Why do you want to precompile the function ? I don't see the point. It goes against the JIT (Just-in-Time compilation) principle. I would use instead AOT (Ahead-of-Time compilation) if you really think it is better to pre-compile whose methods.

Here is the official documentation, including examples. Note that it is possible to define the shape of the input arrays without passing examples for both JIT and AOT.

I post the reason in the docstring. Since I want to compare the training time of several versions, if we do not compile the main function pattern ahead of time, the first-time compilation time will add onto the training time in the collector, and thus it will show that Numba slows down the whole process (from 2050 fps to 1800 fps).

I don't think AOT is a good choice since we cannot determine some arguments' type ahead of time, for example, v_s_ in GAE.

Then I would rather move _compile at the beginning of the brenchmark, rather than here. I thing it makes more sense, because it is really just a benchmarking issue.

I don't think AOT is a good choice since we cannot determine some arguments' type ahead of time, for example, v_s_ in GAE.

I don't now, it is strange. Compilation support variable size numpy array, just the number of dimensions must be known in advance. Knowing that you can also precompile several signatures if necessary.

That's a bit strange for newcomers. Maybe pre_compile can be called during tianshou.init .

I think so, but this conflicts with @duburcqa #193 (comment)

Then I would rather move _compile at the beginning of the brenchmark, rather than here. I thing it makes more sense, because it is really just a benchmarking issue.

There are many benchmarks in tianshou, but some of the benchmarks (in test folder) are also examples for beginners. Therefore, I think it is not a good solution to add pre-compile for every benchmark. Pre-compile in tianshou.__init__.py should be enough.

@duburcqa

I get your point but it is an anti-pattern. JIT is for jit. If precompile is actually able to precompile all you need, then I would use, it works just the same as jit and it is actually intenses to do what you are looking for.

Well AOT is not exactly as portable as JIT in practice because of some limitations. You can put it in init if you want but I really don't like the idea. I would rather keep it as it is right now. Maybe a more comprehensive name for beginner would be enough, like 'init_profiler'. It is not exactly what it does but it is what it is designed for. From this name plus a comment it would be clear it is there on the CI.

But beginners may build up their own code and benchmark and say, "Hey, your provided result is not compatible with my own test", something like #169 and #196. That's a sad thing.
Also, maybe more functions will be added to the future version. Perhaps they need to be compiled too.

I don't think it is a sad thing. People have never been able to properly benchmark a code, you can see this in all but one in a million posts related to stuff this on stackoverflow. It is a petty to lower our standard to match people programming knowledge, especially from people not contributing but only complaining. But I get your point too and I understand your point of view. At the end it can be changed in the future pretty easily, it is not an code design issue so it is not a big deal. As I see you seem to both agree on this point, maybe you should go for it.