Add Cross-Iteration Batch Normalization and Accumulate Gradient

[iumyx2612]

Describe the feature Add Cross-Iteration Batch Normalization in: https://arxiv.org/abs/2002.05712
And Accumulate Gradient for training: https://github.com/WongKinYiu/ScaledYOLOv4/blob/yolov4-large/train.py#L77

Cross-Iteration BN helps model with small batch-size to achieve better results.
And Accumulate Gradient helps me compares to other papers when batch-size is not the same. Motivation A clear and concise description of the motivation of the feature. I don't have enough computation power to train with a big enough batch-size, and it's really hard to compare results to other papers when batch-size is not the same

Related resources If there is an official code release or third-party implementations, please also provide the information here, which would be very helpful. Cross-Iteration BN: https://github.com/Howal/Cross-iterationBatchNorm
Accumulate Gradient: https://github.com/WongKinYiu/ScaledYOLOv4/blob/yolov4-large/train.py#L77

Additional context Add any other context or screenshots about the feature request here. If you would like to implement the feature and create a PR, please leave a comment here and that would be much appreciated.

[HAOCHENYE]

Hi~ thanks for your suggestions, Gradient accumulative has been implemented here https://github.com/open-mmlab/mmcv/blob/1f2500102834a01b86bf9ae4db227cd8d724fa6e/mmcv/runner/hooks/optimizer.py#L99

I think it is a good idea to add Cross-Iteration Batch Normalization into NORM_LAYERS。

[iumyx2612]

Hi~ thanks for your suggestions, Gradient accumulative has been implemented here

https://github.com/open-mmlab/mmcv/blob/1f2500102834a01b86bf9ae4db227cd8d724fa6e/mmcv/runner/hooks/optimizer.py#L99

I think it is a good idea to add Cross-Iteration Batch Normalization into NORM_LAYERS。

I used GradientCumulativeOptimizerHook get this stack trace

2022-06-26 09:33:16,631 - mmseg - WARNING - GradientCumulativeOptimizerHook may slightly decrease performance if the model has BatchNorm layers.
Traceback (most recent call last):
  File "E:/Work work/Python/Work/Practice/Segmentation/mmsegmentation/tools/train.py", line 242, in <module>
    main()
  File "E:/Work work/Python/Work/Practice/Segmentation/mmsegmentation/tools/train.py", line 231, in main
    train_segmentor(
  File "E:\Work work\Python\Work\Practice\Segmentation\mmsegmentation\mmseg\apis\train.py", line 194, in train_segmentor
    runner.run(data_loaders, cfg.workflow)
  File "E:\Anaconda\envs\openmmlab\lib\site-packages\mmcv\runner\iter_based_runner.py", line 135, in run
    iter_runner(iter_loaders[i], **kwargs)
  File "E:\Anaconda\envs\openmmlab\lib\site-packages\mmcv\runner\iter_based_runner.py", line 68, in train
    self.call_hook('after_train_iter')
  File "E:\Anaconda\envs\openmmlab\lib\site-packages\mmcv\runner\base_runner.py", line 309, in call_hook
    getattr(hook, fn_name)(self)
  File "E:\Anaconda\envs\openmmlab\lib\site-packages\mmcv\runner\hooks\optimizer.py", line 163, in after_train_iter
    loss.backward()
  File "E:\Anaconda\envs\openmmlab\lib\site-packages\torch\_tensor.py", line 363, in backward
    torch.autograd.backward(self, gradient, retain_graph, create_graph, inputs=inputs)
  File "E:\Anaconda\envs\openmmlab\lib\site-packages\torch\autograd\__init__.py", line 173, in backward
    Variable._execution_engine.run_backward(  # Calls into the C++ engine to run the backward pass
RuntimeError: Trying to backward through the graph a second time (or directly access saved tensors 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 tensors after calling backward.

