|
| 1 | +import math |
| 2 | +from argparse import ArgumentParser |
| 3 | +from pathlib import Path |
| 4 | +from time import perf_counter_ns |
| 5 | + |
| 6 | +import torch |
| 7 | +from torch import Tensor |
| 8 | +from torchcodec._core import add_video_stream, create_from_file, get_frames_by_pts |
| 9 | +from torchcodec.decoders import VideoDecoder |
| 10 | +from torchvision.transforms import v2 |
| 11 | + |
| 12 | +DEFAULT_NUM_EXP = 20 |
| 13 | + |
| 14 | + |
| 15 | +def bench(f, *args, num_exp=DEFAULT_NUM_EXP, warmup=1) -> Tensor: |
| 16 | + |
| 17 | + for _ in range(warmup): |
| 18 | + f(*args) |
| 19 | + |
| 20 | + times = [] |
| 21 | + for _ in range(num_exp): |
| 22 | + start = perf_counter_ns() |
| 23 | + f(*args) |
| 24 | + end = perf_counter_ns() |
| 25 | + times.append(end - start) |
| 26 | + return torch.tensor(times).float() |
| 27 | + |
| 28 | + |
| 29 | +def report_stats(times: Tensor, unit: str = "ms", prefix: str = "") -> float: |
| 30 | + mul = { |
| 31 | + "ns": 1, |
| 32 | + "µs": 1e-3, |
| 33 | + "ms": 1e-6, |
| 34 | + "s": 1e-9, |
| 35 | + }[unit] |
| 36 | + times = times * mul |
| 37 | + std = times.std().item() |
| 38 | + med = times.median().item() |
| 39 | + mean = times.mean().item() |
| 40 | + min = times.min().item() |
| 41 | + max = times.max().item() |
| 42 | + print( |
| 43 | + f"{prefix:<45} {med = :.2f}, {mean = :.2f} +- {std:.2f}, {min = :.2f}, {max = :.2f} - in {unit}" |
| 44 | + ) |
| 45 | + |
| 46 | + |
| 47 | +def torchvision_resize( |
| 48 | + path: Path, pts_seconds: list[float], dims: tuple[int, int] |
| 49 | +) -> None: |
| 50 | + decoder = create_from_file(str(path), seek_mode="approximate") |
| 51 | + add_video_stream(decoder) |
| 52 | + raw_frames, *_ = get_frames_by_pts(decoder, timestamps=pts_seconds) |
| 53 | + return v2.functional.resize(raw_frames, size=dims) |
| 54 | + |
| 55 | + |
| 56 | +def torchvision_crop( |
| 57 | + path: Path, pts_seconds: list[float], dims: tuple[int, int], x: int, y: int |
| 58 | +) -> None: |
| 59 | + decoder = create_from_file(str(path), seek_mode="approximate") |
| 60 | + add_video_stream(decoder) |
| 61 | + raw_frames, *_ = get_frames_by_pts(decoder, timestamps=pts_seconds) |
| 62 | + return v2.functional.crop(raw_frames, top=y, left=x, height=dims[0], width=dims[1]) |
| 63 | + |
| 64 | + |
| 65 | +def decoder_native_resize( |
| 66 | + path: Path, pts_seconds: list[float], dims: tuple[int, int] |
| 67 | +) -> None: |
| 68 | + decoder = create_from_file(str(path), seek_mode="approximate") |
| 69 | + add_video_stream(decoder, transform_specs=f"resize, {dims[0]}, {dims[1]}") |
| 70 | + return get_frames_by_pts(decoder, timestamps=pts_seconds)[0] |
| 71 | + |
| 72 | + |
| 73 | +def decoder_native_crop( |
| 74 | + path: Path, pts_seconds: list[float], dims: tuple[int, int], x: int, y: int |
| 75 | +) -> None: |
| 76 | + decoder = create_from_file(str(path), seek_mode="approximate") |
| 77 | + add_video_stream(decoder, transform_specs=f"crop, {dims[0]}, {dims[1]}, {x}, {y}") |
