Skip to content

Latest commit

 

History

History
executable file
·
445 lines (301 loc) · 26.1 KB

results.md

File metadata and controls

executable file
·
445 lines (301 loc) · 26.1 KB

Image classification task results

ResNet

Paper:https://arxiv.org/abs/1512.03385

DarkNet

Paper:https://arxiv.org/abs/1804.02767?e05802c1_page=1

RepVGG

Paper:https://arxiv.org/abs/2101.03697

RegNet

Paper:https://arxiv.org/abs/2003.13678

ViT

Paper:https://arxiv.org/abs/2010.11929

ResNetCifar training from scratch on CIFAR100

ResNetCifar is different from ResNet in the first few layers.

Network macs params input size gpu num batch epochs Top-1
ResNet18Cifar 557.935M 11.220M 32x32 1 RTX A5000 128 200 77.110
ResNet34Cifar 1.164G 21.328M 32x32 1 RTX A5000 128 200 78.140
ResNet50Cifar 1.312G 23.705M 32x32 1 RTX A5000 128 200 75.610
ResNet101Cifar 2.531G 42.697M 32x32 1 RTX A5000 128 200 76.970
ResNet152Cifar 3.751G 58.341M 32x32 1 RTX A5000 128 200 77.710

You can find more model training details in classification_training/cifar100/.

ResNet training from scratch on ImageNet1K(ILSVRC2012)

Network macs params input size gpu num batch epochs Top-1
ResNet18 1.819G 11.690M 224x224 2 RTX A5000 256 100 70.512
ResNet34 3.671G 21.798M 224x224 2 RTX A5000 256 100 73.680
ResNet50 4.112G 25.557M 224x224 2 RTX A5000 256 100 76.300
ResNet101 7.834G 44.549M 224x224 2 RTX A5000 256 100 77.380
ResNet152 11.559G 60.193M 224x224 2 RTX A5000 256 100 77.542

You can find more model training details in classification_training/imagenet/.

DarkNet training from scratch on ImageNet1K(ILSVRC2012)

Network macs params input size gpu num batch epochs Top-1
DarkNetTiny 412.537M 2.087M 256x256 2 RTX A5000 256 100 57.786
DarkNet19 3.663G 20.842M 256x256 2 RTX A5000 256 100 74.248
DarkNet53 9.322G 41.610M 256x256 2 RTX A5000 256 100 76.352

You can find more model training details in classification_training/imagenet/.

RepVGG training from scratch on ImageNet1K(ILSVRC2012)

Network macs params input size gpu num batch epochs Top-1
RepVGG_A0_deploy 1.362G 8.309M 224x224 2 RTX A5000 256 120 72.010
RepVGG_A1_deploy 2.364G 12.790M 224x224 2 RTX A5000 256 120 74.032
RepVGG_A2_deploy 5.117G 25.500M 224x224 2 RTX A5000 256 120 76.078
RepVGG_B0_deploy 3.058G 14.339M 224x224 2 RTX A5000 256 120 74.880
RepVGG_B1_deploy 11.816G 51.829M 224x224 2 RTX A5000 256 120 77.790
RepVGG_B2_deploy 18.377G 80.315M 224x224 2 RTX A5000 256 120 78.120

You can find more model training details in classification_training/imagenet/.

RegNet training from scratch on ImageNet1K(ILSVRC2012)

Network macs params input size gpu num batch epochs Top-1
RegNetX_400MF 410.266M 5.158M 224x224 2 RTX A5000 4096 300 69.466
RegNetX_600MF 616.813M 6.196M 224x224 2 RTX A5000 4096 300 71.754
RegNetX_800MF 820.324M 7.260M 224x224 2 RTX A5000 4096 300 73.148
RegNetX_1_6GF 1.635G 9.190M 224x224 2 RTX A5000 4096 300 76.142
RegNetX_3_2GF 3.222G 15.297M 224x224 2 RTX A5000 4096 300 78.244
RegNetX_4_0GF 4.013G 22.118M 224x224 2 RTX A5000 4096 300 78.916

You can find more model training details in classification_training/imagenet/.

U2NetBackbone training from scratch on ImageNet1K(ILSVRC2012)

Network macs params input size gpu num batch epochs Top-1
U2NetBackbone 13.097G 26.181M 224x224 2 RTX A5000 256 100 76.038

You can find more model training details in classification_training/imagenet/.

