MODEL_ZOO.md

FAIR Self-Supervision Benchmark Model Zoo

We provide all the models used in our paper. Further, we also provide the input files to vary the problem complexity of Jigsaw and Colorization approaches by varying the number of permutations and number of color bins respectively.

Models

AlexNet

• ImageNet-1K supervised: https://dl.fbaipublicfiles.com/fair_self_supervision_benchmark/models/caffenet_bvlc_in1k_supervised.npy
• Places-205 supervised: https://dl.fbaipublicfiles.com/fair_self_supervision_benchmark/models/caffenet_bvlc_places205_supervised.pkl
• Jigsaw ImageNet-1K self-supervised: https://dl.fbaipublicfiles.com/fair_self_supervision_benchmark/models/alexnet_jigsaw_in1k_pretext.pkl
• Jigsaw ImageNet-22K self-supervised: https://dl.fbaipublicfiles.com/fair_self_supervision_benchmark/models/alexnet_jigsaw_in22k_pretext.pkl
• Jigsaw YFCC100M self-supervised: https://dl.fbaipublicfiles.com/fair_self_supervision_benchmark/models/alexnet_jigsaw_yfcc100m_pretext.pkl
• Colorization ImageNet-1K self-supervised: https://dl.fbaipublicfiles.com/fair_self_supervision_benchmark/models/alexnet_colorization_in1k_pretext.pkl
• Colorization ImageNet-22K self-supervised: https://dl.fbaipublicfiles.com/fair_self_supervision_benchmark/models/alexnet_colorization_in22k_pretext.pkl
• Colorization YFCC100M self-supervised: https://dl.fbaipublicfiles.com/fair_self_supervision_benchmark/models/alexnet_colorization_yfcc100m_pretext.pkl

ResNet-50

• ImageNet-1K supervised: https://dl.fbaipublicfiles.com/fair_self_supervision_benchmark/models/resnet50_in1k_supervised.pkl
• Places-205 supervised: https://dl.fbaipublicfiles.com/fair_self_supervision_benchmark/models/resnet50_places205_supervised.pkl
• Jigsaw ImageNet-1K self-supervised: https://dl.fbaipublicfiles.com/fair_self_supervision_benchmark/models/resnet50_jigsaw_in1k_pretext.pkl
• Jigsaw ImageNet-22K self-supervised: https://dl.fbaipublicfiles.com/fair_self_supervision_benchmark/models/resnet50_jigsaw_in22k_pretext.pkl
• Jigsaw YFCC100M self-supervised: https://dl.fbaipublicfiles.com/fair_self_supervision_benchmark/models/resnet50_jigsaw_yfcc100m_pretext.pkl
• Jigsaw ImageNet-22K self-supervised for object detection: https://dl.fbaipublicfiles.com/fair_self_supervision_benchmark/models/resnet50_jigsaw_detection_in22k_pretext.pkl
• Colorization ImageNet-1K self-supervised: https://dl.fbaipublicfiles.com/fair_self_supervision_benchmark/models/resnet50_colorization_in1k_pretext.pkl
• Colorization ImageNet-22K self-supervised: https://dl.fbaipublicfiles.com/fair_self_supervision_benchmark/models/resnet50_colorization_in22k_pretext.pkl
• Colorization ImageNet-22K self-supervised for Low-shot training: https://dl.fbaipublicfiles.com/fair_self_supervision_benchmark/models/resnet50_colorization_lowShot_in22k_pretext.pkl
• Colorization YFCC100M self-supervised: https://dl.fbaipublicfiles.com/fair_self_supervision_benchmark/models/resnet50_colorization_yfcc100m_pretext.pkl

Object Detection Models and Proposals

• ImageNet-1K supervised: https://dl.fbaipublicfiles.com/fair_self_supervision_benchmark/models/detection/resnet50_in1k_supervised.pkl

• Places-205 supervised: https://dl.fbaipublicfiles.com/fair_self_supervision_benchmark/models/detection/resnet50_places205_supervised.pkl

• Jigsaw ImageNet-1K self-supervised: https://dl.fbaipublicfiles.com/fair_self_supervision_benchmark/models/detection/resnet50_jigsaw_in1k_pretext.pkl

• Jigsaw ImageNet-22K self-supervised: https://dl.fbaipublicfiles.com/fair_self_supervision_benchmark/models/detection/resnet50_jigsaw_in22k_pretext.pkl

