1]	1	4	1e-3	-	-	- hr	- MB	52.1%
[YoloV4-tiny-608]	1	4	1e-4	-	-	- hr	23.7 MB	53.1%
[YoloV4-608]	1	4	1e-4	-	-	- hr	257 MB	70.3%
[YoloV3-608]	1	4	1e-4	-	-	- hr	247 MB	65.7%

VOC2007+2012train/VOC2007Test

model	#GPUs	batch size	lr	lr_decay	max_epoch	time/epoch	mem/GPU	mAP
[Res-101-FPN-800]	1	1	1e-3	-	-	- hr	- MB	82.9%
[Res-101-Faster-Rcnn-800]	1	1	1e-3	-	-	- hr	- MB	78.8%

COCO Dataloader is also support, you can modified it yourself