Exemplo n.º 1
0
def get_model(args):
    # initiate model
    input_channels = int(args.use_multiview) * 128 + int(
        args.use_normal) * 3 + int(
            args.use_color) * 3 + int(not args.no_height)
    model = RefNet(num_class=DC.num_class,
                   num_heading_bin=DC.num_heading_bin,
                   num_size_cluster=DC.num_size_cluster,
                   mean_size_arr=DC.mean_size_arr,
                   input_feature_dim=input_channels,
                   num_proposal=args.num_proposals,
                   use_lang_classifier=(not args.no_lang_cls),
                   use_bidir=args.use_bidir,
                   no_reference=args.no_reference,
                   batch_size=args.batch_size)

    # trainable model
    if args.use_pretrained:
        # load model
        print("loading pretrained VoteNet...")
        pretrained_model = RefNet(num_class=DC.num_class,
                                  num_heading_bin=DC.num_heading_bin,
                                  num_size_cluster=DC.num_size_cluster,
                                  mean_size_arr=DC.mean_size_arr,
                                  num_proposal=args.num_proposals,
                                  input_feature_dim=input_channels,
                                  use_bidir=args.use_bidir,
                                  no_reference=True)

        pretrained_path = os.path.join(CONF.PATH.OUTPUT, args.use_pretrained,
                                       "model_last.pth")
        pretrained_model.load_state_dict(torch.load(pretrained_path),
                                         strict=False)

        # mount
        model.backbone_net = pretrained_model.backbone_net
        model.vgen = pretrained_model.vgen
        model.proposal = pretrained_model.proposal

        if args.no_detection:
            # freeze pointnet++ backbone
            for param in model.backbone_net.parameters():
                param.requires_grad = False

            # freeze voting
            for param in model.vgen.parameters():
                param.requires_grad = False

            # freeze detector
            for param in model.proposal.parameters():
                param.requires_grad = False

    # to CUDA
    model = model.cuda()

    return model
Exemplo n.º 2
0
def get_model(args):
    # initiate model
    input_channels = int(args.use_multiview) * 128 + int(
        args.use_normal) * 3 + int(
            args.use_color) * 3 + int(not args.no_height)
    model_fn = model_fn_decorator()
    model = RefNet(num_class=DC.num_class,
                   num_heading_bin=DC.num_heading_bin,
                   num_size_cluster=DC.num_size_cluster,
                   mean_size_arr=DC.mean_size_arr,
                   input_feature_dim=input_channels,
                   num_proposal=args.num_proposals,
                   use_lang_classifier=(not args.no_lang_cls),
                   use_bidir=args.use_bidir,
                   no_reference=args.no_reference,
                   batch_size=args.batch_size,
                   fix_match_module_input=args.fix_match_module_input,
                   model_fn=model_fn)
    # trainable model
    if args.use_pretrained:
        # load model
        print("loading pretrained PointGroup...")
        pretrained_model = RefNet(num_class=DC.num_class,
                                  num_heading_bin=DC.num_heading_bin,
                                  num_size_cluster=DC.num_size_cluster,
                                  mean_size_arr=DC.mean_size_arr,
                                  num_proposal=args.num_proposals,
                                  input_feature_dim=input_channels,
                                  use_bidir=args.use_bidir,
                                  no_reference=True)
        if args.use_pretrained[-4:] != ".pth":
            pretrained_path = os.path.join(CONF.PATH.OUTPUT,
                                           args.use_pretrained,
                                           "model_last.pth")
        else:
            pretrained_path = os.path.join(CONF.PATH.BASE, args.use_pretrained)
        pretrained_model.load_state_dict(torch.load(pretrained_path),
                                         strict=False)

        # mount
        model.pointgroup = pretrained_model.pointgroup
        print("loaded pretrained PG model: ", pretrained_path)
        if args.no_pg:
            # freeze PG
            for param in model.pointgroup.parameters():
                param.requires_grad = False
            print("freezed pg params")

    # to CUDA
    model = model.cuda()

    return model