RuntimeError: Trying to backward through the graph a second time

[FrancescoSaverioZuppichini]

Thanks for reporting the unexpected results and we appreciate it a lot.

Describe the Issue
GradientCumulativeOptimizerHook doesn't work.

Reproduction

1. What command, code, or script did you run?
   Add GradientCumulativeOptimizerHook to your *_config.py file

custom_hooks = [
    dict(type="GradientCumulativeOptimizerHook", cumulative_iters=4),
]

Output

2021-09-27 14:21:26,889 - mmdet - WARNING - GradientCumulativeOptimizerHook may slightly decrease performance if the model has BatchNorm layers.
Traceback (most recent call last):
  File "/home/zuppif/integration-object-detection/playground.py", line 81, in <module>
    main(Args(config_file, cfg_options=options))
  File "/home/zuppif/integration-object-detection/src/train.py", line 185, in main
    train_detector(
  File "/home/zuppif/integration-object-detection/.venv/lib/python3.9/site-packages/mmdet/apis/train.py", line 174, in train_detector
    runner.run(data_loaders, cfg.workflow)
  File "/home/zuppif/integration-object-detection/.venv/lib/python3.9/site-packages/mmcv/runner/epoch_based_runner.py", line 127, in run
    epoch_runner(data_loaders[i], **kwargs)
  File "/home/zuppif/integration-object-detection/.venv/lib/python3.9/site-packages/mmcv/runner/epoch_based_runner.py", line 51, in train
    self.call_hook('after_train_iter')
  File "/home/zuppif/integration-object-detection/.venv/lib/python3.9/site-packages/mmcv/runner/base_runner.py", line 307, in call_hook
    getattr(hook, fn_name)(self)
  File "/home/zuppif/integration-object-detection/.venv/lib/python3.9/site-packages/mmcv/runner/hooks/optimizer.py", line 115, in after_train_iter
    loss.backward()
  File "/home/zuppif/integration-object-detection/.venv/lib/python3.9/site-packages/torch/_tensor.py", line 255, in backward
    torch.autograd.backward(self, gradient, retain_graph, create_graph, inputs=inputs)
  File "/home/zuppif/integration-object-detection/.venv/lib/python3.9/site-packages/torch/autograd/__init__.py", line 147, in backward
    Variable._execution_engine.run_backward(
RuntimeError: Trying to backward through the graph a second time (or directly access saved variables after they have already been freed). Saved intermediate values of the graph are freed when you call .backward() or autograd.grad(). Specify retain_graph=True if you need to backward through the graph a second time or if you need to access saved variables after calling backward.

Environment

1. Please run python -c "from mmcv.utils import collect_env; print(collect_env())"

{'sys.platform': 'linux', 'Python': '3.9.5 (default, Jun  4 2021, 12:28:51) [GCC 7.5.0]', 'CUDA available': True, 'GPU 0,1,2': 'GeForce GTX 1080 Ti', 'CUDA_HOME': '/usr/local/cuda', 'NVCC': 'Build cuda_11.2.r11.2/compiler.29373293_0', 'GCC': 'gcc (Ubuntu 9.3.0-17ubuntu1~20.04) 9.3.0', 'PyTorch': '1.9.1+cu102', 'PyTorch compiling details': 'PyTorch built with:\n  - GCC 7.3\n  - C++ Version: 201402\n  - Intel(R) Math Kernel Library Version 2020.0.0 Product Build 20191122 for Intel(R) 64 architecture applications\n  - Intel(R) MKL-DNN v2.1.2 (Git Hash 98be7e8afa711dc9b66c8ff3504129cb82013cdb)\n  - OpenMP 201511 (a.k.a. OpenMP 4.5)\n  - NNPACK is enabled\n  - CPU capability usage: AVX2\n  - CUDA Runtime 10.2\n  - 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\n  - CuDNN 7.6.5\n  - Magma 2.5.2\n  - Build settings: BLAS_INFO=mkl, BUILD_TYPE=Release, CUDA_VERSION=10.2, CUDNN_VERSION=7.6.5, CXX_COMPILER=/opt/rh/devtoolset-7/root/usr/bin/c++, CXX_FLAGS= -Wno-deprecated -fvisibility-inlines-hidden -DUSE_PTHREADPOOL -fopenmp -DNDEBUG -DUSE_KINETO -DUSE_FBGEMM -DUSE_QNNPACK -DUSE_PYTORCH_QNNPACK -DUSE_XNNPACK -DSYMBOLICATE_MOBILE_DEBUG_HANDLE -O2 -fPIC -Wno-narrowing -Wall -Wextra -Werror=return-type -Wno-missing-field-initializers -Wno-type-limits -Wno-array-bounds -Wno-unknown-pragmas -Wno-sign-compare -Wno-unused-parameter -Wno-unused-variable -Wno-unused-function -Wno-unused-result -Wno-unused-local-typedefs -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 -Wno-stringop-overflow, LAPACK_INFO=mkl, PERF_WITH_AVX=1, PERF_WITH_AVX2=1, PERF_WITH_AVX512=1, TORCH_VERSION=1.9.1, 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=ON, USE_NNPACK=ON, USE_OPENMP=ON, \n', 'TorchVision': '0.10.1+cu102', 'OpenCV': '4.5.3', 'MMCV': '1.3.13', 'MMCV Compiler': 'GCC 9.3', 'MMCV CUDA Compiler': '11.2'}

Error traceback
If applicable, paste the error traceback here.