Here's my config file

_base_ = [
    '../_base_/datasets/ade20k.py',
    '../_base_/default_runtime.py',
    '../_base_/schedules/schedule_40k.py',
]

custom_imports = dict(imports=['mmcls.models'], allow_failed_imports=False)
pretrained =\
    "https://download.openmmlab.com/mmclassification/v0/efficientnet/efficientnet-b1_3rdparty_8xb32-aa-advprop_in1k_20220119-5715267d.pth"

model = dict(
    type='EncoderDecoder',
    backbone=dict(
        type='mmcls.EfficientNet',
        arch='b1',
        out_indices=(2, 3, 4, 5),
        init_cfg=dict(
            type='Pretrained',
            checkpoint=pretrained,
            prefix='backbone.'
        )
    ),
    neck=dict(
        type='FPN',
        in_channels=[24, 40, 112, 320],
        out_channels=256,
        num_outs=4
    ),
    decode_head=dict(
        type='FCNHead',
        in_channels=[256, 256, 256, 256],
        channels=128,
        num_classes=3,
        in_index=[0, 1, 2, 3],
        input_transform='resize_concat',
        concat_input=False,
        loss_decode=dict(
            type='FocalLoss',
            use_sigmoid=True
        )
    ),
)

# dataset settings
dataset_type = 'Secret'
data_root = '../Dataset'

img_norm_cfg = dict(
    mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
crop_size = (128, 128)
train_pipeline = [
    dict(type='LoadImageFromFile'),
    dict(type='LoadAnnotations', reduce_zero_label=False),
    dict(type='Resize', img_scale=crop_size, keep_ratio=False, ratio_range=(1, 1)),
    dict(type='RandomFlip', prob=0.5),
    #dict(type='PhotoMetricDistortion'),
    dict(type='Normalize', **img_norm_cfg),
    dict(type='DefaultFormatBundle'),
    dict(type='Collect', keys=['img', 'gt_semantic_seg']),
]
test_pipeline = [
    dict(type='LoadImageFromFile'),
    dict(
        type='MultiScaleFlipAug',
        img_scale=crop_size,
        # img_ratios=[0.5, 0.75, 1.0, 1.25, 1.5, 1.75],
        flip=False,
        transforms=[
            dict(type='Resize', keep_ratio=False),
            dict(type='RandomFlip'),
            dict(type='Normalize', **img_norm_cfg),
            dict(type='ImageToTensor', keys=['img']),
            dict(type='Collect', keys=['img']),
        ])
]


data = dict(
    samples_per_gpu=1,
    workers_per_gpu=2,
    train=dict(
        type=dataset_type,
        data_root=data_root,
        img_dir='train/train',
        ann_dir='train_seg_map',
        pipeline=train_pipeline
    ),
    val=dict(
        type=dataset_type,
        data_root=data_root,
        img_dir='val',
        ann_dir='val_seg_map',
        pipeline=test_pipeline
    ),
    test=dict(
        type=dataset_type,
        data_root=data_root,
        img_dir='val',
        ann_dir='val_seg_map',
        pipeline=test_pipeline
    )
)

checkpoint_config = dict(by_epoch=False, interval=500)
evaluation = dict(interval=500, metric='mIoU', pre_eval=True)

log_config = dict(
    interval=1,
    hooks=[
        dict(type='TextLoggerHook', by_epoch=False),
        # dict(type='TensorboardLoggerHook')
    ])

custom_hooks = [
    dict(
        type='GradientCumulativeOptimizerHook',
        cumulative_iters=2
    )
]

[HAOCHENYE]

Hi~ thanks for your suggestions, Gradient accumulative has been implemented here https://github.com/open-mmlab/mmcv/blob/1f2500102834a01b86bf9ae4db227cd8d724fa6e/mmcv/runner/hooks/optimizer.py#L99

I think it is a good idea to add Cross-Iteration Batch Normalization into NORM_LAYERS。

