On 9 May 2022, at 03:41, 'lala z' via TVB Users <tvb-...@googlegroups.com> wrote:
Hello, is there any code related to advi algorithm in bvep library? I didn't find it
--
You received this message because you are subscribed to the Google Groups "TVB Users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to tvb-users+...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/tvb-users/bb669850-338d-4720-856c-c85c3ea36d76n%40googlegroups.com.
On 12 May 2022, at 15:35, 'lala z' via TVB Users <tvb-...@googlegroups.com> wrote:
OK, thank you very much for your reply.I'm using ADVI algorithm with pymc3 in BVEP model recently, but I found a new problem when I use non-central parametrization and give a distribution to the parameters eps and sig, As shown below:
<7273c6e6df6d77a9870e974706113a6.png>
the following error will appear:
<06ccb1a27b367d18e75a2dafd461caa.png>But when I give eps and sig a certain value, for example, eps = 0.06, sig = 0.16, there will be no such problem, and there will be no such problem when using central parameterization?As you said, would it be better to use central parameterization for ADVI?
在2022年5月9日星期一 UTC+8 15:45:04<marmaduke.woodman> 写道:
hi
ADVI is a variation inference algorithm implemented by Stan. PyMC3 and other toolboxes demo'd in the BVEP repository also implement variational inference. In other words, it is the toolbox, but not the modeler who implements ADVI, and all the models in BVEP should already be usable with ADVI.
cheers,Marmaduke
On 9 May 2022, at 03:41, 'lala z' via TVB Users <tvb-...@googlegroups.com> wrote:
Hello, is there any code related to advi algorithm in bvep library? I didn't find it
--
You received this message because you are subscribed to the Google Groups "TVB Users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to tvb-users+...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/tvb-users/bb669850-338d-4720-856c-c85c3ea36d76n%40googlegroups.com.
--
You received this message because you are subscribed to the Google Groups "TVB Users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to tvb-users+...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/tvb-users/c67fa74d-b759-4f0f-9d75-dfae0489ac6dn%40googlegroups.com.
<06ccb1a27b367d18e75a2dafd461caa.png><7273c6e6df6d77a9870e974706113a6.png>
To view this discussion on the web visit https://groups.google.com/d/msgid/tvb-users/c180111c-4c82-43ce-a91f-b49682112809n%40googlegroups.com.
On 30 May 2022, at 14:56, 'lala z' via TVB Users <tvb-...@googlegroups.com> wrote:
Hello, I'm trying to build the BVEP model with Pyro and use Bayesian inference. I use SVI in Pyro to approximate the posterior distribution. I use the central parameterization because the result of non-central parameterization seems very bad, but there is also a strange problem. The result is as follows:
To view this discussion on the web visit https://groups.google.com/d/msgid/tvb-users/d750b49d-9740-4241-a833-70f941766cden%40googlegroups.com.
<46681e6bebf00d81c4eb0de0457e668.png>
On 6 Jun 2022, at 09:54, 'lala z' via TVB Users <tvb-...@googlegroups.com> wrote:
Hello, thank you very much for your patient reply, which is very helpful to me, but this concept is still a little abstract to me. so I tried to change the code and use the vectorization method. I don't know if this is correct.The result is still bad after this modification .Can you give me some suggestions?
<310221577cc8b869bb096d2efe4b6ec.png>
在2022年6月2日星期四 UTC+8 23:27:59<marmaduke.woodman> 写道:
hi
I should have added that the line below
x_{t+1} ~ N(a * x_t, 1)
is not an assignment in code, it's a probability evaluation. In Python, it would be equivalent to something like
logp += normal.lpdf(x[1:], a * x[:-1], 1)
where logp is the objective function to maximize (for optimization).
cheers,Marmaduke
On 2 Jun 2022, at 17:21, WOODMAN Michael <marmaduk...@univ-amu.fr> wrote:
instead the code computes how well the values of x_{t} match the AR(1) process described above, in terms of a normal distribution,
x_{t+1} ~ N(a * x_t, 1)
This is where the dependence in time enters explicitly, but because all values of x_t are available, this can be done in a vectorized fashion without the for loop in time.
--
You received this message because you are subscribed to the Google Groups "TVB Users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to tvb-users+...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/tvb-users/222ee41b-1444-434a-b015-72a7104d061dn%40googlegroups.com.
<310221577cc8b869bb096d2efe4b6ec.png>
On 9 Jun 2022, at 15:03, 'lala z' via TVB Users <tvb-...@googlegroups.com> wrote:
Hello,I have corrected it, but the result is still not ideal. At first, I set both x and z as uninformational priors, as shown below:
<881e3c689efeb2845eb087397ab372b.png>Then I printed out that all values of X sampled from the posterior distribution are 0.When the priors of X and Z are sampled from Normal (-1.5 , 0) and Normal (-3.5 , 0) as follows:
<fad9bcab5dec1731aa832c21cc4e948.png>
all values of X sampled from the posterior distribution are -1.5, as follows:
<b144634ba05bf86a2771126c9dd09f8.png>
I feel like step_ ode() doesn't work, Do you have any ideas?
Thank you ,lala在2022年6月9日星期四 UTC+8 15:30:35<marmaduke.woodman> 写道:
hi,
yes, the difference coupling is wrong, it should read
diff_state = from_state - to_state
as the summand is x_j - x_i with a sum over j.
see the Stan code or others for reference,
cheers,Marmaduke
On 9 Jun 2022, at 05:01, 'lala z' via TVB Users <tvb-...@googlegroups.com> wrote:
Is there something wrong with my code?
--
You received this message because you are subscribed to the Google Groups "TVB Users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to tvb-users+...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/tvb-users/bd800fb3-00dc-4463-9e78-bb071f922a1fn%40googlegroups.com.
<881e3c689efeb2845eb087397ab372b.png><fad9bcab5dec1731aa832c21cc4e948.png><b144634ba05bf86a2771126c9dd09f8.png>