Debugging odin models can be challenging because:

They’re not composable - you end up with a fairly large set of equations that govern your system, and you can’t easily split this into smaller testable units and compose them together.
You are writing in the DSL but the model runs in some other language; this sometimes behaves unexpectedly and is much less inspectable than just using R
You can’t (easily) interrupt the running of the program at any point and inspect it

Here, we outline some strategies for debugging, and describe the new features that aim to make this easier.

library(odin2)
library(dust2)

Using `print()`

As of odin 1.4.5, you can print the value of some variables in the middle of running your model. We will expand and change this functionality in future versions, your feedback is very welcome.

Consider the simple model below, which illustrates the idea:

gen <- odin2::odin({
  update(x) <- Normal(x, 1)
  initial(x) <- 1
  print("x: {x}")
}, quiet = TRUE, debug = TRUE)
sys <- dust_system_create(gen(), list(), 1)
dust_system_run_to_time(sys, 10)
#> [0.000000] x: 0.000000
#> [1.000000] x: 0.367074
#> [2.000000] x: -0.682729
#> [3.000000] x: -2.373219
#> [4.000000] x: -3.115467
#> [5.000000] x: -3.114798
#> [6.000000] x: -2.126540
#> [7.000000] x: -2.125140
#> [8.000000] x: -3.459709
#> [9.000000] x: -4.160030

Here we’ve told odin that we want to watch the variable x and print its value at every evaluation (the third line of the model code. When we run the model it prints out the time in square brackets then the debug information following. Notice that we only requested the solution at times 0 and 0.1 but the debug information shows every point in time that the ODE solver evaluated this system of equations.

While this function shares its name with R’s print() it has entirely different functionality.

`print` format strings

For print formatting, we use glue to drive the formatting, and if you have used that package the format will feel familiar.

The most simple usage is as above; you can refer to variables within {curly braces}; so long as your variable is a scalar this will work. Outside of curly braces the string is printed verbatim.

Conditional display

If your model takes many steps, or if you want to narrow down on a problem, you may want to enable conditional display of your debug information. Use the argument when = to control display, such as

print("x: {x}", when = x > 1)

which will display the value of x when it is greater than 1. You can chain together expressions with parentheses and && or || and reference any value in your system. For example:

printf("{x} {y} {z}", when = x > (x + y + z) / 2 && a < 1)

Controlling precision

You can control the way that quantities are displayed through the use of formatting options. The formatting is the same as used by R, so you can experiment in the console easily. The default is to print as a generic floating point number, so this:

print("x: {x}, y: {y}")

is roughly equivalent to writing

sprintf("x: %f, y: %f", x, y)

See ?sprintf for more information; but this defaults to 6 decimal places of precision. This may not be appropriate if you are dealing with numbers that are very large or very small; these both look a bit silly:

sprintf("x: %f, y: %f", 1e-7, 1e7)
#> [1] "x: 0.000000, y: 10000000.000000"

The first loses all information - the only non-zero parts of the number fall after the precision cut-off, while in the second the 6 decimal places just add noise. So above we might prefer:

sprintf("x: %g, y: %g", 1e-7, 1e7)
#> [1] "x: 1e-07, y: 1e+07"

which we could write in odin’s approach as

print("x: {x; g}, y: {y; g}")

Anything after the ; is interpreted as a format specifier. You could also do

print("x: {x; g}, y: {y; .2f}")

which would format y to 2 decimal places. We follow here the example of the sprintf transformer example in glue by not including the % placeholder, but allow all formats that the underlying library supports.

Current limitations

This is an experimental interface, and it has not been exposed to much real-world use. As such it is possible that you might write fairly innocent looking code and it produce a compiler error rather than a nicer R error - please let us know so we can fix this.

There’s no good way of printing out the contents of an array aside from indexing into it. That’s possibly a reasonable thing to do though, given that most arrays get very large very quickly.
You can’t yet control the way that time is formatted (e.g., disabling it or changing the precision)
The print statement only runs in your right-hand-side function (ODE models) or update function (discrete time models) and so it’s possible that some variables that you refer to in your print statements won’t exist in this function (e.g., transient quantities used only to compute some initial condition). We hope this is rare in real-use examples but welcome minimal examples that show where this causes problems (likely you will see a compiler error)
We print the result at the end of the rhs/update function; if you have a crash (or are writing off the end of memory) then this might not be what you want (e.g., the variables you see are the ones in the iteration prior to the crash, or after they have been overwritten by junk). We may support printing more eagerly, after all dependencies in the expression are satisfied, with an additional option to print
Be careful of using integer printing (e.g., {x; d}) for variables that are merely integer-like, or you will get unexpected junk output out. You can however write {as.integer(x); d} which will do a conversion to integer and then print that

Show the generated code

Sometimes just looking at the generated code can be helpful. You can do this with odin_show:

odin_show({
  initial(x) <- 0
  update(x) <- Normal(x, 1)
})
#> 
#> ── odin code: ──────────────────────────────────────────────────────────────────
#> #include <dust2/common.hpp>
#> // [[dust2::class(odin)]]
#> // [[dust2::time_type(discrete)]]
#> class odin {
#> public:
#>   odin() = delete;
#>   using real_type = double;
#>   using rng_state_type = monty::random::generator<real_type>;
#>   struct shared_state {
#>     struct offset_type {
#>       struct {
#>         size_t x;
#>       } state;
#>     } offset;
#>   };
#>   struct internal_state {};
#>   using data_type = dust2::no_data;
#>   static dust2::packing packing_state(const shared_state& shared) {
#>     return dust2::packing{
#>       {"x", {}}
#>     };
#>   }
#>   static dust2::packing packing_gradient(const shared_state& shared) {
#>     return dust2::packing{
#>     };
#>   }
#>   static shared_state build_shared(cpp11::list parameters) {
#>     shared_state::offset_type offset;
#>     offset.state.x = 0;
#>     return shared_state{offset};
#>   }
#>   static internal_state build_internal(const shared_state& shared) {
#>     return internal_state{};
#>   }
#>   static void update_shared(cpp11::list parameters, shared_state& shared) {
#>   }
#>   static void update_internal(const shared_state& shared, internal_state& internal) {
#>   }
#>   static void initial(real_type time, const shared_state& shared, internal_state& internal, rng_state_type& rng_state, real_type* state) {
#>     state[0] = 0;
#>   }
#>   static void update(real_type time, real_type dt, const real_type* state, const shared_state& shared, internal_state& internal, rng_state_type& rng_state, real_type* state_next) {
#>     const auto x = state[0];
#>     state_next[0] = monty::random::normal<real_type>(rng_state, x, 1);
#>   }
#>   static auto zero_every(const shared_state& shared) {
#>     return dust2::zero_every_type<real_type>();
#>   }
#> };

- Using print()
  - print format strings
  - Current limitations
- Show the generated code

Debugging