MODEL_ZOO.md

May 21, 2020 ยท View on GitHub

Model Zoo

(1) ResNet-50 with different methods

The pretrained weights for ResNet-50 can be found via the dropbox link here. Both the weights of backbone and projection heads are provided.

NameArch# ParametersEpochs
InsDis.pthResNet-5024M200
CMC.pthResNet-5012M200
MoCo.pthResNet-5024M200
PIRL.pthResNet-5024M200
MoCov2.pthResNet-5024M200
InfoMin_200.pthResNet-5024M200
InfoMin_800.pthResNet-5024M800
InfoMin_800_run2.pth*ResNet-5024M800

*: compared with InfoMin_800.pth, model InfoMin_800_run2.pth has slightly stronger transfer performance on COCO, and slightly weaker linear accuracy on ImageNet.

(2) ResNet and ResNeXt with InfoMin

The pretrained weights for InfoMin with other architectures can be found via the dropbox link here. Both the weights of backbone and projection heads are provided.

NameArch(bsz, gpus)Epochs
InfoMin_resnet101_e300.pthResNet-101(256, 8)300
InfoMin_resnet152_e200.pthResNet-152(256, 8)200
InfoMin_resnext101v1_e200.pthResNeXt-101 (32x4d)(512, 32)200
InfoMin_resnext101v2_e200.pth*ResNeXt-101 (32x8d)(512, 32)200
InfoMin_resnext101v3_e200.pthResNeXt-101 (64x4d)(512, 32)200
InfoMin_resnext152v1_e200.pthResNeXt-152 (32x4d)(512, 32)200
InfoMin_resnext152v2_e200.pth*ResNeXt-152 (32x8d)(512, 32)200

*: used to compare with ImageNet pretrained models on COCO.

(3) ResNeSt with InfoMin (TODO if resources suffice)