[Bug] Loading dataset takes really long

Prerequisite

[X] I have searched Issues and Discussions but cannot get the expected help.
[X] I have read the FAQ documentation but cannot get the expected help.
[X] The bug has not been fixed in the latest version (dev-1.x) or latest version (dev-1.0).

Task

I'm using the official example scripts/configs for the officially supported tasks/models/datasets.

Branch

main branch https://github.com/open-mmlab/mmdetection3d

Environment

sys.platform: linux Python: 3.8.17 (default, Jul 5 2023, 21:04:15) [GCC 11.2.0] CUDA available: True numpy_random_seed: 2147483648 GPU 0,1: NVIDIA TITAN RTX CUDA_HOME: /usr/local/cuda NVCC: Cuda compilation tools, release 12.1, V12.1.105 GCC: gcc (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0 PyTorch: 2.0.1+cu118 PyTorch compiling details: PyTorch built with:

GCC 9.3
C++ Version: 201703
Intel(R) oneAPI Math Kernel Library Version 2023.1-Product Build 20230303 for Intel(R) 64 architecture applications
Intel(R) MKL-DNN v2.7.3 (Git Hash 6dbeffbae1f23cbbeae17adb7b5b13f1f37c080e)
OpenMP 201511 (a.k.a. OpenMP 4.5)
LAPACK is enabled (usually provided by MKL)
NNPACK is enabled
CPU capability usage: AVX2
CUDA Runtime 11.8
NVCC architecture flags: -gencode;arch=compute_37,code=sm_37;-gencode;arch=compute_50,code=sm_50;-gencode;arch=compute_60,code=sm_60;-gencode;arch=compute_70,code=sm_70;-gencode;arch=compute_75,code=sm_75;-gencode;arch=compute_80,code=sm_80;-gencode;arch=compute_86,code=sm_86;-gencode;arch=compute_90,code=sm_90
CuDNN 8.7
Magma 2.6.1
Build settings: BLAS_INFO=mkl, BUILD_TYPE=Release, CUDA_VERSION=11.8, CUDNN_VERSION=8.7.0, CXX_COMPILER=/opt/rh/devtoolset-9/root/usr/bin/c++, CXX_FLAGS= -D_GLIBCXX_USE_CXX11_ABI=0 -fabi-version=11 -Wno-deprecated -fvisibility-inlines-hidden -DUSE_PTHREADPOOL -DNDEBUG -DUSE_KINETO -DLIBKINETO_NOROCTRACER -DUSE_FBGEMM -DUSE_QNNPACK -DUSE_PYTORCH_QNNPACK -DUSE_XNNPACK -DSYMBOLICATE_MOBILE_DEBUG_HANDLE -O2 -fPIC -Wall -Wextra -Werror=return-type -Werror=non-virtual-dtor -Werror=bool-operation -Wnarrowing -Wno-missing-field-initializers -Wno-type-limits -Wno-array-bounds -Wno-unknown-pragmas -Wunused-local-typedefs -Wno-unused-parameter -Wno-unused-function -Wno-unused-result -Wno-strict-overflow -Wno-strict-aliasing -Wno-error=deprecated-declarations -Wno-stringop-overflow -Wno-psabi -Wno-error=pedantic -Wno-error=redundant-decls -Wno-error=old-style-cast -fdiagnostics-color=always -faligned-new -Wno-unused-but-set-variable -Wno-maybe-uninitialized -fno-math-errno -fno-trapping-math -Werror=format -Werror=cast-function-type -Wno-stringop-overflow, LAPACK_INFO=mkl, PERF_WITH_AVX=1, PERF_WITH_AVX2=1, PERF_WITH_AVX512=1, TORCH_DISABLE_GPU_ASSERTS=ON, TORCH_VERSION=2.0.1, USE_CUDA=ON, USE_CUDNN=ON, USE_EXCEPTION_PTR=1, USE_GFLAGS=OFF, USE_GLOG=OFF, USE_MKL=ON, USE_MKLDNN=ON, USE_MPI=OFF, USE_NCCL=1, USE_NNPACK=ON, USE_OPENMP=ON, USE_ROCM=OFF,

OpenCV: 4.8.0 MMEngine: 0.8.4 MMDetection: 3.1.0 MMDetection3D: 1.2.0+9f40d4f spconv2.0: False

Reproduces the problem - code sample

import timeit, functools, copy, pickle

if __name__ == '__main__':
    class MyDataset:
        def __init__(self):
            self._metainfo = dict(
                classes=['car', 'truck', 'trailer', 'bus', 'construction_vehicle', 'bicycle',
                                           'motorcycle', 'pedestrian', 'traffic_cone', 'barrier'],
                version='v1.0-trainval',
                palette=[
                    (255, 158, 0),  # Orange
                    (255, 99, 71),  # Tomato
                    (255, 140, 0),  # Darkorange
                    (255, 127, 80),  # Coral
                    (233, 150, 70),  # Darksalmon
                    (220, 20, 60),  # Crimson
                    (255, 61, 99),  # Red
                    (0, 0, 230),  # Blue
                    (47, 79, 79),  # Darkslategrey
                    (112, 128, 144),  # Slategrey
                ])

        @property
        def metainfo(self):
            return copy.deepcopy(self._metainfo)

    def deepcopy(dataset):
        classes = dataset.metainfo['classes']

    def directly(dataset):
        classes = dataset._metainfo['classes']

    dataset = MyDataset()
    t = timeit.Timer(functools.partial(deepcopy, dataset))
    print(f'metainfo deepcopy: {t.timeit()}')
    t = timeit.Timer(functools.partial(directly, dataset))
    print(f'metainfo directly: {t.timeit()}')

    def deepcopy(dataset):
        dc = copy.deepcopy(dataset._metainfo)

    def _pickle(dataset):
        dc = pickle.loads(pickle.dumps(dataset._metainfo))

    dataset = MyDataset()
    t = timeit.Timer(functools.partial(deepcopy, dataset))
    print(f'Data deepcopy: {t.timeit()}')
    t = timeit.Timer(functools.partial(_pickle, dataset))
    print(f'Data pickle: {t.timeit()}')

Reproduces the problem - command or script

Reproduces the problem - error message

Additional information

Loading the dataset either during training/testing or running browse_dataset takes a really long time for large datasets, such as nuScenes.

I found out that it is spending a lot of time here: https://github.com/open-mmlab/mmdetection3d/blob/0f9dfa97a35ef87e16b700742d3c358d0ad15452/mmdet3d/datasets/det3d_dataset.py#L259 This is called for all instances in each frame and since accessing metadata always does a deepcopy of _metainfo it takes a lot of time. To see how slow deepcopy is I did some experimentation and found out, that it is significantly slower than directly accessing the elements:

metainfo deepcopy: 46.957518891999825
metainfo directly: 0.11945905200082052

I think directly accessing self._metainfo['classes'] could help improving loading times.

Another bottleneck is here: https://github.com/open-mmlab/mmdetection3d/blob/0f9dfa97a35ef87e16b700742d3c358d0ad15452/mmdet3d/datasets/det3d_dataset.py#L374 also caused by a deepcopy. Using input_dict = pickle.loads(pickle.dumps(ori_input_dict)) should also help reducing the runtime:

Data deepcopy: 49.796394890000556
Data pickle: 6.8219003000003795

[UPDATE] The deepcopy isn't actually required anymore, since the deepcopy is already performed by the base dataset https://github.com/open-mmlab/mmengine/pull/471/files

open-mmlab / mmdetection3d