Demo Issue

[liang315]

I encountered this problem when running the following script

python demo.py --config-file .\configs\CoDet_OVCOCO_R50_1x.yaml --input ..\VLDet\heatmap_inputs\000000059598.jpg --output .\demo_outputs\ --vocabulary co
co --opts MODEL.WEIGHTS .\models\CoDet_OVCOCO_R50_1x..pth

Traceback (most recent call last):
  File "demo.py", line 140, in <module>
    demo = VisualizationDemo(cfg, args)
  File "E:\VLDet\CoDet\codet\predictor.py", line 68, in __init__
    reset_cls_infer(self.predictor.model, classifier, num_classes)
  File "E:\VLDet\CoDet\codet\modeling\utils.py", line 63, in reset_cls_infer
    if model.roi_heads.box_predictor[0].cls_score.norm_weight:
TypeError: 'CoDetFastRCNNOutputLayers' object is not subscriptable

Afterwards, I modified the code in the reset_cls_infer function and it can be run

def reset_cls_infer(model, cls_path, num_classes):
    model.roi_heads.num_classes = num_classes
    if type(cls_path) == str:
        print('Resetting zs_weight', cls_path)
        zs_weight = torch.tensor(
            np.load(cls_path),
            dtype=torch.float32).permute(1, 0).contiguous() # D x C
    else:
        zs_weight = cls_path
    zs_weight = torch.cat(
        [zs_weight, zs_weight.new_zeros((zs_weight.shape[0], 1))],
        dim=1) # D x (C + 1)
    if model.roi_heads.box_predictor.cls_score.norm_weight:
    # if model.roi_heads.box_predictor[0].cls_score.norm_weight:
        zs_weight = F.normalize(zs_weight, p=2, dim=0)
    zs_weight = zs_weight.to(model.device)
    # for k in range(len(model.roi_heads.box_predictor)):
    #     del model.roi_heads.box_predictor[k].cls_score.detection_weight
    #     model.roi_heads.box_predictor[k].cls_score.detection_weight = zs_weight
    del model.roi_heads.box_predictor.cls_score.detection_weight
    model.roi_heads.box_predictor.cls_score.detection_weight = zs_weight

I want to know if what I'm doing is correct

[machuofan]

Yes, it looks good to me.

[sssssshf]

I encountered this problem when running the following script

python demo.py --config-file .\configs\CoDet_OVCOCO_R50_1x.yaml --input ..\VLDet\heatmap_inputs\000000059598.jpg --output .\demo_outputs\ --vocabulary co
co --opts MODEL.WEIGHTS .\models\CoDet_OVCOCO_R50_1x..pth

Traceback (most recent call last):
  File "demo.py", line 140, in <module>
    demo = VisualizationDemo(cfg, args)
  File "E:\VLDet\CoDet\codet\predictor.py", line 68, in __init__
    reset_cls_infer(self.predictor.model, classifier, num_classes)
  File "E:\VLDet\CoDet\codet\modeling\utils.py", line 63, in reset_cls_infer
    if model.roi_heads.box_predictor[0].cls_score.norm_weight:
TypeError: 'CoDetFastRCNNOutputLayers' object is not subscriptable

Afterwards, I modified the code in the reset_cls_infer function and it can be run

def reset_cls_infer(model, cls_path, num_classes):
    model.roi_heads.num_classes = num_classes
    if type(cls_path) == str:
        print('Resetting zs_weight', cls_path)
        zs_weight = torch.tensor(
            np.load(cls_path),
            dtype=torch.float32).permute(1, 0).contiguous() # D x C
    else:
        zs_weight = cls_path
    zs_weight = torch.cat(
        [zs_weight, zs_weight.new_zeros((zs_weight.shape[0], 1))],
        dim=1) # D x (C + 1)
    if model.roi_heads.box_predictor.cls_score.norm_weight:
    # if model.roi_heads.box_predictor[0].cls_score.norm_weight:
        zs_weight = F.normalize(zs_weight, p=2, dim=0)
    zs_weight = zs_weight.to(model.device)
    # for k in range(len(model.roi_heads.box_predictor)):
    #     del model.roi_heads.box_predictor[k].cls_score.detection_weight
    #     model.roi_heads.box_predictor[k].cls_score.detection_weight = zs_weight
    del model.roi_heads.box_predictor.cls_score.detection_weight
    model.roi_heads.box_predictor.cls_score.detection_weight = zs_weight

I want to know if what I'm doing is correct

我也是这么改的 应该没什么问题 ，请问你有没有尝试训练起来。