Arulkumar03/Wheat_HEAD_Detection_Counting_ComputerVision_Model

1# Copyright (c) Facebook, Inc. and its affiliates.2import logging3import numpy as np4from itertools import count5from typing import List, Tuple6import torch7import tqdm8from fvcore.common.timer import Timer9 10from detectron2.utils import comm11 12from .build import build_batch_data_loader13from .common import DatasetFromList, MapDataset14from .samplers import TrainingSampler15 16logger = logging.getLogger(__name__)17 18 19class _EmptyMapDataset(torch.utils.data.Dataset):20    """21    Map anything to emptiness.22    """23 24    def __init__(self, dataset):25        self.ds = dataset26 27    def __len__(self):28        return len(self.ds)29 30    def __getitem__(self, idx):31        _ = self.ds[idx]32        return [0]33 34 35def iter_benchmark(36    iterator, num_iter: int, warmup: int = 5, max_time_seconds: float = 6037) -> Tuple[float, List[float]]:38    """39    Benchmark an iterator/iterable for `num_iter` iterations with an extra40    `warmup` iterations of warmup.41    End early if `max_time_seconds` time is spent on iterations.42 43    Returns:44        float: average time (seconds) per iteration45        list[float]: time spent on each iteration. Sometimes useful for further analysis.46    """47    num_iter, warmup = int(num_iter), int(warmup)48 49    iterator = iter(iterator)50    for _ in range(warmup):51        next(iterator)52    timer = Timer()53    all_times = []54    for curr_iter in tqdm.trange(num_iter):55        start = timer.seconds()56        if start > max_time_seconds:57            num_iter = curr_iter58            break59        next(iterator)60        all_times.append(timer.seconds() - start)61    avg = timer.seconds() / num_iter62    return avg, all_times63 64 65class DataLoaderBenchmark:66    """67    Some common benchmarks that help understand perf bottleneck of a standard dataloader68    made of dataset, mapper and sampler.69    """70 71    def __init__(72        self,73        dataset,74        *,75        mapper,76        sampler=None,77        total_batch_size,78        num_workers=0,79        max_time_seconds: int = 90,80    ):81        """82        Args:83            max_time_seconds (int): maximum time to spent for each benchmark84            other args: same as in `build.py:build_detection_train_loader`85        """86        if isinstance(dataset, list):87            dataset = DatasetFromList(dataset, copy=False, serialize=True)88        if sampler is None:89            sampler = TrainingSampler(len(dataset))90 91        self.dataset = dataset92        self.mapper = mapper93        self.sampler = sampler94        self.total_batch_size = total_batch_size95        self.num_workers = num_workers96        self.per_gpu_batch_size = self.total_batch_size // comm.get_world_size()97 98        self.max_time_seconds = max_time_seconds99 100    def _benchmark(self, iterator, num_iter, warmup, msg=None):101        avg, all_times = iter_benchmark(iterator, num_iter, warmup, self.max_time_seconds)102        if msg is not None:103            self._log_time(msg, avg, all_times)104        return avg, all_times105 106    def _log_time(self, msg, avg, all_times, distributed=False):107        percentiles = [np.percentile(all_times, k, interpolation="nearest") for k in [1, 5, 95, 99]]108        if not distributed:109            logger.info(110                f"{msg}: avg={1.0/avg:.1f} it/s, "111                f"p1={percentiles[0]:.2g}s, p5={percentiles[1]:.2g}s, "112                f"p95={percentiles[2]:.2g}s, p99={percentiles[3]:.2g}s."113            )114            return115        avg_per_gpu = comm.all_gather(avg)116        percentiles_per_gpu = comm.all_gather(percentiles)117        if comm.get_rank() > 0:118            return119        for idx, avg, percentiles in zip(count(), avg_per_gpu, percentiles_per_gpu):120            logger.info(121                f"GPU{idx} {msg}: avg={1.0/avg:.1f} it/s, "122                f"p1={percentiles[0]:.2g}s, p5={percentiles[1]:.2g}s, "123                f"p95={percentiles[2]:.2g}s, p99={percentiles[3]:.2g}s."124            )125 126    def benchmark_dataset(self, num_iter, warmup=5):127        """128        Benchmark the speed of taking raw samples from the dataset.129        """130 131        def loader():132            while True:133                for k in self.sampler:134                    yield self.dataset[k]135 136        self._benchmark(loader(), num_iter, warmup, "Dataset Alone")137 138    def benchmark_mapper(self, num_iter, warmup=5):139        """140        Benchmark the speed of taking raw samples from the dataset and map141        them in a single process.142        """143 144        def loader():145            while True:146                for k in self.sampler:147                    yield self.mapper(self.dataset[k])148 149        self._benchmark(loader(), num_iter, warmup, "Single Process Mapper (sec/sample)")150 151    def benchmark_workers(self, num_iter, warmup=10):152        """153        Benchmark the dataloader by tuning num_workers to [0, 1, self.num_workers].154        """155        candidates = [0, 1]156        if self.num_workers not in candidates:157            candidates.append(self.num_workers)158 159        dataset = MapDataset(self.dataset, self.mapper)160        for n in candidates:161            loader = build_batch_data_loader(162                dataset,163                self.sampler,164                self.total_batch_size,165                num_workers=n,166            )167            self._benchmark(168                iter(loader),169                num_iter * max(n, 1),170                warmup * max(n, 1),171                f"DataLoader ({n} workers, bs={self.per_gpu_batch_size})",172            )173            del loader174 175    def benchmark_IPC(self, num_iter, warmup=10):176        """177        Benchmark the dataloader where each worker outputs nothing. This178        eliminates the IPC overhead compared to the regular dataloader.179 180        PyTorch multiprocessing's IPC only optimizes for torch tensors.181        Large numpy arrays or other data structure may incur large IPC overhead.182        """183        n = self.num_workers184        dataset = _EmptyMapDataset(MapDataset(self.dataset, self.mapper))185        loader = build_batch_data_loader(186            dataset, self.sampler, self.total_batch_size, num_workers=n187        )188        self._benchmark(189            iter(loader),190            num_iter * max(n, 1),191            warmup * max(n, 1),192            f"DataLoader ({n} workers, bs={self.per_gpu_batch_size}) w/o comm",193        )194 195    def benchmark_distributed(self, num_iter, warmup=10):196        """197        Benchmark the dataloader in each distributed worker, and log results of198        all workers. This helps understand the final performance as well as199        the variances among workers.200 201        It also prints startup time (first iter) of the dataloader.202        """203        gpu = comm.get_world_size()204        dataset = MapDataset(self.dataset, self.mapper)205        n = self.num_workers206        loader = build_batch_data_loader(207            dataset, self.sampler, self.total_batch_size, num_workers=n208        )209 210        timer = Timer()211        loader = iter(loader)212        next(loader)213        startup_time = timer.seconds()214        logger.info("Dataloader startup time: {:.2f} seconds".format(startup_time))215 216        comm.synchronize()217 218        avg, all_times = self._benchmark(loader, num_iter * max(n, 1), warmup * max(n, 1))219        del loader220        self._log_time(221            f"DataLoader ({gpu} GPUs x {n} workers, total bs={self.total_batch_size})",222            avg,223            all_times,224            True,225        )226