Replicating features in resources/Xreducedall.2002.npy

[AvantiShri]

Hello,

I'm trying to replicate the features in resources/Xreducedall.2002.npy. I am able to get within 99% spearman correlation, but I'm not able to replicate the features exactly. Would you be able to provide the script that was used to produce these features? My attempt at replication is at https://github.com/kundajelab/ExPecto/blob/0a337ec04b451ebdadc62368bec8f05f37b8d6cf/example/Replicate%20ExPecto%20Features.ipynb

[jzthree]

Hello-

My script loads the chromatin feature predictions and do this

rank = np.arange(100)
Xreducedall = [np.tensordot(Xfor[:,:100,:],np.exp(i*rank),axes=([1],[0])) for i in [-0.01, -0.02, -0.05, -0.1, -0.2] ] + \
[np.tensordot(Xrev[:,:100,:],np.exp(i*rank),axes=([1],[0])) for i in [-0.01, -0.02, -0.05, -0.1, -0.2] ]
Xreducedall=np.concatenate(Xreducedall,axis=1)

Xfor and Xrev are numpy arrays each with dimensions [batch, 100 spatial bins, 2002 features] for downstream and upsteam of TSS respectively. I used lua code to generate Xfor and Xrev at the time but it should be no different from what you can do with pytorch.

Hope this helps! Jian

[szhang0112]

@AvantiShri I found if you change this line in block 5 shifts = np.array(list(range(-20000,20000,200)))+100 to shifts = np.array(list(range(-20000,20000,200))), the discrepancy would be almost zero.