|
| 1 | +import logging |
| 2 | +import numpy as np |
| 3 | +from itertools import count |
| 4 | +from typing import List, Tuple |
| 5 | +import torch |
| 6 | +import tqdm |
| 7 | +from fvcore.common.timer import Timer |
| 8 | + |
| 9 | +from detectron2.utils import comm |
| 10 | + |
| 11 | +from .build import build_batch_data_loader |
| 12 | +from .common import DatasetFromList, MapDataset |
| 13 | +from .samplers import TrainingSampler |
| 14 | + |
| 15 | +logger = logging.getLogger(__name__) |
| 16 | + |
| 17 | + |
| 18 | +class _EmptyMapDataset(torch.utils.data.Dataset): |
| 19 | + """ |
| 20 | + Map anything to emptiness. |
| 21 | + """ |
| 22 | + |
| 23 | + def __init__(self, dataset): |
| 24 | + self.ds = dataset |
| 25 | + |
| 26 | + def __len__(self): |
| 27 | + return len(self.ds) |
| 28 | + |
| 29 | + def __getitem__(self, idx): |
| 30 | + _ = self.ds[idx] |
| 31 | + return [0] |
| 32 | + |
| 33 | + |
| 34 | +def iter_benchmark( |
| 35 | + iterator, num_iter: int, warmup: int = 5, max_time_seconds: float = 60 |
| 36 | +) -> Tuple[float, List[float]]: |
| 37 | + """ |
| 38 | + Benchmark an iterator/iterable for `num_iter` iterations with an extra |
| 39 | + `warmup` iterations of warmup. |
| 40 | + End early if `max_time_seconds` time is spent on iterations. |
| 41 | +
|
| 42 | + Returns: |
| 43 | + float: average time (seconds) per iteration |
| 44 | + list[float]: time spent on each iteration. Sometimes useful for further analysis. |
| 45 | + """ |
| 46 | + num_iter, warmup = int(num_iter), int(warmup) |
| 47 | + |
| 48 | + iterator = iter(iterator) |
| 49 | + for _ in range(warmup): |
| 50 | + next(iterator) |
| 51 | + timer = Timer() |
| 52 | + all_times = [] |
| 53 | + for curr_iter in tqdm.trange(num_iter): |
| 54 | + start = timer.seconds() |
| 55 | + if start > max_time_seconds: |
| 56 | + num_iter = curr_iter |
| 57 | + break |
| 58 | + next(iterator) |
| 59 | + all_times.append(timer.seconds() - start) |
| 60 | + avg = timer.seconds() / num_iter |
| 61 | + return avg, all_times |
| 62 | + |
| 63 | + |
| 64 | +class DataLoaderBenchmark: |
| 65 | + """ |
| 66 | + Some common benchmarks that help understand perf bottleneck of a standard dataloader |
| 67 | + made of dataset, mapper and sampler. |
| 68 | + """ |
| 69 | + |
| 70 | + def __init__( |
| 71 | + self, |
| 72 | + dataset, |
| 73 | + *, |
| 74 | + mapper, |
| 75 | + sampler=None, |
| 76 | + total_batch_size, |
| 77 | + num_workers=0, |
| 78 | + max_time_seconds: int = 90, |
| 79 | + ): |
| 80 | + """ |
| 81 | + Args: |
| 82 | + max_time_seconds (int): maximum time to spent for each benchmark |
| 83 | + other args: same as in `build.py:build_detection_train_loader` |
| 84 | + """ |
| 85 | + if isinstance(dataset, list): |
| 86 | + dataset = DatasetFromList(dataset, copy=False, serialize=True) |
| 87 | + if sampler is None: |
| 88 | + sampler = TrainingSampler(len(dataset)) |
| 89 | + |
| 90 | + self.dataset = dataset |
| 91 | + self.mapper = mapper |
| 92 | + self.sampler = sampler |
| 93 | + self.total_batch_size = total_batch_size |
| 94 | + self.num_workers = num_workers |
| 95 | + self.per_gpu_batch_size = self.total_batch_size // comm.get_world_size() |
| 96 | + |
| 97 | + self.max_time_seconds = max_time_seconds |
| 98 | + |
| 99 | + def _benchmark(self, iterator, num_iter, warmup, msg=None): |
| 100 | + avg, all_times = iter_benchmark(iterator, num_iter, warmup, self.max_time_seconds) |
| 101 | + if msg is not None: |
| 102 | + self._log_time(msg, avg, all_times) |
| 103 | + return avg, all_times |
| 104 | + |
| 105 | + def _log_time(self, msg, avg, all_times, distributed=False): |
| 106 | + percentiles = [np.percentile(all_times, k, interpolation="nearest") for k in [1, 5, 95, 99]] |
| 107 | + if not distributed: |
| 108 | + logger.info( |
| 109 | + f"{msg}: avg={1.0/avg:.1f} it/s, " |
| 110 | + f"p1={percentiles[0]:.2g}s, p5={percentiles[1]:.2g}s, " |
| 111 | + f"p95={percentiles[2]:.2g}s, p99={percentiles[3]:.2g}s." |
| 112 | + ) |
| 113 | + return |
| 114 | + avg_per_gpu = comm.all_gather(avg) |
| 115 | + percentiles_per_gpu = comm.all_gather(percentiles) |
| 116 | + if comm.get_rank() > 0: |
| 117 | + return |