I used GradientCumulativeOptimizerHook get this stack trace

2022-06-26 09:33:16,631 - mmseg - WARNING - GradientCumulativeOptimizerHook may slightly decrease performance if the model has BatchNorm layers.
Traceback (most recent call last):
  File "E:/Work work/Python/Work/Practice/Segmentation/mmsegmentation/tools/train.py", line 242, in <module>
    main()
  File "E:/Work work/Python/Work/Practice/Segmentation/mmsegmentation/tools/train.py", line 231, in main
    train_segmentor(
  File "E:\Work work\Python\Work\Practice\Segmentation\mmsegmentation\mmseg\apis\train.py", line 194, in train_segmentor
    runner.run(data_loaders, cfg.workflow)
  File "E:\Anaconda\envs\openmmlab\lib\site-packages\mmcv\runner\iter_based_runner.py", line 135, in run
    iter_runner(iter_loaders[i], **kwargs)
  File "E:\Anaconda\envs\openmmlab\lib\site-packages\mmcv\runner\iter_based_runner.py", line 68, in train
    self.call_hook('after_train_iter')
  File "E:\Anaconda\envs\openmmlab\lib\site-packages\mmcv\runner\base_runner.py", line 309, in call_hook
    getattr(hook, fn_name)(self)
  File "E:\Anaconda\envs\openmmlab\lib\site-packages\mmcv\runner\hooks\optimizer.py", line 163, in after_train_iter
    loss.backward()
  File "E:\Anaconda\envs\openmmlab\lib\site-packages\torch\_tensor.py", line 363, in backward
    torch.autograd.backward(self, gradient, retain_graph, create_graph, inputs=inputs)
  File "E:\Anaconda\envs\openmmlab\lib\site-packages\torch\autograd\__init__.py", line 173, in backward
    Variable._execution_engine.run_backward(  # Calls into the C++ engine to run the backward pass
RuntimeError: Trying to backward through the graph a second time (or directly access saved tensors 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 tensors after calling backward.

Here's my config file

_base_ = [
    '../_base_/datasets/ade20k.py',
    '../_base_/default_runtime.py',
    '../_base_/schedules/schedule_40k.py',
]

custom_imports = dict(imports=['mmcls.models'], allow_failed_imports=False)
pretrained =\
    "https://download.openmmlab.com/mmclassification/v0/efficientnet/efficientnet-b1_3rdparty_8xb32-aa-advprop_in1k_20220119-5715267d.pth"

model = dict(
    type='EncoderDecoder',
    backbone=dict(
        type='mmcls.EfficientNet',
        arch='b1',
        out_indices=(2, 3, 4, 5),
        init_cfg=dict(
            type='Pretrained',
            checkpoint=pretrained,
            prefix='backbone.'
        )
    ),
    neck=dict(
        type='FPN',
        in_channels=[24, 40, 112, 320],
        out_channels=256,
        num_outs=4
    ),
    decode_head=dict(
        type='FCNHead',
        in_channels=[256, 256, 256, 256],
        channels=128,
        num_classes=3,
        in_index=[0, 1, 2, 3],
        input_transform='resize_concat',
        concat_input=False,
        loss_decode=dict(
            type='FocalLoss',
            use_sigmoid=True
        )
    ),
)

# dataset settings
dataset_type = 'Secret'
data_root = '../Dataset'

img_norm_cfg = dict(
    mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
crop_size = (128, 128)
train_pipeline = [
    dict(type='LoadImageFromFile'),
    dict(type='LoadAnnotations', reduce_zero_label=False),
    dict(type='Resize', img_scale=crop_size, keep_ratio=False, ratio_range=(1, 1)),
    dict(type='RandomFlip', prob=0.5),
    #dict(type='PhotoMetricDistortion'),
    dict(type='Normalize', **img_norm_cfg),
    dict(type='DefaultFormatBundle'),
    dict(type='Collect', keys=['img', 'gt_semantic_seg']),
]
test_pipeline = [
    dict(type='LoadImageFromFile'),
    dict(
        type='MultiScaleFlipAug',
        img_scale=crop_size,
        # img_ratios=[0.5, 0.75, 1.0, 1.25, 1.5, 1.75],
        flip=False,
        transforms=[
            dict(type='Resize', keep_ratio=False),
            dict(type='RandomFlip'),
            dict(type='Normalize', **img_norm_cfg),
            dict(type='ImageToTensor', keys=['img']),
            dict(type='Collect', keys=['img']),
        ])
]


