MLVLab/Human_Object_Interaction

1# ------------------------------------------------------------------------2# HOTR official code : engine/arg_parser.py3# Copyright (c) Kakao Brain, Inc. and its affiliates. All Rights Reserved4# Modified arguments are represented with *5# ------------------------------------------------------------------------6# Modified from DETR (https://github.com/facebookresearch/detr)7# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved8# ------------------------------------------------------------------------9import argparse10import hotr.util.misc as utils11 12def get_args_parser():13    parser = argparse.ArgumentParser('Set transformer detector', add_help=False)14    parser.add_argument('--lr', default=1e-4, type=float)15    parser.add_argument('--lr_backbone', default=1e-5, type=float)16    parser.add_argument('--batch_size', default=2, type=int)17    parser.add_argument('--weight_decay', default=1e-4, type=float)18    parser.add_argument('--epochs', default=100, type=int)19    parser.add_argument('--lr_drop', default=80, type=int)20    parser.add_argument('--clip_max_norm', default=0.1, type=float,21                        help='gradient clipping max norm')22 23    # DETR Model parameters24    parser.add_argument('--frozen_weights', type=str, default=None,25                        help="Path to the pretrained model. If set, only the mask head will be trained")26    parser.add_argument('--pretrain_interaction_tf', type=str, default=None,27                        help="Path to the pretrained model. If set, only the mask head will be trained")28                        29    # DETR Backbone30    parser.add_argument('--backbone', default='resnet50', type=str,31                        help="Name of the convolutional backbone to use")32    parser.add_argument('--dilation', action='store_true',33                        help="If true, we replace stride with dilation in the last convolutional block (DC5)")34    parser.add_argument('--position_embedding', default='sine', type=str, choices=('sine', 'learned'),35                        help="Type of positional embedding to use on top of the image features")36 37    # DETR Transformer (= Encoder, Instance Decoder)38    parser.add_argument('--enc_layers', default=6, type=int,39                        help="Number of encoding layers in the transformer")40    parser.add_argument('--dec_layers', default=6, type=int,41                        help="Number of decoding layers in the transformer")42    parser.add_argument('--dim_feedforward', default=2048, type=int,43                        help="Intermediate size of the feedforward layers in the transformer blocks")44    parser.add_argument('--hidden_dim', default=256, type=int,45                        help="Size of the embeddings (dimension of the transformer)")46    parser.add_argument('--dropout', default=0.1, type=float,47                        help="Dropout applied in the transformer")48    parser.add_argument('--nheads', default=8, type=int,49                        help="Number of attention heads inside the transformer's attentions")50    parser.add_argument('--num_queries', default=100, type=int,51                        help="Number of query slots")52    parser.add_argument('--pre_norm', action='store_true')53    parser.add_argument('--decoder_form', default=2, type=int,54                        help="1-decoder or 2-decoder")55    # Segmentation56    parser.add_argument('--masks', action='store_true',57                        help="Train segmentation head if the flag is provided")58 59    # Loss Option60    parser.add_argument('--no_aux_loss', dest='aux_loss', action='store_false',61                        help="Disables auxiliary decoding losses (loss at each layer)")62 63    # Loss coefficients (DETR)64    parser.add_argument('--mask_loss_coef', default=1, type=float)65    parser.add_argument('--dice_loss_coef', default=1, type=float)66    parser.add_argument('--bbox_loss_coef', default=5, type=float)67    parser.add_argument('--giou_loss_coef', default=2, type=float)68    parser.add_argument('--eos_coef', default=0.1, type=float,69                        help="Relative classification weight of the no-object class")70 71    # Matcher (DETR)72    parser.add_argument('--set_cost_class', default=1, type=float,73                        help="Class coefficient in the matching cost")74    parser.add_argument('--set_cost_bbox', default=5, type=float,75                        help="L1 box coefficient in the matching cost")76    parser.add_argument('--set_cost_giou', default=2, type=float,77                        help="giou box coefficient in the matching cost")78 79    # * HOI Detection80    parser.add_argument('--HOIDet', action='store_true',81                        help="Train HOI Detection head if the flag is provided")82    parser.add_argument('--share_enc', action='store_true',83                        help="Share the Encoder in DETR for HOI