The simulation data here is 19K (11K urban+8.8K suburban). I guess the performance of these two are similar. Please find the links of the experiments below:

pretrained: https://www.comet.ml/tianyu-z/advent-ccai/1e62883c43814d40b9da12e13970bacf

non-pretrained: https://www.comet.ml/tianyu-z/advent-ccai/b9b25cbeb47843dda48f4ce593797b66

P.S. They are not trained at the same starting time. The pretrained one started 5 hrs ahead of the non-pretrained one and the training process is still ongoing.