data = dict(
    samples_per_gpu=1,
    workers_per_gpu=2,
    train=dict(
        type=dataset_type,
        data_root=data_root,
        img_dir='train/train',
        ann_dir='train_seg_map',
        pipeline=train_pipeline
    ),
    val=dict(
        type=dataset_type,
        data_root=data_root,
        img_dir='val',
        ann_dir='val_seg_map',
        pipeline=test_pipeline
    ),
    test=dict(
        type=dataset_type,
        data_root=data_root,
        img_dir='val',
        ann_dir='val_seg_map',
        pipeline=test_pipeline
    )
)

checkpoint_config = dict(by_epoch=False, interval=500)
evaluation = dict(interval=500, metric='mIoU', pre_eval=True)

log_config = dict(
    interval=1,
    hooks=[
        dict(type='TextLoggerHook', by_epoch=False),
        # dict(type='TensorboardLoggerHook')
    ])

custom_hooks = [
    dict(
        type='GradientCumulativeOptimizerHook',
        cumulative_iters=2
    )
]

Hi~ It seems that you use loss.backward() mannualy, and GradientOptimizerHook execute the backward the second time and raise the error. Do you execute loss.backward in model mannuly?

[iumyx2612]

Hi~ thanks for your suggestions, Gradient accumulative has been implemented here https://github.com/open-mmlab/mmcv/blob/1f2500102834a01b86bf9ae4db227cd8d724fa6e/mmcv/runner/hooks/optimizer.py#L99

I think it is a good idea to add Cross-Iteration Batch Normalization into NORM_LAYERS。

I used GradientCumulativeOptimizerHook get this stack trace

2022-06-26 09:33:16,631 - mmseg - WARNING - GradientCumulativeOptimizerHook may slightly decrease performance if the model has BatchNorm layers.
Traceback (most recent call last):
  File "E:/Work work/Python/Work/Practice/Segmentation/mmsegmentation/tools/train.py", line 242, in <module>
    main()
  File "E:/Work work/Python/Work/Practice/Segmentation/mmsegmentation/tools/train.py", line 231, in main
    train_segmentor(
  File "E:\Work work\Python\Work\Practice\Segmentation\mmsegmentation\mmseg\apis\train.py", line 194, in train_segmentor
    runner.run(data_loaders, cfg.workflow)
  File "E:\Anaconda\envs\openmmlab\lib\site-packages\mmcv\runner\iter_based_runner.py", line 135, in run
    iter_runner(iter_loaders[i], **kwargs)
  File "E:\Anaconda\envs\openmmlab\lib\site-packages\mmcv\runner\iter_based_runner.py", line 68, in train
    self.call_hook('after_train_iter')
  File "E:\Anaconda\envs\openmmlab\lib\site-packages\mmcv\runner\base_runner.py", line 309, in call_hook
    getattr(hook, fn_name)(self)
  File "E:\Anaconda\envs\openmmlab\lib\site-packages\mmcv\runner\hooks\optimizer.py", line 163, in after_train_iter
    loss.backward()
  File "E:\Anaconda\envs\openmmlab\lib\site-packages\torch\_tensor.py", line 363, in backward
    torch.autograd.backward(self, gradient, retain_graph, create_graph, inputs=inputs)
  File "E:\Anaconda\envs\openmmlab\lib\site-packages\torch\autograd\__init__.py", line 173, in backward
    Variable._execution_engine.run_backward(  # Calls into the C++ engine to run the backward pass
RuntimeError: Trying to backward through the graph a second time (or directly access saved tensors 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 tensors after calling backward.

