Modifications to the simulator

Heat Signature Zero
1

Hi,

is it allowed to modify the given simulator/solver (e.g., rewrite it in jax/pytorch to compute gradients, or use a different solver altogether), or do we have to use the provided simulator exactly as it is given?

All the best,
Thomas

1 Like
2

Hi @Th204c great to see you on the forums again :waving_hand: . Please use the simulator as it is provided for this Challenge.

ThinkOnward Team

3

Hi @discourse-admin , how about making new supplementary function for the simulation itself (eg calculating derivative, etc) ?

1 Like
4

Hi @Leo_D good to see you on the forums again. If it happens outside the simulation engine, it is fair game! It just shouldn’t use anything from inside then engine (e.g. the diffusion eq.). If it uses input-output of the engine to calculate a derivative it is totally fine.

Happy heat seaking

ThinkOnward Team

5

Hi @discourse-admin , Can we use GPU accelerated simulator (rewritten with pytorch) for training and use the original simulator for the Inference? Because training with the current simulator is very slow.

6

Hi @pranay and welcome to the ThinkOnward Challenges forum . Please use the simulator as it is provided for this Challenge.

ThinkOnward Team

7

Hi, I do not fully understand this point. First, the guidelines do not specify anything regarding this matter. Second, I do not see an inherent issue with modifying the simulation engine if doing so allows us to improve speed or precision, it is fair (this is my personal opinion).

Additionally, in the notebook, under the section “Example Approaches We Encourage”, you mention that the “Differentiable Simulation or Surrogate Models” approach is permitted. This seems to contradict what is stated in this post. That is, if we use a Differentiable Simulation, we would essentially be constructing a differentiable function based on the original simulator.

I would appreciate it if you could clarify this point in detail. And please do not simply respond with “Please use the simulator as it is provided for this Challenge.” =) Many of us are investing a significant amount of time in this very engaging competition.

Thank you very much.

8

One more thing, and I do not want to be repetitive… I am quite sure that the top positions on the leaderboard are using an improved simulator. As mentioned in another post, the baseline simulator is very slow, and with only a few dozen executions it is not possible to achieve those scores.

Thank you

9

Couldn’t have said it better