ResNet finetune from ImageNet21k pretrain weight on ImageNet1K(ILSVRC2012)

Network macs params input size gpu num batch epochs Top-1
ResNet18 1.819G 11.690M 224x224 2 RTX A5000 4096 300 71.580
ResNet34 3.671G 21.798M 224x224 2 RTX A5000 4096 300 76.316
ResNet50 4.112G 25.557M 224x224 2 RTX A5000 4096 300 79.484
ResNet101 7.834G 44.549M 224x224 2 RTX A5000 4096 300 80.940
ResNet152 11.559G 60.193M 224x224 2 RTX A5000 4096 300 81.236

You can find more model training details in classification_training/imagenet/.

ResNet finetune from DINO pretrain weight on ImageNet1K(ILSVRC2012)

Network macs params input size gpu num batch epochs Top-1
ResNet50 4.112G 25.557M 224x224 2 RTX A5000 256 100 77.114
ResNet50 4.112G 25.557M 224x224 2 RTX A5000 4096 300 79.418

You can find more model training details in classification_training/imagenet/.

ViT finetune from self-trained MAE pretrain weight(400epoch) on ImageNet1K(ILSVRC2012)

Network macs params input size gpu num batch epochs Top-1
ViT-Small-Patch16 4.241G 21.955M 224x224 2 RTX A5000 4096 100 79.006
ViT-Base-Patch16 16.849G 86.377M 224x224 2 RTX A5000 4096 100 83.204
ViT-Large-Patch16 59.647G 304.024M 224x224 2 RTX A5000 4096 100 85.020

You can find more model training details in classification_training/imagenet/.

ViT finetune from offical MAE pretrain weight(800 epoch) on ImageNet1K(ILSVRC2012)

Network macs params input size gpu num batch epochs Top-1
ViT-Base-Patch16 16.849G 86.377M 224x224 2 RTX A5000 4096 100 83.290
ViT-Large-Patch16 59.647G 304.024M 224x224 2 RTX A5000 4096 100 85.876

You can find more model training details in classification_training/imagenet/.

ResNet train from ImageNet1K pretrain weight on ImageNet21K(Winter 2021 release)

Network macs params input size gpu num batch epochs Semantic Softmax Acc
ResNet18 1.819G 11.690M 224x224 2 RTX A5000 4096 80 68.639
ResNet34 3.671G 21.798M 224x224 2 RTX A5000 4096 80 71.873
ResNet50 4.112G 25.557M 224x224 2 RTX A5000 4096 80 74.664
ResNet101 7.834G 44.549M 224x224 2 RTX A5000 4096 80 76.136
ResNet152 11.559G 60.193M 224x224 2 RTX A5000 4096 80 75.731

You can find more model training details in classification_training/imagenet21k/.

Object detection task results

RetinaNet

Paper:https://arxiv.org/abs/1708.02002

FCOS

Paper:https://arxiv.org/abs/1904.01355

CenterNet

Paper:https://arxiv.org/abs/1904.07850

TTFNet

Paper:https://arxiv.org/abs/1909.00700

All detection models training from scratch on COCO2017

ResNet50 backbone using DINO pretrained weight on ImageNet1K.

Trained on COCO2017 train dataset, tested on COCO2017 val dataset.

mAP is IoU=0.5:0.95,area=all,maxDets=100,mAP(COCOeval,stats[0]).

Network resize-style input size macs params gpu num batch epochs mAP
ResNet50-RetinaNet YoloStyle-640 640x640 95.558G 37.969M 2 RTX A5000 32 13 34.459
ResNet50-RetinaNet YoloStyle-800 800x800 149.522G 37.969M 2 RTX A5000 32 13 36.023
ResNet50-RetinaNet RetinaStyle-800 800x1333 250.069G 37.969M 2 RTX A5000 8 13 35.434
ResNet50-FCOS YoloStyle-640 640x640 81.943G 32.291M 2 RTX A5000 32 13 37.176
ResNet50-FCOS YoloStyle-800 800x800 128.160G 32.291M 2 RTX A5000 32 13 38.745
ResNet50-FCOS RetinaStyle-800 800x1333 214.406G 32.291M 2 RTX A5000 8 13 39.649
ResNet18DCN-CenterNet YoloStyle-512 512x512 14.854G 12.889M 2 RTX A5000 64 140 26.209
ResNet18DCN-TTFNet-3x YoloStyle-512 512x512 16.063G 13.737M 2 RTX A5000 64 39 27.054

You can find more model training details in detection_training/coco/.

All detection models finetune from objects365 pretrain weight on COCO2017

Trained on COCO2017 train dataset, tested on COCO2017 val dataset.

mAP is IoU=0.5:0.95,area=all,maxDets=100,mAP(COCOeval,stats[0]).