Here's my config file

_base_ = [
    '../_base_/datasets/ade20k.py',
    '../_base_/default_runtime.py',
    '../_base_/schedules/schedule_40k.py',
]

custom_imports = dict(imports=['mmcls.models'], allow_failed_imports=False)
pretrained =\
    "https://download.openmmlab.com/mmclassification/v0/efficientnet/efficientnet-b1_3rdparty_8xb32-aa-advprop_in1k_20220119-5715267d.pth"

model = dict(
    type='EncoderDecoder',
    backbone=dict(
        type='mmcls.EfficientNet',
        arch='b1',
        out_indices=(2, 3, 4, 5),
        init_cfg=dict(
            type='Pretrained',
            checkpoint=pretrained,
            prefix='backbone.'
        )
    ),
    neck=dict(
        type='FPN',
        in_channels=[24, 40, 112, 320],
        out_channels=256,
        num_outs=4
    ),
    decode_head=dict(
        type='FCNHead',
        in_channels=[256, 256, 256, 256],
        channels=128,
        num_classes=3,
        in_index=[0, 1, 2, 3],
        input_transform='resize_concat',
        concat_input=False,
        loss_decode=dict(
            type='FocalLoss',
            use_sigmoid=True
        )
    ),
)

# dataset settings
dataset_type = 'Secret'
data_root = '../Dataset'

img_norm_cfg = dict(
    mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
crop_size = (128, 128)
train_pipeline = [
    dict(type='LoadImageFromFile'),
    dict(type='LoadAnnotations', reduce_zero_label=False),
    dict(type='Resize', img_scale=crop_size, keep_ratio=False, ratio_range=(1, 1)),
    dict(type='RandomFlip', prob=0.5),
    #dict(type='PhotoMetricDistortion'),
    dict(type='Normalize', **img_norm_cfg),
    dict(type='DefaultFormatBundle'),
    dict(type='Collect', keys=['img', 'gt_semantic_seg']),
]
test_pipeline = [
    dict(type='LoadImageFromFile'),
    dict(
        type='MultiScaleFlipAug',
        img_scale=crop_size,
        # img_ratios=[0.5, 0.75, 1.0, 1.25, 1.5, 1.75],
        flip=False,
        transforms=[
            dict(type='Resize', keep_ratio=False),
            dict(type='RandomFlip'),
            dict(type='Normalize', **img_norm_cfg),
            dict(type='ImageToTensor', keys=['img']),
            dict(type='Collect', keys=['img']),
        ])
]


data = dict(
    samples_per_gpu=1,
    workers_per_gpu=2,
    train=dict(
        type=dataset_type,
        data_root=data_root,
        img_dir='train/train',
        ann_dir='train_seg_map',
        pipeline=train_pipeline
    ),
    val=dict(
        type=dataset_type,
        data_root=data_root,
        img_dir='val',
        ann_dir='val_seg_map',
        pipeline=test_pipeline
    ),
    test=dict(
        type=dataset_type,
        data_root=data_root,
        img_dir='val',
        ann_dir='val_seg_map',
        pipeline=test_pipeline
    )
)

checkpoint_config = dict(by_epoch=False, interval=500)
evaluation = dict(interval=500, metric='mIoU', pre_eval=True)

log_config = dict(
    interval=1,
    hooks=[
        dict(type='TextLoggerHook', by_epoch=False),
        # dict(type='TensorboardLoggerHook')
    ])

custom_hooks = [
    dict(
        type='GradientCumulativeOptimizerHook',
        cumulative_iters=2
    )
]

Hi~ It seems that you use loss.backward() mannualy, and GradientOptimizerHook execute the backward the second time and raise the error. Do you execute loss.backward in model mannuly?

Hi~ I assume I don't execute loss.backward mannualy. I use all the predefined components in MMSegmentation and didn't use any custom components. I only modified the config file

[HAOCHENYE]

Hi~ thanks for your suggestions, Gradient accumulative has been implemented here https://github.com/open-mmlab/mmcv/blob/1f2500102834a01b86bf9ae4db227cd8d724fa6e/mmcv/runner/hooks/optimizer.py#L99

I think it is a good idea to add Cross-Iteration Batch Normalization into NORM_LAYERS。

