Greetings,
I come seeking guidance on a perplexing issue regarding data transformation for efficiently predicting the phase of complex numbers using PyTorch. Despite extensive experimentation, I have yet to find a transformation method that yields satisfactory results.
I have explored various techniques including Cox-Box, Yeo-Johnson, and quantile transformations, but unfortunately, none of them have provided the desired outcome for predicting the phase of complex numbers accurately.
In my pursuit of a solution, I devised a custom transformation function as follows:
def transform_y(y, a, b):
return np.log(y**a + 1)**b
def inverse_transform(y_transformed, a, b):
return (np.exp(y_transformed**(1/b)) - 1)**(1/a)
I then fine-tuned the parameters a and b to minimize the Root Mean Squared Error (RMSE) of the transformed data. Surprisingly, this transformation appears to perform well in predicting the transformed data, with an RMSE of approximately 10e-4.
However, the issue arises when applying the inverse transformation. Despite the promising results obtained during prediction, the application of the inverse transformation results in an infinite RMSE, indicating a severe discrepancy.
I am reaching out to the PyTorch community for insights, suggestions, or alternative approaches to tackle this challenge. Any advice or recommendations on more effective data transformation techniques or adjustments to the existing transformation method would be greatly appreciated.
Thank you in advance for your assistance and expertise.