Network resize-style input size macs params gpu num batch epochs mAP
ResNet50-RetinaNet YoloStyle-640 640x640 95.558G 37.969M 2 RTX A5000 32 13 38.930
ResNet50-RetinaNet YoloStyle-800 800x800 149.522G 37.969M 2 RTX A5000 32 13 40.483
ResNet50-RetinaNet RetinaStyle-800 800x1333 250.069G 37.969M 2 RTX A5000 8 13 40.424
ResNet50-FCOS YoloStyle-640 640x640 81.943G 32.291M 2 RTX A5000 32 13 42.871
ResNet50-FCOS YoloStyle-800 800x800 128.160G 32.291M 2 RTX A5000 32 13 44.526
ResNet50-FCOS RetinaStyle-800 800x1333 214.406G 32.291M 2 RTX A5000 8 13 42.848

You can find more model training details in detection_training/coco/.

All detection models train from COCO2017 pretrain weight on Objects365(v2,2020)

Trained on objects365 train dataset, tested on objects365 val dataset.

mAP is IoU=0.5:0.95,area=all,maxDets=100,mAP(COCOeval,stats[0]).

Network resize-style input size macs params gpu num batch epochs mAP
ResNet50-RetinaNet YoloStyle-800 800x800 149.522G 37.969M 8 RTX A5000 32 13 16.360
ResNet50-FCOS RetinaStyle-800 800x1333 214.406G 32.291M 8 RTX A5000 32 13 17.068

All detection models training from scratch on VOC2007 and VOC2012

Trained on VOC2007 trainval dataset + VOC2012 trainval dataset, tested on VOC2007 test dataset.

mAP is IoU=0.50,area=all,maxDets=100,mAP.

Network resize-style input size macs params gpu num batch epochs mAP
ResNet50-RetinaNet YoloStyle-640 640x640 84.947G 36.724M 2 RTX A5000 32 13 81.948
ResNet50-FCOS YoloStyle-640 640x640 80.764G 32.153M 2 RTX A5000 32 13 81.624

You can find more model training details in detection_training/voc/.

All detection models finetune from objects365 pretrain weight on VOC2007 and VOC2012

Trained on VOC2007 trainval dataset + VOC2012 trainval dataset, tested on VOC2007 test dataset.

mAP is IoU=0.50,area=all,maxDets=100,mAP.

Network resize-style input size macs params gpu num batch epochs mAP
ResNet50-RetinaNet YoloStyle-640 640x640 84.947G 36.724M 2 RTX A5000 32 13 90.220
ResNet50-FCOS YoloStyle-640 640x640 80.764G 32.153M 2 RTX A5000 32 13 90.371

You can find more model training details in detection_training/voc/.

Semantic Segmentation task results

DeepLabv3+

Paper:https://arxiv.org/abs/1802.02611

U2Net

Paper:https://arxiv.org/abs/2005.09007

All semantic segmentation models training from scratch on ADE20K

Network input size macs params gpu num batch epochs miou
ResNet50-DeepLabv3+ 512x512 25.548G 26.738M 2 RTX A5000 8 128 34.659
U2Net 512x512 219.012G 46.191M 2 RTX A5000 8 128 39.046

You can find more model training details in semantic_segmentation_training/ade20k/.

All semantic segmentation models training from scratch on COCO2017

Network input size macs params gpu num batch epochs miou
ResNet50-DeepLabv3+ 512x512 25.548G 26.738M 2 RTX A5000 32 64 64.176
U2Net 512x512 219.012G 46.191M 4 RTX A5000 32 64 66.529

You can find more model training details in semantic_segmentation_training/coco/.

Knowledge distillation task results

KD loss

Paper:https://arxiv.org/abs/1503.02531

DML loss

Paper:https://arxiv.org/abs/1706.00384

ResNet training from pretrain weight on ImageNet1K(ILSVRC2012)

Teacher Network Student Network method Freeze Teacher input size gpu num batch epochs Teacher Top-1 Student Top-1
ResNet152 ResNet50 CE+KD True 224x224 2 RTX A5000 256 100 / 77.352
ResNet152 ResNet50 CE+DML False 224x224 2 RTX A5000 256 100 79.274 78.122
ResNet152 ResNet50 CE+KD+Vit Aug True 224x224 2 RTX A5000 4096 300 / 80.168
ResNet152 ResNet50 CE+DML+Vit Aug False 224x224 2 RTX A5000 4096 300 81.508 79.810

You can find more model training details in distillation_training/imagenet/.

