(self)
| 72 | |
| 73 | |
| 74 | def validate(self): |
| 75 | args, config = self.args, self.config |
| 76 | |
| 77 | rank = torch.distributed.get_rank() |
| 78 | model, model_cond, __, __, __, __, __, __, __, __, __ = get_model(config, is_train=False, resume = True, resume_path = config.training.scorenet.test_path) |
| 79 | model = nn.parallel.DistributedDataParallel(model, device_ids=[args.device], output_device=args.local_rank) |
| 80 | model_cond = nn.parallel.DistributedDataParallel(model_cond, device_ids=[args.device], output_device=args.local_rank) |
| 81 | model_score, model_score_cond, __, __, __, __, epoch, step, __ = get_model_score(config, is_train=False, resume = True, resume_path = self.config.sampling.ckpt) |
| 82 | model_score = nn.parallel.DistributedDataParallel(model_score, device_ids=[args.device], output_device=args.local_rank) |
| 83 | model_score_cond = nn.parallel.DistributedDataParallel(model_score_cond, device_ids=[args.device], output_device=args.local_rank) |
| 84 | state = dict(model = model, model_cond = model_cond, model_score = model_score, epoch = epoch, model_score_cond = model_score_cond) |
| 85 | |
| 86 | KST = datetime.timezone(datetime.timedelta(hours=8)) |
| 87 | if config.inference.input_type == 'video': |
| 88 | dname = config.inference.input_name |
| 89 | else: |
| 90 | dname = 'image' |
| 91 | config.inference.out_dir = os.path.join(config.inference.out_dir,str(datetime.datetime.now(tz=KST))[5:-16]+dname) |
| 92 | |
| 93 | list_name = [config.inference.input_type,'mesh'] |
| 94 | for name in list_name: |
| 95 | path = os.path.join(config.inference.out_dir,name) |
| 96 | if os.path.exists(path) and rank == 0: |
| 97 | shutil.rmtree(path) |
| 98 | os.makedirs(path,exist_ok=True) |
| 99 | |
| 100 | img_path_list, img_dir, fps = get_image_path(config) |
| 101 | |
| 102 | |
| 103 | virtualpose_name = 'VirtualPose' |
| 104 | det_update_config(f'{virtualpose_name}/configs/images/images_inference.yaml') |
| 105 | |
| 106 | cur_path = config.inference.det_dir |
| 107 | img_dir = img_dir |
| 108 | |
| 109 | det_model = eval('det_models.multi_person_posenet.get_multi_person_pose_net')(det_cfg, is_train=False) |
| 110 | with torch.no_grad(): |
| 111 | det_model = torch.nn.DataParallel(det_model,device_ids=[rank]) |
| 112 | |
| 113 | pretrained_file = osp.join(cur_path, f'{virtualpose_name}', det_cfg.NETWORK.PRETRAINED) |
| 114 | state_dict = torch.load(pretrained_file) |
| 115 | new_state_dict = {k:v for k, v in state_dict.items() if 'backbone.pose_branch.' not in k} |
| 116 | det_model.module.load_state_dict(new_state_dict, strict = False) |
| 117 | pretrained_file = osp.join(cur_path, f'{virtualpose_name}', det_cfg.NETWORK.PRETRAINED_BACKBONE) |
| 118 | det_model = load_backbone_validate(det_model, pretrained_file) |
| 119 | |
| 120 | # prepare detection dataset |
| 121 | infer_dataset = det_dataset.images( |
| 122 | det_cfg, img_dir, focal_length=1700, |
| 123 | transform=transforms.Compose([ |
| 124 | transforms.ToTensor(), |
| 125 | transforms.Normalize( |
| 126 | mean=[0.485, 0.456, 0.406], |
| 127 | std=[0.229, 0.224, 0.225]), |
| 128 | ])) |
| 129 | infer_loader = torch.utils.data.DataLoader( |
| 130 | infer_dataset, |
| 131 | batch_size=config.inference.det_bs, |
nothing calls this directly
no test coverage detected