I used GradientCumulativeOptimizerHook get this stack trace

2022-06-26 09:33:16,631 - mmseg - WARNING - GradientCumulativeOptimizerHook may slightly decrease performance if the model has BatchNorm layers.
Traceback (most recent call last):
  File "E:/Work work/Python/Work/Practice/Segmentation/mmsegmentation/tools/train.py", line 242, in <module>
    main()
  File "E:/Work work/Python/Work/Practice/Segmentation/mmsegmentation/tools/train.py", line 231, in main
    train_segmentor(
  File "E:\Work work\Python\Work\Practice\Segmentation\mmsegmentation\mmseg\apis\train.py", line 194, in train_segmentor
    runner.run(data_loaders, cfg.workflow)
  File "E:\Anaconda\envs\openmmlab\lib\site-packages\mmcv\runner\iter_based_runner.py", line 135, in run
    iter_runner(iter_loaders[i], **kwargs)
  File "E:\Anaconda\envs\openmmlab\lib\site-packages\mmcv\runner\iter_based_runner.py", line 68, in train
    self.call_hook('after_train_iter')
  File "E:\Anaconda\envs\openmmlab\lib\site-packages\mmcv\runner\base_runner.py", line 309, in call_hook
    getattr(hook, fn_name)(self)
  File "E:\Anaconda\envs\openmmlab\lib\site-packages\mmcv\runner\hooks\optimizer.py", line 163, in after_train_iter
    loss.backward()
  File "E:\Anaconda\envs\openmmlab\lib\site-packages\torch\_tensor.py", line 363, in backward
    torch.autograd.backward(self, gradient, retain_graph, create_graph, inputs=inputs)
  File "E:\Anaconda\envs\openmmlab\lib\site-packages\torch\autograd\__init__.py", line 173, in backward
    Variable._execution_engine.run_backward(  # Calls into the C++ engine to run the backward pass
RuntimeError: Trying to backward through the graph a second time (or directly access saved tensors 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 tensors after calling backward.

Here's my config file

_base_ = [
    '../_base_/datasets/ade20k.py',
    '../_base_/default_runtime.py',
    '../_base_/schedules/schedule_40k.py',
]

custom_imports = dict(imports=['mmcls.models'], allow_failed_imports=False)
pretrained =\
    "https://download.openmmlab.com/mmclassification/v0/efficientnet/efficientnet-b1_3rdparty_8xb32-aa-advprop_in1k_20220119-5715267d.pth"

model = dict(
    type='EncoderDecoder',
    backbone=dict(
        type='mmcls.EfficientNet',
        arch='b1',
        out_indices=(2, 3, 4, 5),
        init_cfg=dict(
            type='Pretrained',
            checkpoint=pretrained,
            prefix='backbone.'
        )
    ),
    neck=dict(
        type='FPN',
        in_channels=[24, 40, 112, 320],
        out_channels=256,
        num_outs=4
    ),
    decode_head=dict(
        type='FCNHead',
        in_channels=[256, 256, 256, 256],
        channels=128,
        num_classes=3,
        in_index=[0, 1, 2, 3],
        input_transform='resize_concat',
        concat_input=False,
        loss_decode=dict(
            type='FocalLoss',
            use_sigmoid=True
        )
    ),
)

# dataset settings
dataset_type = 'Secret'
data_root = '../Dataset'

img_norm_cfg = dict(
    mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
crop_size = (128, 128)
train_pipeline = [
    dict(type='LoadImageFromFile'),
    dict(type='LoadAnnotations', reduce_zero_label=False),
    dict(type='Resize', img_scale=crop_size, keep_ratio=False, ratio_range=(1, 1)),
    dict(type='RandomFlip', prob=0.5),
    #dict(type='PhotoMetricDistortion'),
    dict(type='Normalize', **img_norm_cfg),
    dict(type='DefaultFormatBundle'),
    dict(type='Collect', keys=['img', 'gt_semantic_seg']),
]
test_pipeline = [
    dict(type='LoadImageFromFile'),
    dict(
        type='MultiScaleFlipAug',
        img_scale=crop_size,
        # img_ratios=[0.5, 0.75, 1.0, 1.25, 1.5, 1.75],
        flip=False,
        transforms=[
            dict(type='Resize', keep_ratio=False),
            dict(type='RandomFlip'),
            dict(type='Normalize', **img_norm_cfg),
            dict(type='ImageToTensor', keys=['img']),
            dict(type='Collect', keys=['img']),
        ])
]


