Skip to content

Latest commit

 

History

History
314 lines (221 loc) · 29.9 KB

README.md

File metadata and controls

314 lines (221 loc) · 29.9 KB
Raw

Model Zoo

All experiments are conducted on servers with 8 NVIDIA V100 / 2080Ti GPUs (PCIE). The software in use were PyTorch 1.3, CUDA 10.1, cuDNN 7.6.3.

ImageNet Classification

Comming Soon.

Self-supervised Learning

Comming Soon.

Object Detection

COCO

Faster R-CNN

Name input size lr sched train time (s/iter) train mem (GB) box AP Trained Model
FasterRCNN-R50-FPN 640-800 90k 0.225(2080ti) 2.82 38.1 LINK
FasterRCNN-R50-FPN-SyncBN 640-800 180k 0.546 5.23 39.9 LINK
FasterRCNN-ResNeSt50-FPN 800 90k 0.416 3.53 39.9 LINK
FasterRCNN-ResNeSt50-FPN-SyncBN 640-800 90k 0.661 5.35 42.5 LINK
FasterRCNN-MOBILENET-FPN 640-800 90k 0.279(2080ti) 3.47 29.27 LINK
FasterRCNN-MOBILENET-FPN-NoP2 640-800 90k 0.227(2080ti) 2.49 29.57 LINK

RetinaNet

Name input size lr sched train time (s/iter) train mem (GB) box AP Trained Model
RetinaNet-R50 800 90k 0.3593 3.85 35.9 LINK
RetinaNet-R50 640-800 90k 0.244 3.84 36.5 LINK
RetinaNet-R50 640-800 90k 0.344(2080ti) 3.96 37.2 LINK
RetinaNet-R50-DRLoss 800 90k 0.357(2080ti) 3.72 37.4 LINK
RetinaNet-MOBILENET 640-800 90k 0.266 4.46 28.5 LINK

FCOS

Name input size lr sched train time (s/iter) train mem (GB) box AP Trained Model
FCOS-R50-FPN 800 90k 0.334(2080ti) 3.09 38.8 LINK

ATSS

Name input size lr sched train time (s/iter) train mem (GB) box AP Trained Model
ATSS-R50-FPN 800 90k 0.340(2080ti) 3.09 39.3 LINK

FreeAnchor

Name input size lr sched train time (s/iter) train mem (GB) box AP Trained Model
FreeAnchor-R50-FPN 800 90k 0.353(2080ti) 4.08 38.3 LINK

TridentNet

Name input size lr sched train time (s/iter) train mem (GB) box AP Trained Model
TridentNet-R50-C4 800 90k 0.754(2080ti) 4.65 37.7 LINK

RepPoints

Name input size lr sched train time (s/iter) train mem (GB) box AP Trained Model
RepPoints-R50-FPN 800 90k 0.415(2080ti) 2.85 38.2 LINK

CenterNet

Name input size lr sched train time (s/iter) train mem (GB) box AP Trained Model
CenterNet-R18 512 126k TBD TBD 29.8
TBD)
CenterNet-R50 512 126k TBD TBD 34.9
TBD)
CenterNet-R101 512 126k TBD TBD 36.8
TBD)

EfficientDet

Name input size lr sched train time (s/iter) train mem (GB) box AP Trained Model
EffDet0-Effnet0-BiFPN 512 562k 0.540(2080ti) 5.77 32.6 LINK
EffDet0-Effnet0-BiFPN-SyncBN 512 562k 0.760(2080ti) 5.77 33.2 LINK
EffDet1-Effnet1-BiFPN 640 562k 0.782(v100) 23.18 38.1 LINK
EffDet1-Effnet1-BiFPN-SyncBN 640 562k 1.182(v100) 23.18 38.0 LINK

YOLO

Name input size lr sched train time (s/iter) train mem (GB) box AP Trained Model
YOLOv3-Darknet53-SyncBN 320-608 470k 0.729 7.45 37.5 LINK

SSD

Name input size lr sched train time (s/iter) train mem (GB) box AP Trained Model
SSD-VGG16 300 200k 0.442 1.93 23.6 LINK
SSD-VGG16-Expand 300 200k 0.448 1.93 24.9 LINK
SSD-VGG16 512 200k 0.487 4.37 26.7 LINK
SSD-VGG16-Expand 512 200k 0.491 4.37 29.0 LINK

DETR

Name input size lr sched train time (s/iter) train mem (GB) box AP Trained Model
DETR-R50-C5 480-800 150e 0.270(v100) 3.62 38.7 LINK

Sparse R-CNN

Name input size lr sched train time (s/iter) train mem (GB) box AP Trained Model
SparseRCNN-R50-FPN 480-800 270k 0.627(2080ti) 4.11 43.2 LINK

PASCAL VOC

Name input size lr sched train time (s/iter) train mem (GB) AP AP50 AP75 Trained Model
FasterRCNN-R50-FPN 480-800 18k 0.377 2.82 54.2 82.1 59.3 LINK

WIDER FACE

Name input size lr sched train time (s/iter) train mem (GB) box AP Trained Model
RetinaNet-R50 600 45k 0.342 4.76 49.4 LINK
FCOS-R50-FPN 600 45k 0.382 5.75 50.8 LINK

