ig248 / pyampd Goto Github PK

Hi, @ig248 ,

I read the original paper and compared it with your implementation, I am wondering why below code reshapes LSM matrix by LSM[0:l_scale, :] but not by LSM[0:l_scale+1, :]?

As the paper said, it reshapes the LSM matrix by removing all elements m_{k,i} for which k > λ holds. So it should lead to new λ ×N matrix. But in your implementation, the new matrix is (λ-1) ×N.

And under certain dataset, the l_scale may equals to 0, then LSM[0:0] will throw an exception...

Looking forward to your response

I wanted to run the script on my data but always got the error
"ValueError: zero-size array to reduction operation minimum which has no identity" in line 107 of ampd.py

The issue is that the largest value in G (for my data) is the first one and thus
l_scale = np.argmax(G)
leads to l_scale = 0

In line 107 the array LSM is reduced to
LSM[0:l_scale, :]
But as l_scale = 0 this leads to an empty array and an error when trying to to use np.min on it.

I think line 107 should be
pks_logical = np.min(LSM[0:l_scale+1, :], axis=0)

Love the performance of this library for my time series analysis project, I would love to use it for my micropython or circuitpython projects (ESP32, Teensy and Arduino) and I think it perfectly complements upcoming egde ML tool chains from Nvidia-ARM and TinyML.

Thank you!

as the title suggests