Detection if the flag is provided")84    parser.add_argument('--pretrained_dec', action='store_true',85                        help="Use Pre-trained Decoder in DETR for Interaction Decoder if the flag is provided")                        86    parser.add_argument('--hoi_enc_layers', default=1, type=int,87                        help="Number of decoding layers in HOI transformer")88    parser.add_argument('--hoi_dec_layers', default=1, type=int,89                        help="Number of decoding layers in HOI transformer")90    parser.add_argument('--hoi_nheads', default=8, type=int,91                        help="Number of decoding layers in HOI transformer")92    parser.add_argument('--hoi_dim_feedforward', default=2048, type=int,93                        help="Number of decoding layers in HOI transformer")94    # parser.add_argument('--hoi_mode', type=str, default=None, help='[inst | pair | all]')95    parser.add_argument('--num_hoi_queries', default=100, type=int,96                        help="Number of Queries for Interaction Decoder")97    parser.add_argument('--hoi_aux_loss', action='store_true')98 99 100    # * HOTR Matcher101    parser.add_argument('--set_cost_idx', default=1, type=float,102                        help="IDX coefficient in the matching cost")103    parser.add_argument('--set_cost_act', default=1, type=float,104                        help="Action coefficient in the matching cost")105    parser.add_argument('--set_cost_tgt', default=1, type=float,106                        help="Target coefficient in the matching cost")107 108    # * HOTR Loss coefficients109    parser.add_argument('--temperature', default=0.05, type=float, help="temperature")110    parser.add_argument('--hoi_consistency_loss_coef', default=1, type=float)111    parser.add_argument('--hoi_idx_loss_coef', default=1, type=float)112    parser.add_argument('--hoi_idx_consistency_loss_coef', default=1, type=float)113    parser.add_argument('--hoi_act_loss_coef', default=1, type=float)114    parser.add_argument('--hoi_act_consistency_loss_coef', default=1, type=float)115    parser.add_argument('--hoi_tgt_loss_coef', default=1, type=float)116    parser.add_argument('--hoi_tgt_consistency_loss_coef', default=1, type=float)117    parser.add_argument('--hoi_eos_coef', default=0.1, type=float, help="Relative classification weight of the no-object class")118    119    parser.add_argument('--ramp_down_epoch',default=10000,type=int)120    parser.add_argument('--ramp_up_epoch',default=0,type=int)121    #consistency122    parser.add_argument('--use_consis',action='store_true',help='use consistency regularization')123    parser.add_argument('--share_dec_param',action='store_true',help = 'share decoder parameters of all stages')124    parser.add_argument("--augpath_name", type=utils.arg_as_list,default=[],125                        help='choose which augmented inference paths to use. (p2:x->HO->I,p3:x->HI->O,p4:x->OI->H)')                       126    parser.add_argument('--stop_grad_stage',action='store_true',help='Do not back propogate loss to previous stage')127    parser.add_argument('--path_id', default=0, type=int)128 129    # * dataset parameters130    parser.add_argument('--dataset_file', help='[coco | vcoco]')131    parser.add_argument('--data_path', type=str)132    parser.add_argument('--object_threshold', type=float, default=0, help='Threshold for object confidence')133 134    # machine parameters135    parser.add_argument('--output_dir', default='',136                        help='path where to save, empty for no saving')137    parser.add_argument('--custom_path', default='',138                        help="Data path for custom inference. Only required for custom_main.py")139    parser.add_argument('--device', default='cuda',140                        help='device to use for training / testing')141    parser.add_argument('--seed', default=42, type=int)142    parser.add_argument('--resume', default='', help='resume from checkpoint')143    parser.add_argument('--start_epoch', default=0, type=int, metavar='N',144                        help='start epoch')145    parser.add_argument('--num_workers', default=2, type=int)146 147    # mode148    parser.add_argument('--eval', action='store_true', help="Only evaluate results if the flag is provided")149    parser.add_argument('--validate', action='store_true', help="Validate after every epoch")150 151    # distributed training parameters152    parser.add_argument('--world_size', default=1, type=int,153                        help='number of distributed processes')154    parser.add_argument('--dist_url', default='env://', help='url used to set up distributed training')155 156    # * WanDB157    parser.add_argument('--wandb', action='store_true')158    parser.add_argument('--project_name', default='hotr_cpc')159    parser.add_argument('--group_name', default='mlv')160    parser.add_argument('--run_name', default='run_000001')161    return parser162