data = dict(
    samples_per_gpu=1,
    workers_per_gpu=2,
    train=dict(
        type=dataset_type,
        data_root=data_root,
        img_dir='train/train',
        ann_dir='train_seg_map',
        pipeline=train_pipeline
    ),
    val=dict(
        type=dataset_type,
        data_root=data_root,
        img_dir='val',
        ann_dir='val_seg_map',
        pipeline=test_pipeline
    ),
    test=dict(
        type=dataset_type,
        data_root=data_root,
        img_dir='val',
        ann_dir='val_seg_map',
        pipeline=test_pipeline
    )
)

checkpoint_config = dict(by_epoch=False, interval=500)
evaluation = dict(interval=500, metric='mIoU', pre_eval=True)

log_config = dict(
    interval=1,
    hooks=[
        dict(type='TextLoggerHook', by_epoch=False),
        # dict(type='TensorboardLoggerHook')
    ])

custom_hooks = [
    dict(
        type='GradientCumulativeOptimizerHook',
        cumulative_iters=2
    )
]

Hi~ It seems that you use loss.backward() mannualy, and GradientOptimizerHook execute the backward the second time and raise the error. Do you execute loss.backward in model mannuly?

Hi~ I assume I don't execute loss.backward mannualy. I use all the predefined components in MMSegmentation and didn't use any custom components. I only modified the config file

https://github.com/open-mmlab/mmcv/issues/1379, it seems GradientOptimizerHook should be set in optimizer_config. Otherwise OptimizerHook and GradientOptimzerHook will be registered both.

[iumyx2612]

Hi~ thanks for your suggestions, Gradient accumulative has been implemented here https://github.com/open-mmlab/mmcv/blob/1f2500102834a01b86bf9ae4db227cd8d724fa6e/mmcv/runner/hooks/optimizer.py#L99

I think it is a good idea to add Cross-Iteration Batch Normalization into NORM_LAYERS。

I used GradientCumulativeOptimizerHook get this stack trace

2022-06-26 09:33:16,631 - mmseg - WARNING - GradientCumulativeOptimizerHook may slightly decrease performance if the model has BatchNorm layers.
Traceback (most recent call last):
  File "E:/Work work/Python/Work/Practice/Segmentation/mmsegmentation/tools/train.py", line 242, in <module>
    main()
  File "E:/Work work/Python/Work/Practice/Segmentation/mmsegmentation/tools/train.py", line 231, in main
    train_segmentor(
  File "E:\Work work\Python\Work\Practice\Segmentation\mmsegmentation\mmseg\apis\train.py", line 194, in train_segmentor
    runner.run(data_loaders, cfg.workflow)
  File "E:\Anaconda\envs\openmmlab\lib\site-packages\mmcv\runner\iter_based_runner.py", line 135, in run
    iter_runner(iter_loaders[i], **kwargs)
  File "E:\Anaconda\envs\openmmlab\lib\site-packages\mmcv\runner\iter_based_runner.py", line 68, in train
    self.call_hook('after_train_iter')
  File "E:\Anaconda\envs\openmmlab\lib\site-packages\mmcv\runner\base_runner.py", line 309, in call_hook
    getattr(hook, fn_name)(self)
  File "E:\Anaconda\envs\openmmlab\lib\site-packages\mmcv\runner\hooks\optimizer.py", line 163, in after_train_iter
    loss.backward()
  File "E:\Anaconda\envs\openmmlab\lib\site-packages\torch\_tensor.py", line 363, in backward
    torch.autograd.backward(self, gradient, retain_graph, create_graph, inputs=inputs)
  File "E:\Anaconda\envs\openmmlab\lib\site-packages\torch\autograd\__init__.py", line 173, in backward
    Variable._execution_engine.run_backward(  # Calls into the C++ engine to run the backward pass
RuntimeError: Trying to backward through the graph a second time (or directly access saved tensors 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 tensors after calling backward.

