Profiling DataLoader

[ndem0]

Describe the solution you'd like
Just a comparison on the performance of the dataloader in these contexts:

• no batches, single condition
• batches, single condition
• no batches, multiple condition
• batches, multiple condition

Hopefully, GPU and CPU

[dario-coscia]

@FilippoOlivo @gc031298 I would do a profiler class for this issue, something as a callback that we can attach to the trainer if needed, what do you think? Once the DataModule is updated we can start!

[FilippoOlivo]

I completely agree! Maybe something inherited from https://github.com/pyutils/line_profiler

[dario-coscia]

After merging #428 PINA 0.2 version is fully compatible with Lightning Profilers which can be easily used by the user.

Example:

from pina import Trainer
from pina.model import FeedForward
from pina.solvers import PINN
from pina.problem.zoo import Poisson2DSquareProblem
from lightning.pytorch.profilers import AdvancedProfiler

# train and profiler
problem = Poisson2DSquareProblem()
problem.discretise_domain(10000)
model = FeedForward(len(problem.input_variables), len(problem.output_variables))
solver = PINN(model, problem)
profiler = AdvancedProfiler(dirpath=".", filename="perf_logs") # also PyTorchProfiler, SimpleProfiler, ....
trainer = Trainer(solver, profiler=profiler)
trainer.train()