[BUG] misunderstanding in create_model_from_config for arg pretrained

[yohann84L]

This may not be a bug/error but in create_model_from_config you have the following:

pretrained_backbone = kwargs.pop('pretrained_backbone', True)
if pretrained or checkpoint_path:
    pretrained_backbone = False  # no point in loading backbone weights

If pretrained is True in create_model_from_config , this lead to not loading pretrained weight for the backbone in the EfficientDet class:

class EfficientDet(nn.Module):
    def __init__(self, config, pretrained_backbone=True, alternate_init=False):
        super(EfficientDet, self).__init__()
        self.config = config
        set_config_readonly(self.config)
        self.backbone = create_model(
            config.backbone_name, features_only=True, out_indices=(2, 3, 4),
            pretrained=pretrained_backbone, **config.backbone_args)

Is there an explanation or it should be changed to the following ?

pretrained_backbone = kwargs.pop('pretrained_backbone', True)
if not pretrained or checkpoint_path:
    pretrained_backbone = False  # no point in loading backbone weights

[yohann84L]

Okay sorry, forgot what I said, pretrained is actually for the whole EfficientDet model and not the backbone.