Question about mean baseline

[Leafaeolian]

Hi author, thanks for your excellent work which really inspire me. However i have a question about mean baseline. This means, instead of using gears's predicted profiles, i mechanically count mean profiles according to profiles in the train set. This sometimes sound a slight stupid solution, but my experiences tell me it works in most of cases. Therefore, i try the mean baseline by simply input “mean_profile” to function(evaluate) and replace "p = model(batch)" with "p = torch.Tensor(mean_profile).repeat(batch_size, 1)"，and the metrics seem no difference with gears's (attached below). Does this mean that gears does not solve the problem of perturbation prediction? Same thing happended in "Single-Cell Perturbations" Competition (https://www.kaggle.com/competitions/open-problems-single-cell-perturbations/overview). But I'm not sure if that's the case here.

All results come from "model_tutorial.ipynb". (dataset = 'norman', epoch = 20)
Gears
{'mse': 0.0046562045,
'delta_pearson': 0.5698893233579129 #
}

Mean baseline
{'mse': 0.005267251,
'delta_pearson': 0.5756597193705737,
}