CityPersons

Name input size lr sched train time (s/iter) train mem (GB) box AP MR Trained Model
FasterRCNN-R50-FPN 640 9K 0.401 3.38 36.1 0.37 LINK
RetinaNet-R50 640 18k 0.349 2.97 33.6 0.42 LINK
FCOS-R50-FPN 640 9K 0.375 3.55 35.7 0.40 LINK

CrowdHuman

Name input size lr sched train time (s/iter) train mem (GB) box AP MR Trained Model
FasterRCNN-R50-FPN 800 2.8K 0.856 4.80 84.1 0.481 LINK

Instance Segmentation

COCO

Mask R-CNN

Name input size lr sched train time (s/iter) train mem (GB) box AP mask AP Trained Model
MaskRCNN-R50-C4 640-800 90k 0.609 5.04 36.8 32.2 LINK
MaskRCNN-R50-C4-SyncBN-ExtraNorm 640-800 90k 0.852 9.82 37.9 33.1 LINK
MaskRCNN-R50-C4-SyncBN 640-800 180k 0.837 9.82 39.9 34.5 LINK)
MaskRCNN-R50-C4-SyncBN-ExtraNorm 640-800 180k 0.853 9.82 40.1 34.7 LINK
MaskRCNN-R50-FPN 640-800 90k 0.297(2080ti) 3.36 38.5 35.2 LINK

TensorMask

Name input size lr sched train time (s/iter) train mem (GB) box AP mask AP Trained Model
TensorMask-R50-FPN 800 90k 0.788(2080ti) 7.83 37.5 32.3 LINK

CascadeRCNN

Name input size lr sched train time (s/iter) train mem (GB) box AP mask AP Trained Model
CascadeRCNN-R50-FPN 800 90k 0.546 3.91 41.7 36.1 LINK

PointRend

Name input size lr sched train time (s/iter) train mem (GB) box AP mask AP Trained Model
PointRend-R50-FPN 640-800 90k 0.439 4.88 38.4 36.2 LINK
PointRend-R50-FPN 640-800 270k 0.416 4.88 41.1 38.2 LINK

SOLO

Name input size lr sched train time (s/iter) train mem (GB) box AP mask AP Trained Model
SOLO-R50-FPN 800 90k 0.970 6.99 33.1 32.7 LINK
SOLO-R50-FPN 640-800 270k 0.950 6.99 35.6 35.2 LINK
DecoupledSOLO-R50-FPN 800 90k 1.097 6.68 34.0 33.7 LINK
DecoupledSOLO-R50-FPN 640-800 270k 0.922 6.47 35.9 35.6 LINK

LVIS

Name input size lr sched train time (s/iter) train mem (GB) box AP mask AP Trained Model
MaskRCNN-R50-FPN 800 90k 0.486 5.26 20.3 21.0 LINK
MaskRCNN-R50-FPN-DataResampling 800 90k 0.500 5.26 23.0 23.1 LINK
MaskRCNN-R50-FPN-DataResampling 640-800 90k 0.485 5.25 24.1 24.7 LINK

CITYSCAPES

Name input size lr sched train time (s/iter) train mem (GB) mask AP
MaskRCNN-R50-FPN 640-800 90k 0.737 5.21 37.4 LINK
PointRend-R50-FPN 800-1024 240k 0.746 8.21 36.0 LINK

Semantic Segmentation

COCO

SemanticFPN

Name input size lr sched train time (s/iter) train mem (GB) mIoU Trained Model
SemanticFPN-R50-FPN 640-800 90k 0.285 6.16 40.3 LINK

CITYSCAPES

PointRend

Name input size lr sched train time (s/iter) train mem (GB) mIoU Trained Model
PointRend-R101-FPN 512-2048 65k 1.900 3.88 78.2 LINK

DynamicRouting

Name input size lr sched train time (s/iter) train mem (GB) mIoU Trained Model
Dynamic-A 512-2048 190k 0.736 8.74 75.7 LINK
Dynamic-B 512-2048 190k 0.706 8.74 75.3 LINK
Dynamic-C 512-2048 190k 0.717 8.74 76.2 LINK
Dynamic-Raw 512-2048 190k 0.757 8.73 76.5 LINK

FCN

Name input size lr sched train time (s/iter) train mem (GB) mIoU Trained Model
FCN-Res101-s32 512-2048 65k 0.605 3.44 71.9 LINK
FCN-Res101-s16 512-2048 65k 0.593 3.41 73.5 LINK
FCN-Res101-s8 512-2048 65k 0.541 3.41 74.0 LINK

Panoptic Segmentation

COCO

PanopticFPN

Name input size lr sched train time (s/iter) train mem (GB) PG Trained Model
PanopticFPN-R50-FPN-800 800 90k 0.4842 4.74 39.4 LINK
PanopticFPN-R50-FPN-MS 640-800 90k 0.4657 4.74 39.5 LINK

Key-Points

COCO_PERSON

Keypoint-RCNN

Named input size lr sched train time (s/iter) train mem (GB) box AP kp AP
RCNN_R50_FPN 480-800 90k 0.4r0(2080Ti) 4.47 53.7 64.2

3D

Comming Soon.