Description
Hi there,
I've noticed something weird with the performance of the EQT model - specifically the original model file.
I tried loading the model and then running it on the 100samples.hdf5 data to seem how it performs.
For this, I noticed the P picks were preferentially happening on 100 sample (1 second) boundaries.
For example, here is the list of the locations of the max probability for each of those 100 samples:
Note that many of them are exactly on 100 sample boundaries - this is without any additional logic, just by selecting the index of max probability on the predictions that come out of the model.
I tested to see whether this was a tensorflow versioning issue - I've tried both tensorflow 2.6 and 2.11 and confirmed that both setups have the same issue. I also tested with some other 100 Hz data and noticed the same behavior.
Any ideas? It doesn't seem to happen quite as much when using the conservative model:
Although it does still happen to some degree.
I did not apply any preprocessing to the data in the hdf file.
Any information you can give me would be much appreciated.
Thanks,
Josh W