search

juglab / cryoCARE_pip

PIP package of cryoCARE
BSD 3-Clause "New" or "Revised" License
25 stars 14 forks source link

documentation? #50

Open McHaillet opened 1 year ago

McHaillet commented 1 year ago

I was wondering if there is documentation on what the optimal training parameters are? For example, how many tomograms should I provide for network training (I now provided 3, from my dataset of 100)? Same for batch size, kernel size, etc. for which I stuck to the values in the README.

Any help would be appreciated!

JulienMaufront commented 1 year ago

Hi,

Like Marten, I'm also interested in getting an idea of how many tomograms should be used on average for training. Initially, we were told to use all the data (up to 100 tomograms here). But this takes a considerable amount of time (6 days) and even with a loss value that tends to increase over the last few epochs. It's very different from what Marten uses here, so I'd imagine a dozen tomograms would be already enough, isn't it ?

Thank you