the configure item "SCORE_THRESH_TEST " doesn't work for a configure file generated by lazyconfig_train_net.py

[proshanm]

I finetuned a MAE pretrained model for object detection using ViTDet project using our own dataset by involing lazyconfig_train_net.py. I wanted to load the finetuned module and visualize the result, but I found a problem that the structure of configure file 'config.yaml' generated by lazyconfig_train_net.py is different from that of configure file of build-in module. So I implemented user-defined Predictor which load module and weight from config.yaml instead of DefaultPredictor and visualize the output using Visualizer. The result could be visualized, but there were many low-score boxes that I didn't want to show. so I add hierarchical property MODEL.ROI_HEADS.SCORE_THRESH_TEST into config.yaml, but it doesn't work. I know the problem can be solved by modifying Visualizer to introduce a score thresh, but i'm not sure if it is necessory for this case or there's another better way for solving the problem. The code of predictor is as follow.
`class ProjectPredictor:

def __init__(self, cfg):
    self.model = instantiate(cfg.model)
    self.model.eval()
    if len(cfg.dataloader.test):
        self.metadata = MetadataCatalog.get(cfg.dataloader.test.dataset['names'])

    checkpointer = DetectionCheckpointer(self.model)
    checkpointer.load(cfg.train.init_checkpoint)

    self.aug = T.ResizeShortestEdge(
        [cfg.dataloader.test.mapper.augmentations[0]['short_edge_length'], cfg.dataloader.test.mapper.augmentations[0]['short_edge_length']], cfg.dataloader.test.mapper.augmentations[0]['max_size']
    )

    self.input_format = cfg.dataloader.test.mapper.image_format
    assert self.input_format in ["RGB", "BGR", 'L'], self.input_format

def __call__(self, original_image):
    """
    Args:
        original_image (np.ndarray): an image of shape (H, W, C) (in BGR order).

    Returns:
        predictions (dict):
            the output of the model for one image only.
            See :doc:`/tutorials/models` for details about the format.
    """
    with torch.no_grad():  # https://github.com/sphinx-doc/sphinx/issues/4258
        # Apply pre-processing to image.
        if self.input_format == "RGB":
            # whether the model expects BGR inputs or RGB
            original_image = original_image[:, :, ::-1]
        height, width = original_image.shape[:2]
        image = self.aug.get_transform(original_image).apply_image(original_image)
        # 
        if self.input_format == "L":
            image = image.astype("float32")
            image = [image]
            image = torch.as_tensor(image)
        else:
            image = torch.as_tensor(image.astype("float32").transpose(2, 0, 1))
        inputs = {"image": image, "height": height, "width": width}
        predictions = self.model([inputs])[0]
        return predictions`

[bmaxdk]

You can filtering out low-scoring bounding boxes with modify the predictor.
You can directly modify the prediction outputs in your custom predictor after calling self.model([inputs])[0]. The returned predictions dict usually contains a key instances, which itself contains various fields like scores. You can filter out boxes with low scores before returning them.

threshold = 0.5  # set your own threshold
instances = predictions["instances"]
keep = instances.scores > threshold  # assuming the field is named 'scores'
instances = instances[keep]
predictions["instances"] = instances