Here's my config file

_base_ = [
    '../_base_/datasets/ade20k.py',
    '../_base_/default_runtime.py',
    '../_base_/schedules/schedule_40k.py',
]

custom_imports = dict(imports=['mmcls.models'], allow_failed_imports=False)
pretrained =\
    "https://download.openmmlab.com/mmclassification/v0/efficientnet/efficientnet-b1_3rdparty_8xb32-aa-advprop_in1k_20220119-5715267d.pth"

model = dict(
    type='EncoderDecoder',
    backbone=dict(
        type='mmcls.EfficientNet',
        arch='b1',
        out_indices=(2, 3, 4, 5),
        init_cfg=dict(
            type='Pretrained',
            checkpoint=pretrained,
            prefix='backbone.'
        )
    ),
    neck=dict(
        type='FPN',
        in_channels=[24, 40, 112, 320],
        out_channels=256,
        num_outs=4
    ),
    decode_head=dict(
        type='FCNHead',
        in_channels=[256, 256, 256, 256],
        channels=128,
        num_classes=3,
        in_index=[0, 1, 2, 3],
        input_transform='resize_concat',
        concat_input=False,
        loss_decode=dict(
            type='FocalLoss',
            use_sigmoid=True
        )
    ),
)

# dataset settings
dataset_type = 'Secret'
data_root = '../Dataset'

img_norm_cfg = dict(
    mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
crop_size = (128, 128)
train_pipeline = [
    dict(type='LoadImageFromFile'),
    dict(type='LoadAnnotations', reduce_zero_label=False),
    dict(type='Resize', img_scale=crop_size, keep_ratio=False, ratio_range=(1, 1)),
    dict(type='RandomFlip', prob=0.5),
    #dict(type='PhotoMetricDistortion'),
    dict(type='Normalize', **img_norm_cfg),
    dict(type='DefaultFormatBundle'),
    dict(type='Collect', keys=['img', 'gt_semantic_seg']),
]
test_pipeline = [
    dict(type='LoadImageFromFile'),
    dict(
        type='MultiScaleFlipAug',
        img_scale=crop_size,
        # img_ratios=[0.5, 0.75, 1.0, 1.25, 1.5, 1.75],
        flip=False,
        transforms=[
            dict(type='Resize', keep_ratio=False),
            dict(type='RandomFlip'),
            dict(type='Normalize', **img_norm_cfg),
            dict(type='ImageToTensor', keys=['img']),
            dict(type='Collect', keys=['img']),
        ])
]


data = dict(
    samples_per_gpu=1,
    workers_per_gpu=2,
    train=dict(
        type=dataset_type,
        data_root=data_root,
        img_dir='train/train',
        ann_dir='train_seg_map',
        pipeline=train_pipeline
    ),
    val=dict(
        type=dataset_type,
        data_root=data_root,
        img_dir='val',
        ann_dir='val_seg_map',
        pipeline=test_pipeline
    ),
    test=dict(
        type=dataset_type,
        data_root=data_root,
        img_dir='val',
        ann_dir='val_seg_map',
        pipeline=test_pipeline
    )
)

checkpoint_config = dict(by_epoch=False, interval=500)
evaluation = dict(interval=500, metric='mIoU', pre_eval=True)

log_config = dict(
    interval=1,
    hooks=[
        dict(type='TextLoggerHook', by_epoch=False),
        # dict(type='TensorboardLoggerHook')
    ])

custom_hooks = [
    dict(
        type='GradientCumulativeOptimizerHook',
        cumulative_iters=2
    )
]

Hi~ It seems that you use loss.backward() mannualy, and GradientOptimizerHook execute the backward the second time and raise the error. Do you execute loss.backward in model mannuly?

Hi~ I assume I don't execute loss.backward mannualy. I use all the predefined components in MMSegmentation and didn't use any custom components. I only modified the config file

#1379, it seems GradientOptimizerHook should be set in optimizer_config. Otherwise OptimizerHook and GradientOptimzerHook will be registered both.

Thank you so much, working now