Questions after beat update
emiliaxin opened this issue · comments
Dear @hvasbath ,
I 'm sorry to disturb you with this question. I redid example 3 and example 4 and I noticed that rake is not the same as before, did I do something wrong?
config_geometry.txt
config_ffi.txt
BEAT_log.txt
Thank you very much
Looking at your config_ffi I would say your SMC sampler configuration is not ensuring convergence of the sampler. You have too little nchains and nsteps in order to make convergence of the sampling likely. Remember you are sampling in this step ~200 parameters. For that 500 chains and 100 steps is way too little! I would say you need at least 3k and 300, respectively. Better even more. The updates did not include any changes in the inference architecture.
Dear @hvasbath ,
I have set up 3k chains and 300 steps when redoing example 4. However, I got the wrong result:
Then, I set rake to "rake: -114.58259929068664" according to example 4 and got the same correct slip distribution as in example 4. Therefore, I guess the rake = -82.96 calculated in Example 3 may not be correct.
My question is:
In example 3, my settings are: 496 chains and 100 steps, previous versions calculated rake at -110 ~ -120 ( beat 1.1 documentation: rake=-114.58259929068664). But after this update, the rake is calculated at -80 ~ -90 (example 3), should I set 3k chains and 300 steps for example 3 as well?
example 3_summary.txt
example 3_config_geometry.txt
That explains it. If your geometry estimation results at the rake of -82 this will be the reference rake to which the static slip estimation is rotated, then allowing only variation in rake according to your uperp prior bound.
Yes also 500 chains and 100 steps are too little to assure convergence for the geometry step as you estimate time-shifts, i.e. an additional parameter for EACH station and EACH channel, may easily result again at 50-100 parameters additional. But 3k and 300 would be to high. But if you have the computational power you can still use it. But sth like 2000 chains and 200 steps, may be enough.
Thank you for your patient reply, I will continue to learn the beat.