Handle videos with begin_stream_seconds > 0

[ahmadsharif1]

Without this PR I get error messages like this:

python benchmarks/decoders/benchmark_decoders.py --decoders=torchcodec_core,torchcodec_core:device=cuda --bm_video_paths=/home/ahmads/jupyter/853.mp4 
video_files_paths=['/home/ahmads/jupyter/853.mp4']
video=/home/ahmads/jupyter/853.mp4, decoder=TorchCodecCore:device=cuda
[AVHWDeviceContext @ 0x9799dc0] Using current CUDA context.
Traceback (most recent call last):
  File "/home/ahmads/personal/torchcodec/benchmarks/decoders/benchmark_decoders.py", line 181, in <module>
    main()
  File "/home/ahmads/personal/torchcodec/benchmarks/decoders/benchmark_decoders.py", line 169, in main
    df_data = run_benchmarks(
              ^^^^^^^^^^^^^^^
  File "/home/ahmads/personal/torchcodec/benchmarks/decoders/benchmark_decoders_library.py", line 538, in run_benchmarks
    seeked_result.blocked_autorange(min_run_time=min_runtime_seconds)
  File "/home/ahmads/.conda/envs/tcgpu1/lib/python3.12/site-packages/torch/utils/benchmark/utils/timer.py", line 372, in blocked_autorange
    number = self._estimate_block_size(min_run_time)
             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/ahmads/.conda/envs/tcgpu1/lib/python3.12/site-packages/torch/utils/benchmark/utils/timer.py", line 319, in _estimate_block_size
    time_taken = self._timeit(number)
                 ^^^^^^^^^^^^^^^^^^^^
  File "/home/ahmads/.conda/envs/tcgpu1/lib/python3.12/site-packages/torch/utils/benchmark/utils/timer.py", line 264, in _timeit
    return max(self._timer.timeit(number), 1e-9)
               ^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/ahmads/.conda/envs/tcgpu1/lib/python3.12/timeit.py", line 180, in timeit
    timing = self.inner(it, self.timer)
             ^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "<timeit-src>", line 6, in inner
  File "/home/ahmads/personal/torchcodec/benchmarks/decoders/benchmark_decoders_library.py", line 130, in get_frames_from_video
    frames, *_ = get_frames_by_pts(
                 ^^^^^^^^^^^^^^^^^^
  File "/home/ahmads/.conda/envs/tcgpu1/lib/python3.12/site-packages/torch/_ops.py", line 667, in __call__
    return self_._op(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^
RuntimeError: frame pts is 0.000000; must be in range [0.015999, 5.004316).

After this PR, I can run the benchmark on that video:

python benchmarks/decoders/benchmark_decoders.py --decoders=torchcodec_core,torchcodec_core:device=cuda --bm_video_paths=/home/ahmads/jupyter/853.mp4 
video_files_paths=['/home/ahmads/jupyter/853.mp4']
video=/home/ahmads/jupyter/853.mp4, decoder=TorchCodecCore
video=/home/ahmads/jupyter/853.mp4, decoder=TorchCodecCore:device=cuda
[AVHWDeviceContext @ 0x7029c00] Using current CUDA context.
[--------------------------------- video=/home/ahmads/jupyter/853.mp4 h264 3840x2160, 4.988317s 59.94005593469701fps ---------------------------------]
                                  |  uniform 10 seek()+next()  |  random 10 seek()+next()  |  1 next()  |  10 next()  |  100 next()  |  create()+next()
1 threads: --------------------------------------------------------------------------------------------------------------------------------------------
      TorchCodecCore              |           1897.6           |           1965.3          |   288.5    |    531.0    |    1701.9    |       279.0     
      TorchCodecCore:device=cuda  |           3260.1           |           2703.9          |   345.7    |    597.0    |    1092.1    |                 

Times are in milliseconds (ms).

[NicolasHug]

Approving to unblock but maybe try arange (or linspace?) as it might provide a cleaner solution

[NicolasHug]

Suggestion: try to rely on torch.arange instead, like we do in the samplers

torchcodec/src/torchcodec/samplers/_time_based.py

Line 198 in c2bea4b

clip_start_seconds = torch.arange(

. It should start exactly at begin which hopefully wouldn't cause the error we're seeing.

[scotts]

I had the same thought when I was refactoring things, and I can't remember why I didn't do it. :) I'll dig in later.

[ahmadsharif1]

Seems like arange doesn't work here, possibly due to rounding:

        uniform_pts_list = list(
            torch.arange(
                metadata.begin_stream_seconds,
                metadata.end_stream_seconds,
                (metadata.end_stream_seconds - metadata.begin_stream_seconds)
                / num_samples,
            )
        )

Gives me the same error:

RuntimeError: frame pts is 0.015999; must be in range [0.015999, 5.004316).