• Jigsaw YFCC100M self-supervised: https://dl.fbaipublicfiles.com/fair_self_supervision_benchmark/models/detection/resnet50_jigsaw_yfcc100m_pretext.pkl

• Selective Search Proposals VOC2007 trainval: https://dl.fbaipublicfiles.com/detectron/selective_search_proposals/selective_search_msra_voc_2007_trainval.pkl

• Selective Search Proposals VOC2007 test: https://dl.fbaipublicfiles.com/detectron/selective_search_proposals/selective_search_msra_voc_2007_test.pkl

• Selective Search Proposals VOC2012 trainval: https://dl.fbaipublicfiles.com/detectron/selective_search_proposals/selective_search_msra_voc_2012_trainval.pkl

PyTorch Models for Visual Navigation

• ImageNet-1K supervised: https://dl.fbaipublicfiles.com/fair_self_supervision_benchmark/models/resnet50_in1k_supervised_gibson.dat
• Jigsaw ImageNet-22K self-supervised: https://dl.fbaipublicfiles.com/fair_self_supervision_benchmark/models/resnet50_jigsaw_in22k_pretext_gibson.dat
• Jigsaw YFCC100M self-supervised: https://dl.fbaipublicfiles.com/fair_self_supervision_benchmark/models/resnet50_in1k_supervised_gibson.dat

Surface Normal Estimation Data

• Ground Truth surface normals and their metadata: https://dl.fbaipublicfiles.com/fair_self_supervision_benchmark/nyuv2_surfacenormal_metadata.zip

Jigsaw Permutations

All permutations below are for 9 patches:

• 100 permutations: https://dl.fbaipublicfiles.com/fair_self_supervision_benchmark/jigsaw_permutations/naroozi_perms_100_patches_9_max.npy
• 701 permutations: https://dl.fbaipublicfiles.com/fair_self_supervision_benchmark/jigsaw_permutations/naroozi_perms_701_patches_9_max.npy
• 2000 permutations: https://dl.fbaipublicfiles.com/fair_self_supervision_benchmark/jigsaw_permutations/hamming_perms_2000_patches_9_max_avg.npy
• 5000 permutations: https://dl.fbaipublicfiles.com/fair_self_supervision_benchmark/jigsaw_permutations/hamming_perms_5000_patches_9_max_avg.npy
• 10000 permutations: https://dl.fbaipublicfiles.com/fair_self_supervision_benchmark/jigsaw_permutations/hamming_perms_10000_patches_9_max_avg.npy

Colorization color bins and priors

Number of color bins vary based on the bin size used. We generate the color bins and priors for various bin sizes:

• Bin size 10 colors: https://dl.fbaipublicfiles.com/fair_self_supervision_benchmark/colorization_bins_priors/priors_train_imagenet1k_bin10.npy

• Bin size 10 color priors: https://dl.fbaipublicfiles.com/fair_self_supervision_benchmark/colorization_bins_priors/pts_in_hull_train_imagenet1k_bin10.npy

• Bin size 15 colors: https://dl.fbaipublicfiles.com/fair_self_supervision_benchmark/colorization_bins_priors/priors_train_imagenet1k_bin15.npy

• Bin size 15 colors priors: https://dl.fbaipublicfiles.com/fair_self_supervision_benchmark/colorization_bins_priors/pts_in_hull_train_imagenet1k_bin15.npy

• Bin size 20 colors: https://dl.fbaipublicfiles.com/fair_self_supervision_benchmark/colorization_bins_priors/priors_train_imagenet1k_bin20.npy

• Bin size 20 colors priors: https://dl.fbaipublicfiles.com/fair_self_supervision_benchmark/colorization_bins_priors/pts_in_hull_train_imagenet1k_bin20.npy

• Bin size 5 colors: https://dl.fbaipublicfiles.com/fair_self_supervision_benchmark/colorization_bins_priors/priors_train_imagenet1k_bin5.npy

• Bin size 5 colors priors: https://dl.fbaipublicfiles.com/fair_self_supervision_benchmark/colorization_bins_priors/pts_in_hull_train_imagenet1k_bin5.npy

• Original work Bin size 10 colors: https://dl.fbaipublicfiles.com/fair_self_supervision_benchmark/colorization_bins_priors/zhang_prior_probs.npy

• Original work Bin size 10 colors priors: https://dl.fbaipublicfiles.com/fair_self_supervision_benchmark/colorization_bins_priors/zhang_pts_in_hull.npy