James-Thorson
commented
7 months ago Ben,
I'm guessing that you have a mismatch between Matrix and TMB. Have you confirmed that TMB is working using the example here ( https://github.com/pfmc-assessments/geostatistical_delta-GLMM/wiki/Steps-to-install-TMB) ...?
Also, feel free to email me about tinyVAST ( https://vast-lib.github.io/tinyVAST/). I'm maintaining VAST for the next couple years, but my development efforts are focused on tinyVAST which has similar functionality using a smaller code base and regression interface.
Jim
On Wed, Jan 31, 2024 at 11:04 AM Blevy2 @.***> wrote:
Hi Jim!
I am running into an issue that I am having trouble understanding that I wanted to run by you.
I am running VAST models on samples taken from spatial population simulation output for fish species that I have developed. In the spatial population models, the probability a fish moves to discrete cell $(i,j)$ in week $w$ is given by probability $Move{w,i,j}$, where $Move{w,i,j}$ depends on factors such as the water temperature in the given cell, $Temp_{w,i,j}$.
I want to compare the performance of VAST models with and without covariates, so I could reasonably provide VAST with either $Move{w,i,j}$ and/or $Temp{w,i,j}$ as the covariate. Since $Temp{w,i,j}$ is just one component of the actual movement probability $Move{w,i,j}$, my assumption is that $Move{w,i,j}$ would provide more information to VAST about species spatial preferences and thus provide a more accurate estimate compared to $Temp{w,i,j}$. I am using a second degree polynomial response when including covariates. For example, if using $Move$ as the covariate I input
X2_formula = ~ poly(Move, degree=2 ) X1_formula = ~ poly(Move, degree=2 )My models without covariates are all converging fine and my models that use $Temp{w,i,j}$ as the covariate are also converging, but models that use $Move{w,i,j}$ as the covariate all seem to run nearly to completion before they produce the error
<simpleError in nlminb(start = startpar, objective = fn, gradient = gr, control = nlminb.control, lower = lower, upper = upper): NA/NaN gradient evaluation>There are a few github issues for VAST that involve NA/NaN function evaluation, but not NA/NaN gradient evaluation. The closest thing I could find related to this is this issue thread in glmmTMB: glmmTMB/glmmTMB#164 https://github.com/glmmTMB/glmmTMB/issues/164
Based on the discussion in the above issue link our best guess is that maybe the gradient function created in VAST has a term something like log(exp(X)) where X is some parameter. During the final stage of a VAST run, possibly when the final Hessian is being calculated, the X value gets really really small (e.g., 1E-320) and the exp(X) goes to zero due to computer rounding and thus the log(exp(X)) is undefined. Does that sound reasonable?
Are you familiar with this error? Do you know how to fix this so we can use $Move$ as a covariate?
Thanks for your help!
Ben
— Reply to this email directly, view it on GitHub https://github.com/James-Thorson-NOAA/VAST/issues/383, or unsubscribe https://github.com/notifications/unsubscribe-auth/AB46UTIAKL5M2WUXRTWHHQTYRKIVLAVCNFSM6AAAAABCTS7YBKVHI2DSMVQWIX3LMV43ASLTON2WKOZSGEYTANZXGU2DSNA . You are receiving this because you are subscribed to this thread.Message ID: @.***>
Blevy2
Hi Jim!
I am running into an issue that I am having trouble understanding that I wanted to run by you.
I am running VAST models on samples taken from spatial population simulation output for fish species that I have developed. In the spatial population models, the probability a fish moves to discrete cell $(i,j)$ in week $w$ is given by probability $Move{w,i,j}$, where $Move{w,i,j}$ depends on factors such as the water temperature in the given cell, $Temp_{w,i,j}$.
I want to compare the performance of VAST models with and without covariates, so I could reasonably provide VAST with either $Move{w,i,j}$ and/or $Temp{w,i,j}$ as the covariate. Since $Temp{w,i,j}$ is just one component of the actual movement probability $Move{w,i,j}$, my assumption is that $Move{w,i,j}$ would provide more information to VAST about species spatial preferences and thus provide a more accurate estimate compared to $Temp{w,i,j}$. I am using a second degree polynomial response when including covariates. For example, if using $Move$ as the covariate I input
My models without covariates are all converging fine and my models that use $Temp{w,i,j}$ as the covariate are also converging, but models that use $Move{w,i,j}$ as the covariate all seem to run nearly to completion before they produce the error
There are a few github issues for VAST that involve NA/NaN function evaluation, but not NA/NaN gradient evaluation. The closest thing I could find related to this is this issue thread in glmmTMB: https://github.com/glmmTMB/glmmTMB/issues/164
Based on the discussion in the above issue link our best guess is that maybe the gradient function created in VAST has a term something like log(exp(X)) where X is some parameter. During the final stage of a VAST run, possibly when the final Hessian is being calculated, the X value gets really really small (e.g., 1E-320) and the exp(X) goes to zero due to computer rounding and thus the log(exp(X)) is undefined. Does that sound reasonable?
Are you familiar with this error? Do you know how to fix this so we can use $Move$ as a covariate?
Here is my Seesion.Info():
attached base packages: [1] stats graphics grDevices utils datasets methods base
other attached packages: [1] dplyr_1.1.3 VAST_3.10.1 FishStatsUtils_2.12.1 [4] marginaleffects_0.15.1 units_0.8-4 TMB_1.9.6
loaded via a namespace (and not attached): [1] utf8_1.2.4 R6_2.5.1 tidyselect_1.2.0 Matrix_1.6-1.1
[5] lattice_0.20-44 magrittr_2.0.3 INLA_23.05.30-1 splines_4.3.2
[9] glue_1.6.2 tibble_3.2.1 pkgconfig_2.0.3 generics_0.1.2
[13] lifecycle_1.0.4 cli_3.6.1 fansi_1.0.5 vctrs_0.6.4
[17] grid_4.3.2 data.table_1.14.4 compiler_4.3.2 sp_1.5-0
[21] pillar_1.9.0 Rcpp_1.0.11 rlang_1.1.2
[1] "input is"
Thanks for your help!
Ben