Function validation_sam

func_3d/function.py:198–314

(args, val_loader, epoch, net: nn.Module, clean_dir=True)

    return epoch_loss / len(train_loader), epoch_prompt_loss / len(train_loader), epoch_non_prompt_loss / len(train_loader)

def validation_sam(args, val_loader, epoch, net: nn.Module, clean_dir=True):
     # eval mode
    net.eval()

    n_val = len(val_loader)  # the number of batch
    mix_res = (0,)*1*2
    tot = 0
    threshold = (0.1, 0.3, 0.5, 0.7, 0.9)
    prompt_freq = args.prompt_freq

    lossfunc = criterion_G
    # lossfunc = paper_loss

    prompt = args.prompt

    with tqdm(total=n_val, desc='Validation round', unit='batch', leave=False) as pbar:
        for pack in val_loader:
            imgs_tensor = pack['image']
            mask_dict = pack['label']
            if prompt == 'click':
                pt_dict = pack['pt']
                point_labels_dict = pack['p_label']
            elif prompt == 'bbox':
                bbox_dict = pack['bbox']
            if len(imgs_tensor.size()) == 5:
                imgs_tensor = imgs_tensor.squeeze(0)
            frame_id = list(range(imgs_tensor.size(0)))
            
            train_state = net.val_init_state(imgs_tensor=imgs_tensor)
            prompt_frame_id = list(range(0, len(frame_id), prompt_freq))
            obj_list = []
            for id in frame_id:
                obj_list += list(mask_dict[id].keys())
            obj_list = list(set(obj_list))
            if len(obj_list) == 0:
                continue

            name = pack['image_meta_dict']['filename_or_obj']

            with torch.no_grad():
                for id in prompt_frame_id:
                    for ann_obj_id in obj_list:
                        try:
                            if prompt == 'click':
                                points = pt_dict[id][ann_obj_id].to(device=GPUdevice)
                                labels = point_labels_dict[id][ann_obj_id].to(device=GPUdevice)
                                _, _, _ = net.train_add_new_points(
                                    inference_state=train_state,
                                    frame_idx=id,
                                    obj_id=ann_obj_id,
                                    points=points,
                                    labels=labels,
                                    clear_old_points=False,
                                )
                            elif prompt == 'bbox':
                                bbox = bbox_dict[id][ann_obj_id]
                                _, _, _ = net.train_add_new_bbox(
                                    inference_state=train_state,