Hierarchical parameter of custom models' forward method

[jan-fathomx]

Hi! I have a question about the custom forward method for models, which includes the hierarchical argument. In the documentation in models/custom.py it says that when hierarchical==False, the model should return the logits of the classification task. However, both the models/resnet.py and models/convnext.py implementations appear incorrect.

• In the former, the forward method skips layers 1-4, resulting in a channel mismatch when passing the output of self.maxpool(x) to the fully connected layer.
• In the latter, the forward method method skips right to the fully connected layer, also resulting in a tensor sizer mismatch error

Is the hierarchical parameter even necessary? Is there a scenario in which it's used during the pretraining and if so, what should the actual behavior be?