Question about DRC Violations Prediction and RouteNet

[wwwwwly]

Thank you very much for your work, which has been incredibly helpful to me. However, I encountered some questions while reading the paper (CircuitNet: An Open-Source Dataset for Machine Learning in VLSI CAD Applications With Improved Domain-Specific Evaluation Metric and Learning Strategies) and exploring the source code of the CircuitNet project.

1. Labels for Model Training
   In the paper, the task is described as: Given a globally routed design, build a model to predict DRC violations after detailed routing. This problem is formulated as a DRC hotspot detection task, with a threshold to categorize DRC hotspot regions and non-hotspot regions.
   The mathematical formulation is given as:
   $g_{DRC}: X \in \mathbb{R}^{w \times h \times f} \rightarrow V_i \in \{0,1\}^{w \times h}$
   However, upon examining the source code, I found that the labels used in the training process are continuous values, rather than binary (0-1) images.

2. Loss Calculation
   In the source code of RouteNet model, the decoder's final layer uses a sigmoid activation function to map the model output to the range [0,1]. Meanwhile, the labels are normalized using min-max scaling. Is it appropriate to use Mean Squared Error (MSE) as the loss function in this case, given the difference in how the output and labels are processed?

[apri0426]

1. In this implementation, the label is binary only in inference. Of course, binary labels can be used for training, and the performance is better, but in this case, the threshold for the model becomes fixed.
2. I don't understand your point. After min-max scaling, the labels are also in the range [0,1]