o ϯi @sddlmZmZddlmZddlmZmZddlm Z  dZ e de de de d d Z d d d efde de de dededejf ddZ dde de dedejfddZ dde de dedejfddZ dde de dedejfddZ dde de dedejfddZdS))AnyCallableN) create_resnetcreate_resnet_with_roi_head)load_state_dict_from_urlz5https://dl.fbaipublicfiles.com/pytorchvideo/model_zooz/kinetics/SLOW_8x8_R50.pythz!/ava/SLOW_4x16_R50_DETECTION.pythz/kinetics/C2D_8x8_R50.pythz/kinetics/I3D_8x8_R50.pyth)slow_r50slow_r50_detectionc2d_r50i3d_r50FT pretrainedprogresscheckpoint_path model_builderkwargsreturncKs6|di|}|rt||dd}|d}|||S)Ncpu)r map_location model_state)rload_state_dict)r r rrrmodel checkpoint state_dictrrR/home/ubuntu/.local/lib/python3.10/site-packages/pytorchvideo/models/hub/resnet.py_resnets rc Ks td||tddddd|S)a4 Slow R50 model architecture [1] with pretrained weights based on 8x8 setting on the Kinetics dataset. Model with pretrained weights has top1 accuracy of 74.58. [1] "SlowFast Networks for Video Recognition" Christoph Feichtenhofer et al https://arxiv.org/pdf/1812.03982.pdf Args: pretrained (bool): If True, returns a model pre-trained on the Kinetics dataset progress (bool): If True, displays a progress bar of the download to stderr kwargs: use these to modify any of the other model settings. All the options are defined in pytorchvideo/models/resnet.py NOTE: to use the pretrained model, do not modify the model configuration via the kwargs. Only modify settings via kwargs to initialize a new model without pretrained weights. rr)rr2)r r rstem_conv_kernel_sizehead_pool_kernel_size model_depthNr)rcheckpoint_pathsr r rrrrr)srcKstd||tdtd|S)aW Slow R50 model architecture [1] with pretrained weights based on 4x16 setting. The model is initially trained on Kinetics dataset for classification and later finetuned on AVA dataset for detection. [1] Christoph Feichtenhofer et al, "SlowFast Networks for Video Recognition" https://arxiv.org/pdf/1812.03982.pdf r)r r rrNr)rr$rr%rrrrIs rc Ks"td||tddtjdd|S)a C2D R50 model architecture with pretrained weights based on 8x8 setting on the Kinetics dataset. Model with pretrained weights has top1 accuracy of 71.46. Args: pretrained (bool): If True, returns a model pre-trained on the Kinetics dataset progress (bool): If True, displays a progress bar of the download to stderr kwargs: use these to modify any of the other model settings. All the options are defined in pytorchvideo/models/resnet.py NOTE: to use the pretrained model, do not modify the model configuration via the kwargs. Only modify settings via kwargs to initialize a new model without pretrained weights. r r)rrrr&r&r&r r rr! stage1_poolstage_conv_a_kernel_sizeNrrr$nn MaxPool3dr%rrrr ]s r c Ks6td||tddtjdddgddgddgfd|S)aR I3D R50 model architecture from [1] with pretrained weights based on 8x8 setting on the Kinetics dataset. Model with pretrained weights has top1 accuracy of 73.27. [1] "Quo Vadis, Action Recognition? A New Model and the Kinetics Dataset" Joao Carreira, Andrew Zisserman https://arxiv.org/abs/1705.07750 Args: pretrained (bool): If True, returns a model pre-trained on the Kinetics dataset progress (bool): If True, displays a progress bar of the download to stderr kwargs: use these to modify any of the other model settings. All the options are defined in pytorchvideo/models/resnet.py NOTE: to use the pretrained model, do not modify the model configuration via the kwargs. Only modify settings via kwargs to initialize a new model without pretrained weights. r )rr)rrr&r'Nrr*r%rrrr ~s r )FT)typingrrtorch.nnr+pytorchvideo.models.resnetrr torch.hubrroot_dirr$boolstrModulerrrr r rrrrs     !  "