2021-09-27 14:21:26,889 - mmdet - WARNING - GradientCumulativeOptimizerHook may slightly decrease performance if the model has BatchNorm layers.
Traceback (most recent call last):
  File "/home/zuppif/integration-object-detection/playground.py", line 81, in <module>
    main(Args(config_file, cfg_options=options))
  File "/home/zuppif/integration-object-detection/src/train.py", line 185, in main
    train_detector(
  File "/home/zuppif/integration-object-detection/.venv/lib/python3.9/site-packages/mmdet/apis/train.py", line 174, in train_detector
    runner.run(data_loaders, cfg.workflow)
  File "/home/zuppif/integration-object-detection/.venv/lib/python3.9/site-packages/mmcv/runner/epoch_based_runner.py", line 127, in run
    epoch_runner(data_loaders[i], **kwargs)
  File "/home/zuppif/integration-object-detection/.venv/lib/python3.9/site-packages/mmcv/runner/epoch_based_runner.py", line 51, in train
    self.call_hook('after_train_iter')
  File "/home/zuppif/integration-object-detection/.venv/lib/python3.9/site-packages/mmcv/runner/base_runner.py", line 307, in call_hook
    getattr(hook, fn_name)(self)
  File "/home/zuppif/integration-object-detection/.venv/lib/python3.9/site-packages/mmcv/runner/hooks/optimizer.py", line 115, in after_train_iter
    loss.backward()
  File "/home/zuppif/integration-object-detection/.venv/lib/python3.9/site-packages/torch/_tensor.py", line 255, in backward
    torch.autograd.backward(self, gradient, retain_graph, create_graph, inputs=inputs)
  File "/home/zuppif/integration-object-detection/.venv/lib/python3.9/site-packages/torch/autograd/__init__.py", line 147, in backward
    Variable._execution_engine.run_backward(
RuntimeError: Trying to backward through the graph a second time (or directly access saved variables after they have already been freed). Saved intermediate values of the graph are freed when you call .backward() or autograd.grad(). Specify retain_graph=True if you need to backward through the graph a second time or if you need to access saved variables after calling backward.

[zhouzaida]

Hi, @FrancescoSaverioZuppichini , GradientCumulativeOptimizerHook and OptimizerHook are both set so RuntimeError was raised. OptimizerHook is set here

You can modify the config like this

optimizer_config = dict(type="GradientCumulativeOptimizerHook", cumulative_iters=4)

[FrancescoSaverioZuppichini]

Super! Thank you. @zhouzaida would it make sense to add an example in the doc to avoid this confusion?

[zhouzaida]

yet, great suggestion. We have a plan to add more examples or tutorials about how to use all of the hooks like OptimizerHook or GradientCumulativeOptimizerHook

[xiviax]

Hi @zhouzaida

I replaced optimizer_config = dict(grad_clip=None) in the linked file with
optimizer_config = dict(type="GradientCumulativeOptimizerHook", cumulative_iters=4) as mentioned above.

However, I obtain the following error:

How can I change the Hook from the OptimizerHook to the GradientCumulativeOptimizerHook correctly?

Thanks already in advance for your help!

[zhouzaida]

Hi @xiviax , could you provide the versions of mmcv and mmdet?

[xiviax]

Hi @zhouzaida
Thanks for your fast reply! :)
I am using mmcv version 1.6.1 and for mmdet I am cloning the mmdetection repository to colab - so version 2.25.1.

[xiviax]

These are the hooks that are being executed when I am not trying to use the GradientCumulativeOptimizerHook:

[zhouzaida]

It seems like you set the fp16 like here so optimizer_config = dict(type="GradientCumulativeOptimizerHook", cumulative_iters=4) is invalid. Could you provide the config you used?

