ConvNeXt Weights from PyTorch

ConvNeXt Weights from PyTorch

ViT Weights from PyTorch

Fork weigths of ViT from torchvison.

ShuffleNets

Model	Top-1@Paper	Top-1	Top-5
ShuffleNetV2 x2.0	74.9	74.368	92.050
ShuffleNetV2 x2.0 + SDKs	-	73.870	91.674

EfficientNets

Model	Top-1@Paper	Top-1	Top-5
EfficientNet-B0	77.3	76.256	93.174
EfficientNet-B0 + SDKs	-	75.764	92.830

ResNets Weights

Model	Top-1@Paper	Top-1	Top-5
ResNet18 V1	69.7	71.060	89.922

Training

Procedure
learning rate	0.2
batch size	512
weight decay	1e-4
lr scheduler	cosine
min lr	1e-6
optimizer	SGD with momentum 0.9
epochs	100
warmup epochs	5
lable smoothing	0.1
mixup	-
cutmix	-
amp	True
image processing	RandomResizedCrop/RandomHorizontalFlip
data augment	-
dropout rate	0.0
drop path rate	-
crop size	224
val-resize size	256
val-resize size	224

VGNets Weights

Model	Params(M)	FLOPS(M)	Top-1	Top-5
VGNetG-1.0MP	0.999	144	68.128	88.312
VGNetG-1.0MP + SE	1.143	145	70.122	89.524
VGNetG-1.5MP	1.502	191	70.494	89.684
VGNetG-1.5MP + SE	1.702	192	72.422	90.664
VGNetG-2.0MP	2.006	304	72.314	90.730
VGNetG-2.0MP + SE	2.345	306	74.324	91.788
VGNetG-2.5MP	2.493	399	73.740	91.516
VGNetG-2.5MP + SE	2.940	401	75.590	92.568

Training

Procedure
learning rate	0.2
batch size	512
weight decay	1e-4
no bias & bn decay	True
lr scheduler	cosine
min lr	1e-6
optimizer	SGD with momentum 0.9
warmup epochs	5
lable smoothing	0.1
mixup	-
cutmix	-
amp	True
image processing	RandomResizedCrop/RandomHorizontalFlip
data augment	-
dropout rate	0.0
drop path rate	-
crop size	192
val-resize size	232
val-resize size	224

RegNets Weights

Model	Top-1@Paper	Top-1	Top-5
RegNetX 400MF	72.7	73.156	91.320

Training

V1 & V2

Procedure
learning rate	0.2
batch size	512
weight decay	1e-4
lr scheduler	cosine
min lr	1e-6
optimizer	SGD with momentum 0.9
epochs	100
warmup epochs	5
lable smoothing	0.1
mixup	-
cutmix	-
amp	True
image processing	RandomResizedCrop/RandomHorizontalFlip
data augment	-
dropout rate	0.0
drop path rate	-
crop size	224
val-resize size	256
val-resize size	224

MobileNets

Accuracy

Model	Parms(M)	FLOPS(M)	Top-1@Paper	Top-1	Top-5
MobileNet V1 x1.0	4.232	579	70.6	72.966(73.264@224/232)	91.190
MobileNet V1 x0.75	2.586	333	68.4	69.894	89.520
MobileNet V1 x0.5	1.332	155	63.7	65.590	86.336
MobileNet V1 x0.35	0.766	79	-	59.328	82.068
MobileNet V2 x1.0	3.505	314	72.0	72.154(224/232)	90.736
MobileNet V2 x0.75	2.636	221	-	68.738	88.330
MobileNet V2 x0.5	1.969	104	-	63.926	85.154
MobileNet V2 x0.35	1.677	65.074	-	59.240	81.782

Training

Procedure
learning rate	0.2
batch size	512
lr scheduler	cosine
min lr	1e-6
optimizer	SGD with momentum 0.9
weight decay	1e-4
epochs	100
warmup epochs	5
lable smoothing	0.1
mixup	-
cutmix	-
amp	True
image processing	RandomResizedCrop/RandomHorizontalFlip
data augment	-
dropout rate	0.0
drop path rate	-
crop size	224
val-resize size	256
val-resize size	224

MobileNet V1 x1.0

Procedure
crop size	192
val-resize size	232
val-resize size	192

V2

Procedure
weight decay	1e-5
data augment	normal