PySlowFast Model Zoo and Baselines

August 25, 2022 ยท View on GitHub

Kinetics 400 and 600

architecturesizecrops x clipsframe length x sample ratetop1top5modelconfigdataset
C2DR503 x 108 x 867.287.8linkKinetics/c2/C2D_NOPOOL_8x8_R50K400
I3DR503 x 108 x 873.590.8linkKinetics/c2/I3D_8x8_R50K400
I3D NLNR503 x 108 x 874.091.1linkKinetics/c2/I3D_NLN_8x8_R50K400
SlowR503 x 104 x 1672.790.3linkKinetics/c2/SLOW_4x16_R50K400
SlowR503 x 108 x 874.891.6linkKinetics/c2/SLOW_8x8_R50K400
SlowFastR503 x 104 x 1675.692.0linkKinetics/c2/SLOWFAST_4x16_R50K400
SlowFastR503 x 108 x 877.092.6linkKinetics/c2/SLOWFAST_8x8_R50K400
MViTv1B-Conv1 x 516 x 478.493.5linkKinetics/MVIT_B_16x4_CONVK400
rev-MViTB-Conv1 x 516 x 478.493.4linkKinetics/REV_MVIT_B_16x4_CONVK400
MViTv1B-Conv1 x 532 x 380.494.8linkKinetics/MVIT_B_32x3_CONVK400
MViTv1B-Conv1 x 532 x 383.996.5linkKinetics/MVIT_B_32x3_CONV_K600K600
MViTv2S1 x 516 x 481.094.6linkKinetics/MVITv2_S_16x4K400
MViTv2B1 x 532 x 382.995.7linkKinetics/MVITv2_B_32x3K400

X3D models (details in projects/x3d)

architecturesizepretrainframe length x sample ratetop1 10-viewtop1 30-viewparameters (M)FLOPs (G)modelconfig
X3DXS-4 x 1268.769.53.80.60linkKinetics/X3D_XS
X3DS-13 x 673.173.53.81.96linkKinetics/X3D_S
X3DM-16 x 575.176.23.84.73linkKinetics/X3D_M
X3DL-16 x 576.977.56.218.37linkKinetics/X3D_L

AVA

architecturesizePretrain Modelframe length x sample rateMAPAVA versionmodel
SlowR50Kinetics 4004 x 1619.52.2link
SlowFastR101Kinetics 6008 x 828.22.1link
SlowFastR101Kinetics 6008 x 829.12.2link
SlowFastR101Kinetics 60016 x 829.42.2link

Multigrid Training

Update June, 2020: In the following we provide (reimplemented) models from "A Multigrid Method for Efficiently Training Video Models " paper. The multigrid method trains about 3-6x faster than the original training on multiple datasets. See projects/multigrid for more information. The following provides models, results, and example config files.

Kinetics:

architecturesizepretrainframe length x sample ratetrainingtop1top5modelconfig
SlowFastR50-8 x 8Standard76.892.7linkKinetics/SLOWFAST_8x8_R50_stepwise
SlowFastR50-8 x 8Multigrid76.692.7linkKinetics/SLOWFAST_8x8_R50_stepwise_multigrid

(Here we use stepwise learning rate schedule.)

Something-Something V2:

architecturesizepretrainframe length x sample ratetrainingtop1top5modelconfig
SlowFastR50Kinetics 40016 x 8Standard63.088.5linkSSv2/SLOWFAST_16x8_R50
SlowFastR50Kinetics 40016 x 8Multigrid63.588.7linkSSv2/SLOWFAST_16x8_R50_multigrid

Charades

architecturesizepretrainframe length x sample ratetrainingmAPmodelconfig
SlowFastR50Kinetics 40016 x 8Standard38.9linkSSv2/SLOWFAST_16x8_R50
SlowFastR50Kinetics 40016 x 8Multigrid38.6linkSSv2/SLOWFAST_16x8_R50_multigrid

ImageNet

We also release the imagenet pretrained model if finetuning from ImageNet is preferred. The reported accuracy is obtained by center crop testing on the validation set.

architecturesizeTop1Top5modelConfig
ResNetR5076.493.2linkImageNet/RES_R50
MVITB-16-Conv82.996.3linkImageNet/MVIT_B_16_CONV
rev-VITSmall79.994.9linkImageNet/REV_VIT_S.yaml
rev-VITBase81.895.6linkImageNet/REV_VIT_B.yaml
rev-MVITBase82.9*96.3linkImageNet/REV_MVIT_B_16_CONV.yaml

*please refer to Reversible Model Zoo.

PyTorchVideo

We support and benchmark PyTorchVideo models and datasets in PySlowFast. See projects/pytorchvideo for more information about PyTorchVideo Model Zoo.