| 78 | + return get_frames_by_pts(decoder, timestamps=pts_seconds)[0] |
| 79 | + |
| 80 | + |
| 81 | +def main(): |
| 82 | + parser = ArgumentParser() |
| 83 | + parser.add_argument("--path", type=str, help="path to file", required=True) |
| 84 | + parser.add_argument( |
| 85 | + "--num-exp", |
| 86 | + type=int, |
| 87 | + default=DEFAULT_NUM_EXP, |
| 88 | + help="number of runs to average over", |
| 89 | + ) |
| 90 | + |
| 91 | + args = parser.parse_args() |
| 92 | + path = Path(args.path) |
| 93 | + |
| 94 | + metadata = VideoDecoder(path).metadata |
| 95 | + duration = metadata.duration_seconds |
| 96 | + |
| 97 | + print( |
| 98 | + f"Benchmarking {path.name}, duration: {duration}, codec: {metadata.codec}, averaging over {args.num_exp} runs:" |
| 99 | + ) |
| 100 | + |
| 101 | + input_height = metadata.height |
| 102 | + input_width = metadata.width |
| 103 | + fraction_of_total_frames_to_sample = [0.005, 0.01, 0.05, 0.1] |
| 104 | + fraction_of_input_dimensions = [0.5, 0.25, 0.125] |
| 105 | + |
| 106 | + for num_fraction in fraction_of_total_frames_to_sample: |
| 107 | + num_frames_to_sample = math.ceil(metadata.num_frames * num_fraction) |
| 108 | + print( |
| 109 | + f"Sampling {num_fraction * 100}%, {num_frames_to_sample}, of {metadata.num_frames} frames" |
| 110 | + ) |
| 111 | + uniform_timestamps = [ |
| 112 | + i * duration / num_frames_to_sample for i in range(num_frames_to_sample) |
| 113 | + ] |
| 114 | + |
| 115 | + for dims_fraction in fraction_of_input_dimensions: |
| 116 | + dims = (int(input_height * dims_fraction), int(input_width * dims_fraction)) |
| 117 | + |
| 118 | + times = bench( |
| 119 | + torchvision_resize, path, uniform_timestamps, dims, num_exp=args.num_exp |
| 120 | + ) |
| 121 | + report_stats(times, prefix=f"torchvision_resize({dims})") |
| 122 | + |
| 123 | + times = bench( |
| 124 | + decoder_native_resize, |
| 125 | + path, |
| 126 | + uniform_timestamps, |
| 127 | + dims, |
| 128 | + num_exp=args.num_exp, |
| 129 | + ) |
| 130 | + report_stats(times, prefix=f"decoder_native_resize({dims})") |
| 131 | + print() |
| 132 | + |
| 133 | + center_x = (input_height - dims[0]) // 2 |
| 134 | + center_y = (input_width - dims[1]) // 2 |
| 135 | + times = bench( |
| 136 | + torchvision_crop, |
| 137 | + path, |
| 138 | + uniform_timestamps, |
| 139 | + dims, |
| 140 | + center_x, |
| 141 | + center_y, |
| 142 | + num_exp=args.num_exp, |
| 143 | + ) |
| 144 | + report_stats( |
| 145 | + times, prefix=f"torchvision_crop({dims}, {center_x}, {center_y})" |
| 146 | + ) |
| 147 | + |
| 148 | + times = bench( |
| 149 | + decoder_native_crop, |
| 150 | + path, |
| 151 | + uniform_timestamps, |
| 152 | + dims, |
| 153 | + center_x, |
| 154 | + center_y, |
| 155 | + num_exp=args.num_exp, |
| 156 | + ) |
| 157 | + report_stats( |
| 158 | + times, prefix=f"decoder_native_crop({dims}, {center_x}, {center_y})" |
| 159 | + ) |
| 160 | + print() |
| 161 | + |
| 162 | + |
| 163 | +if __name__ == "__main__": |
| 164 | + main() |
0 commit comments