[xiviax]

Ah I see! This is the config that I used:

#base config file
_base_='/content/mmdetection/configs/swin/mask_rcnn_swin-s-p4-w7_fpn_fp16_ms-crop-3x_coco.py'

#declare self-customed dataset type
dataset_type='CocoDataset'

CATEGORY = 'bottletype'
BASE_WORK_DIR = '/content/drive/MyDrive/Uni Materialien/Master Thesis/PoC no label/Output Bottletype'

#images directory
img_dir_synthetic='/content/drive/MyDrive/Uni Materialien/Master Thesis/Ubuntu_Master_Thesis/Datasets/synthetic/50_bottles_randomized_labels/images'
img_dir_real='/content/drive/MyDrive/Uni Materialien/Master Thesis/Ubuntu_Master_Thesis/Datasets/real/2022_08_11-POC_bottles_conveyor/images'
img_dir_real_test='/content/drive/MyDrive/Uni Materialien/Master Thesis/Ubuntu_Master_Thesis/Datasets/real/2022_08_23-POC_testset/images'

#classes list
classes = ("beverage PET", "milk & dairy", "food", "beauty care", "home care",)

#model configuration
model = dict(
    roi_head=dict(
        bbox_head=dict(num_classes=5),#change bbox and mask head numbers here.
        mask_head=dict(num_classes=5)))

# data setup
DATA_BASE_PATH_SYNTHETIC = '/content/drive/MyDrive/Uni Materialien/Master Thesis/Ubuntu_Master_Thesis/Datasets/synthetic/50_bottles_randomized_labels/coco_annotations'
DATA_BASE_PATH_REAL = '/content/drive/MyDrive/Uni Materialien/Master Thesis/Ubuntu_Master_Thesis/Datasets/real/2022_08_11-POC_bottles_conveyor/coco_annotations'
DATA_BASE_PATH_REAL_TEST = '/content/drive/MyDrive/Uni Materialien/Master Thesis/Ubuntu_Master_Thesis/Datasets/real/2022_08_23-POC_testset/coco_annotations'

data=dict(
  
    train=dict(
        type=dataset_type,
        classes=classes,
        ann_file=[
            f'{DATA_BASE_PATH_SYNTHETIC}/plastic_bottles_annotations_{CATEGORY}_train.json',
        ],
        img_prefix=img_dir_synthetic
    ),

    val=dict(
        type=dataset_type,
        classes=classes,
        ann_file=[
            f'{DATA_BASE_PATH_SYNTHETIC}/plastic_bottles_annotations_{CATEGORY}_val.json',
        ],
        img_prefix=img_dir_synthetic
    ),

    multi_val={
        'real_valset':dict(
            type=dataset_type,
            classes=classes,
            ann_file=[
                f'{DATA_BASE_PATH_REAL}/plastic_bottles_annotations_{CATEGORY}_val.json',
            ],
            img_prefix=img_dir_real
        ),
        'real_testset':dict(
            type=dataset_type,
            classes=classes,
            ann_file=[
                f'{DATA_BASE_PATH_REAL_TEST}/plastic_bottles_annotations_{CATEGORY}_val.json',
            ],
            img_prefix=img_dir_real_test
        )
    },

    test=dict(
        type=dataset_type,
        classes=classes,
        ann_file=[
            f'{DATA_BASE_PATH_REAL}/plastic_bottles_annotations_{CATEGORY}_val.json',#validate file path
        ],
        img_prefix=img_dir_real

    )
)

#optimizer_config = dict(type="GradientCumulativeOptimizerHook", cumulative_iters=4)

#pretrained model
#load_from=f'{BASE_WORK_DIR}//synthetic/epoch_30.pth'

#resume training
#resume_from=f'{BASE_WORK_DIR}/synthetic/epoch_23.pth'

#where to save checkpoint files
work_dir = f'{BASE_WORK_DIR}/synthetic-bs8-hook'

checkpoint_config = dict(
    interval = 2,
    max_keep_ckpts = 1
)

I just saw that there is also a GradientCumulativeFp16OptimizerHook. I tried to put it in my config file the same way aas for the GradientCumulativeOptimizerHook , however I am getting the same error:

Traceback (most recent call last):
  File "/content/mmdetection/tools/train.py", line 242, in <module>
    main()
  File "/content/mmdetection/tools/train.py", line 238, in main
    meta=meta)
  File "/content/mmdetection/mmdet/apis/train.py", line 186, in train_detector
    **cfg.optimizer_config, **fp16_cfg, distributed=distributed)
TypeError: __init__() got an unexpected keyword argument 'type'

Could the GradientCumulativeFp16OptimizerHook work in my case? If yes, how can it be implemented?

Thanks already in advance for your answer!

[zhouzaida]

Hi, you need to remove this line from config.

[xiviax]

Great. It seems to be working now. Thank you so much for your support!

[kazuma-jp]

Hi @zhouzaida.
Thank you for your description of GradientCumulativeOptimizerHook.
Can GradientCumulative be used in the fp16 setting?

I run mask r cnn on following config.But, doesn't work.

fp16 = dict(loss_scale=512.)

optimizer_config = dict(
    type="GradientCumulativeFp16OptimizerHook", cumulative_iters=4)

Detailed settings can be found here.