Contrastive learning task results

DINO:Emerging Properties in Self-Supervised Vision Transformers

Paper:https://arxiv.org/abs/2104.14294

ResNet DINO pretrain on ImageNet1K(ILSVRC2012)

Network input size gpu num batch epochs Loss
ResNet50-DINO 224x224 4 RTX A5000 256 400 1.997

You can find more model training details in contrastive_learning_training/imagenet/.

ResNet finetune from DINO pretrain weight on ImageNet1K(ILSVRC2012)

Network macs params input size gpu num batch epochs Top-1
ResNet50 4.112G 25.557M 224x224 2 RTX A5000 256 100 77.114
ResNet50 4.112G 25.557M 224x224 2 RTX A5000 4096 300 79.418

You can find more model training details in classification_training/imagenet/.

Masked image modeling task results

MAE:Masked Autoencoders Are Scalable Vision Learners

Paper:https://arxiv.org/abs/2111.06377

ViT MAE pretrain on ImageNet1K(ILSVRC2012)

Network input size gpu num batch epochs Loss
ViT-Small-Patch16 224x224 2 RTX A5000 256 400 0.414
ViT-Base-Patch16 224x224 2 RTX A5000 256 400 0.388
ViT-Large-Patch16 224x224 2 RTX A5000 256 400 0.378

You can find more model training details in masked_image_modeling_training/imagenet/.

ViT finetune from self-trained MAE pretrain weight(400epoch) on ImageNet1K(ILSVRC2012)

Network macs params input size gpu num batch epochs Top-1
ViT-Small-Patch16 4.241G 21.955M 224x224 2 RTX A5000 4096 100 79.006
ViT-Base-Patch16 16.849G 86.377M 224x224 2 RTX A5000 4096 100 83.204
ViT-Large-Patch16 59.647G 304.024M 224x224 2 RTX A5000 4096 100 85.020

You can find more model training details in classification_training/imagenet/.

ViT finetune from offical MAE pretrain weight(800 epoch) on ImageNet1K(ILSVRC2012)

Network macs params input size gpu num batch epochs Top-1
ViT-Base-Patch16 16.849G 86.377M 224x224 2 RTX A5000 4096 100 83.290
ViT-Large-Patch16 59.647G 304.024M 224x224 2 RTX A5000 4096 100 85.876

You can find more model training details in classification_training/imagenet/.

Diffusion model task results

Denoising Diffusion Probabilistic Models

Paper:https://arxiv.org/abs/2006.11239

Denoising Diffusion Implicit Models

Paper:https://arxiv.org/abs/2010.02502

High-Resolution Image Synthesis with Latent Diffusion Models

Paper:https://arxiv.org/abs/2112.10752

All diffusion model with different sampling methods on CIFAR10

Trained diffusion unet on CIFAR10 dataset(DDPM method).Test image num=50000.

sampling method input size steps condition label(train/test) FID IS score(mean/std)
DDPM 32x32 1000 False/False 5.394 8.684/0.169
DDIM 32x32 50 False/False 7.644 8.642/0.129
PLMS 32x32 20 False/False 7.027 8.834/0.200
DDPM 32x32 1000 True/True 3.949 8.985/0.139

You can find more model training details in diffusion_model_training/cifar10/.

All diffusion model with different sampling methods on CIFAR100

Trained diffusion unet on CIFAR100 dataset(DDPM method).Test image num=50000.

sampling method input size steps condition label(train/test) FID IS score(mean/std)
DDPM 32x32 1000 False/False 9.620 9.399/0.138
DDIM 32x32 50 False/False 13.250 8.946/0.150
PLMS 32x32 20 False/False 11.854 9.391/0.202
DDPM 32x32 1000 True/True 5.209 10.880/0.180

You can find more model training details in diffusion_model_training/cifar100/.

All diffusion model with different sampling methods on CelebA-HQ

Trained diffusion unet on CelebA-HQ dataset(DDPM method).Test image num=28000.

sampling method input size steps condition label(train/test) FID IS score(mean/std)
DDPM 64x64 1000 False/False 6.491 2.577/0.035
DDIM 64x64 50 False/False 15.195 2.625/0.028
PLMS 64x64 20 False/False 18.061 2.701/0.040

You can find more model training details in diffusion_model_training/celebahq/.

All diffusion model with different sampling methods on FFHQ

Trained diffusion unet on FFHQ dataset(DDPM method).Test image num=60000.

sampling method input size steps condition label(train/test) FID IS score(mean/std)
DDPM 64x64 1000 False/False 6.671 3.399/0.055
DDIM 64x64 50 False/False 10.479 3.431/0.044
PLMS 64x64 20 False/False 12.387 3.462/0.034

You can find more model training details in diffusion_model_training/ffhq/.