bbox prediction

[msdkarimi]

I've tried to use model in evaluation to understand the pipeline.
Sem. Pan. segmentation works for me, but xbox prediction is all zero coordinates(e.g: [0, 0, 0, 0]). Besides that, I think this problem affects the instance predictions and return all possible instances (e.g: 100).

I set MODEL.DECODER.DETECTION True, to set 'bbox' in task_switch dictionary in "modeling/architectures/seem_model_v1.py", but it gives me KeyError in

    if deep_supervision:
        dec_layers = dec_cfg['DEC_LAYERS']
        aux_weight_dict = {}
        for i in range(dec_layers - 1):
            for k, v in weight_dict.items():
                if (i+1) > (top_x_layers[k.split('_')[1]] - 1): <------------------------  KeyError 
                    continue
                aux_weight_dict.update({k.replace('_0', f"_{i+1}"): v})
        weight_dict.update(aux_weight_dict)

To consider bbox, I've done it in wrong way??

Many thanks in advance