| 118 | + for idx, avg, percentiles in zip(count(), avg_per_gpu, percentiles_per_gpu): |
| 119 | + logger.info( |
| 120 | + f"GPU{idx} {msg}: avg={1.0/avg:.1f} it/s, " |
| 121 | + f"p1={percentiles[0]:.2g}s, p5={percentiles[1]:.2g}s, " |
| 122 | + f"p95={percentiles[2]:.2g}s, p99={percentiles[3]:.2g}s." |
| 123 | + ) |
| 124 | + |
| 125 | + def benchmark_dataset(self, num_iter, warmup=5): |
| 126 | + """ |
| 127 | + Benchmark the speed of taking raw samples from the dataset. |
| 128 | + """ |
| 129 | + |
| 130 | + def loader(): |
| 131 | + while True: |
| 132 | + for k in self.sampler: |
| 133 | + yield self.dataset[k] |
| 134 | + |
| 135 | + self._benchmark(loader(), num_iter, warmup, "Dataset Alone") |
| 136 | + |
| 137 | + def benchmark_mapper(self, num_iter, warmup=5): |
| 138 | + """ |
| 139 | + Benchmark the speed of taking raw samples from the dataset and map |
| 140 | + them in a single process. |
| 141 | + """ |
| 142 | + |
| 143 | + def loader(): |
| 144 | + while True: |
| 145 | + for k in self.sampler: |
| 146 | + yield self.mapper(self.dataset[k]) |
| 147 | + |
| 148 | + self._benchmark(loader(), num_iter, warmup, "Single Process Mapper (sec/sample)") |
| 149 | + |
| 150 | + def benchmark_workers(self, num_iter, warmup=10): |
| 151 | + """ |
| 152 | + Benchmark the dataloader by tuning num_workers to [0, 1, self.num_workers]. |
| 153 | + """ |
| 154 | + candidates = [0, 1] |
| 155 | + if self.num_workers not in candidates: |
| 156 | + candidates.append(self.num_workers) |
| 157 | + |
| 158 | + dataset = MapDataset(self.dataset, self.mapper) |
| 159 | + for n in candidates: |
| 160 | + loader = build_batch_data_loader( |
| 161 | + dataset, |
| 162 | + self.sampler, |
| 163 | + self.total_batch_size, |
| 164 | + num_workers=n, |
| 165 | + ) |
| 166 | + self._benchmark( |
| 167 | + iter(loader), |
| 168 | + num_iter * max(n, 1), |
| 169 | + warmup * max(n, 1), |
| 170 | + f"DataLoader ({n} workers, bs={self.per_gpu_batch_size})", |
| 171 | + ) |
| 172 | + del loader |
| 173 | + |
| 174 | + def benchmark_IPC(self, num_iter, warmup=10): |
| 175 | + """ |
| 176 | + Benchmark the dataloader where each worker outputs nothing. This |
| 177 | + eliminates the IPC overhead compared to the regular dataloader. |
| 178 | +
|
| 179 | + PyTorch multiprocessing's IPC only optimizes for torch tensors. |
| 180 | + Large numpy arrays or other data structure may incur large IPC overhead. |
| 181 | + """ |
| 182 | + n = self.num_workers |
| 183 | + dataset = _EmptyMapDataset(MapDataset(self.dataset, self.mapper)) |
| 184 | + loader = build_batch_data_loader( |
| 185 | + dataset, self.sampler, self.total_batch_size, num_workers=n |
| 186 | + ) |
| 187 | + self._benchmark( |
| 188 | + iter(loader), |
| 189 | + num_iter * max(n, 1), |
| 190 | + warmup * max(n, 1), |
| 191 | + f"DataLoader ({n} workers, bs={self.per_gpu_batch_size}) w/o comm", |
| 192 | + ) |
| 193 | + |
| 194 | + def benchmark_distributed(self, num_iter, warmup=10): |
| 195 | + """ |
| 196 | + Benchmark the dataloader in each distributed worker, and log results of |
| 197 | + all workers. This helps understand the final performance as well as |
| 198 | + the variances among workers. |
| 199 | +
|
| 200 | + It also prints startup time (first iter) of the dataloader. |
| 201 | + """ |
| 202 | + gpu = comm.get_world_size() |
| 203 | + dataset = MapDataset(self.dataset, self.mapper) |
| 204 | + n = self.num_workers |
| 205 | + loader = build_batch_data_loader( |
| 206 | + dataset, self.sampler, self.total_batch_size, num_workers=n |
| 207 | + ) |
| 208 | + |
| 209 | + timer = Timer() |
| 210 | + loader = iter(loader) |
| 211 | + next(loader) |
| 212 | + startup_time = timer.seconds() |
| 213 | + logger.info("Dataloader startup time: {:.2f} seconds".format(startup_time)) |
| 214 | + |
| 215 | + comm.synchronize() |
| 216 | + |
| 217 | + avg, all_times = self._benchmark(loader, num_iter * max(n, 1), warmup * max(n, 1)) |
| 218 | + del loader |
| 219 | + self._log_time( |
| 220 | + f"DataLoader ({gpu} GPUs x {n} workers, total bs={self.total_batch_size})", |
| 221 | + avg, |
| 222 | + all_times, |
| 223 | + True, |
| 224 | + ) |
0 commit comments