Code profiling support

We provide built-in performance profiling capabilities to make it easier for you to debug and assess the performance of your code.

In-house profiler

You can use our built-in profiler. By default, it is enabled and will print at the termination of the program. You can disable it via CLI using the flag --logging.no-enable-profiler.

The profiler computes the average total time of execution for any function with the @profiler.time_function decorator. For instance, if you wanted to profile the total time it takes to generate rays given pixel and camera indices via the RayGenerator class, you might want to time its forward() function. In that case, you would need to add the decorator to the function.

"""nerfstudio/model_components/ray_generators.py""""

class RayGenerator(nn.Module):

    ...

    @profiler.time_function     # <-- add the profiler decorator before the function
    def forward(self, ray_indices: TensorType["num_rays", 3]) -> RayBundle:
        # implementation here
        ...

Alternatively, you can also time parts of the code:

...

def forward(self, ray_indices: TensorType["num_rays", 3]) -> RayBundle:
    # implementation here
    with profiler.time_function("code1"):
      # some code here
      ...

    with profiler.time_function("code2"):
      # some code here
      ...
    ...

At termination of training or end of the training run, the profiler will print out the average execution time for all of the functions or code blocks that have the profiler tag.

Tip

Use this profiler if there are specific/individual functions you want to measure the times for.

Profiling with PyTorch profiler

If you want to profile the training or evaluation code and track the memory and CUDA kernel launches, consider using PyTorch profiler. It will run the profiler for some selected step numbers, once with CUDA_LAUNCH_BLOCKING=1, once with CUDA_LAUNCH_BLOCKING=0. The PyTorch profiler can be enabled with --logging.profiler=pytorch flag. The outputs of the profiler are trace files stored in {PATH_TO_MODEL_OUTPUT}/profiler_traces, and can be loaded in Google Chrome by typing chrome://tracing.

Profiling with PySpy

If you want to profile the entire codebase, consider using PySpy.

Install PySpy

pip install py-spy

To perform the profiling, you can either specify that you want to generate a flame graph or generate a live-view of the profiler.

• flame graph: with wandb logging and our inhouse logging disabled

  program="ns-train nerfacto -- --vis=wandb --logging.no-enable-profiler blender-data"
  py-spy record -o {PATH_TO_OUTPUT_SVG} $program
  

• top-down stats: running same program configuration as above

  py-spy top $program
  

Attention

In defining program, you will need to add an extra -- before you specify your program’s arguments.