Several ImageNet pretrained models

Name	Model	Size	Augmentation	Epoch	Top 1	Top 5	Loss	Time	Code
resnetvd50_nls	ResNet-50_vd	224	Standard	120	77.89	93.77	0.8920	354.0	469
resnet50	ResNet-50	224	Standard	120	77.34	93.63	1.9400	289.0	472
resnet50_nls	ResNet-50	224	Standard	120	76.92	93.46	0.9748	289.0	470
resnet50_nls2	ResNet-50	224	Standard	120	76.93	93.28	0.9702	289.0	468

• nls means no label smoothing because it may hurt transfer learning in detection and segmentation
• nls2 does not use zero γ

Models for new Re-ResNet

Name	Model	Size	Augmentation	Epoch	Top 1	Top 5	Loss	Time	Code
re_resnet_s	Re-ResNet-S	224	Standard	120	79.34	94.47	1.6733	392.0	211
re_resnet_s_aug	Re-ResNet-S	224	RA + Mixup	300	81.12	95.45	2.3112	420.9	212
re_resnet_s_aug1	Re-ResNet-S	224	RA + ResizeMix	300	81.67	95.73	2.3008	420.9	213

• nls means no label smoothing because it may hurt transfer learning in detection and segmentation
• hanser uses fixed_padding as the default padding strategy after commint 7bc8ae0 (Sep 3, 2021)
• these models use fixed padding and InplaceABN

Models for new Res2Net and PP-ResNet

New Res2Net and PP-ResNet have convs only, and don't have group conv, which is beneficial for deployment.

Several models for ImageNet-1K

0.1.1

Add convert_checkpoint