fix: store cuda_available variable and extend perf_counter fix to all get_computational_cost functions

[vinitjain2005]

Summary

This PR extends the timing fix originally proposed in the closed PR to all get_computational_cost() functions across the library, and also addresses the review comment to store torch.cuda.is_available() in a variable instead of calling it on every loop iteration.

Closes #453

Problem

The get_computational_cost() functions across multiple files had two issues:

1. torch.cuda.synchronize() was called unconditionally on every iteration — it is a no-op on CPU and was designed only for CUDA devices

2. time.time() was used for timing — it has low resolution on Windows (15ms granularity) causing timing to show 0ms or jump in 15ms chunks

3. torch.cuda.is_available() was being called on every loop iteration instead of being stored once as a variable

Fix

# Before
for _ in range(runs):
    torch.cuda.synchronize()        # unconditional, no-op on CPU
    start = time.time()             # low resolution on Windows
    ...
    torch.cuda.synchronize()
    inference_times.append(time.time() - start)

# After
cuda_available = torch.cuda.is_available()  # stored once
for _ in range(runs):
    if cuda_available:
        torch.cuda.synchronize()
    start = time.perf_counter()     # high resolution on all platforms
    ...
    if cuda_available:
        torch.cuda.synchronize()
    inference_times.append(time.perf_counter() - start)

Files Changed

• perceptionmetrics/models/torch_detection.py

  • Fixed standalone get_computational_cost() function

• perceptionmetrics/models/torch_segmentation.py

  • Fixed TorchImageSegmentationModel.get_computational_cost()
  • Fixed TorchLiDARSegmentationModel.get_computational_cost()

• perceptionmetrics/models/tf_segmentation.py

  • Fixed TensorflowImageSegmentationModel.get_computational_cost()
  • Note: Uses has_gpu variable (already stored before loop)
    and replaces time.time() with time.perf_counter()

Impact

Every user running computational cost estimation on CPU gets accurate timing results. This is especially relevant since CUDA is optional and most contributors and new users run on CPU-only machines.

References

• PyTorch docs: torch.cuda.synchronize() only synchronizes CUDA device operations
• Python docs: time.perf_counter() is the recommended high-resolution timer for benchmarking

Github: @vinitjain2005

[dpascualhe]

Good! Upon fixing these small issues, we can merge.

[vinitjain2005]

@dpascualhe I have updated the changes

[dpascualhe]

Thanks for addressing the previous concerns! Upon closer inspection, there is a last change that would improve your PR. Could you use the self.device attribute in the condition for calling torch.cuda.synchronize instead of the torch.cuda.is_available? In that way, even if CUDA is available, if model is supposed to be CPU, we will avoid calling synchronize unnecessarily. Sorry for the back and forth, just realized that 😅

[vinitjain2005]

I have done the changes and created a new pr request #553 . So i am closing these pr as fix in #553