Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

[Bug] distill for yolox_s failed #184

Closed
tanghy2016 opened this issue Jun 20, 2022 · 8 comments · May be fixed by #192
Closed

[Bug] distill for yolox_s failed #184

tanghy2016 opened this issue Jun 20, 2022 · 8 comments · May be fixed by #192
Assignees
@pppppM
Labels
Bug:P2 usage

Comments

@tanghy2016
Copy link

Describe the bug

The original Backbone for yolox_s is CSPDarknet. I use ResNet18 instead. The detailed configuration is as follows:

algorithm = dict(
    type='GeneralDistill',
    architecture=dict(
        type='MMDetArchitecture',
        model=dict(
            type='mmdet.YOLOX',
            input_size=(640, 640),
            random_size_range=(15, 25),
            random_size_interval=10,
            backbone=dict(
                type='ResNet',
                depth=18,
                num_stages=4,
                out_indices=(1, 2, 3),
                norm_cfg=dict(type='BN', requires_grad=True),
                norm_eval=True,
                style='pytorch',
                init_cfg=dict(
                    type='Pretrained', checkpoint='torchvision://resnet18')),
            neck=dict(
                type='YOLOXPAFPN',
                in_channels=[128, 256, 512],
                out_channels=128,
                num_csp_blocks=1),
            bbox_head=dict(
                type='YOLOXHead',
                num_classes=1,
                in_channels=128,
                feat_channels=128),
            train_cfg=dict(
                assigner=dict(type='SimOTAAssigner', center_radius=2.5)),
            test_cfg=dict(
                score_thr=0.01, nms=dict(type='nms', iou_threshold=0.65)))),
    with_student_loss=True,
    with_teacher_loss=False,
    distiller=dict(
        type='SingleTeacherDistiller',
        teacher=dict(
            type='mmdet.YOLOX',
            init_cfg=dict(
                type='Pretrained',
                checkpoint='/root/minio_model/1518897383465840642/epoch_19.pth'
            ),
            input_size=(640, 640),
            random_size_range=(15, 25),
            random_size_interval=10,
            backbone=dict(
                type='CSPDarknet', deepen_factor=0.33, widen_factor=0.5),
            neck=dict(
                type='YOLOXPAFPN',
                in_channels=[128, 256, 512],
                out_channels=128,
                num_csp_blocks=1),
            bbox_head=dict(
                type='YOLOXHead',
                num_classes=1,
                in_channels=128,
                feat_channels=128),
            train_cfg=dict(
                assigner=dict(type='SimOTAAssigner', center_radius=2.5)),
            test_cfg=dict(
                score_thr=0.01, nms=dict(type='nms', iou_threshold=0.65))),
        teacher_trainable=False,
        components=[
            dict(
                student_module='neck.out_convs.0.conv',
                teacher_module='neck.out_convs.0.conv',
                losses=[
                    dict(
                        type='ChannelWiseDivergence',
                        name='loss_cwd_logits',
                        tau=1,
                        loss_weight=5)
                ]),
            dict(
                student_module='neck.out_convs.1.conv',
                teacher_module='neck.out_convs.1.conv',
                losses=[
                    dict(
                        type='ChannelWiseDivergence',
                        name='loss_cwd_logits',
                        tau=1,
                        loss_weight=5)
                ]),
            dict(
                student_module='neck.out_convs.2.conv',
                teacher_module='neck.out_convs.2.conv',
                losses=[
                    dict(
                        type='ChannelWiseDivergence',
                        name='loss_cwd_logits',
                        tau=1,
                        loss_weight=5)
                ])
        ]))

When it runs to the epoch 2, the following error occurs:

2022-06-20 07:27:45,022 - mmdet - INFO - Epoch [2][2/8]	lr: 6.250e-04, eta: 0:05:27, time: 0.702, data_time: 0.108, memory: 6444, student.loss_cls: 0.4989, student.loss_bbox: 4.8508, student.loss_obj: 19.1068, distiller.loss_cwd_logits.0: 0.2968, loss: 24.7533
Traceback (most recent call last):
  File "MMCV/train.py", line 52, in <module>
    main()
  File "MMCV/train.py", line 46, in main
    train_distill(args)
  File "/root/project/xbrain-ai/XbrainAiModels/MMCV/tools/train_distill.py", line 39, in train_distill
    train_mmdet(args, cfg)
  File "/root/project/xbrain-ai/XbrainAiModels/MMCV/tools/train_mmdet.py", line 43, in train_mmdet
    train_mmdet_model(
  File "/root/project/xbrain-ai/ThirdParty/MMCV/mmrazor/mmrazor/apis/mmdet/train.py", line 206, in train_mmdet_model
    runner.run(data_loader, cfg.workflow)
  File "/opt/conda/lib/python3.8/site-packages/mmcv/runner/epoch_based_runner.py", line 127, in run
    epoch_runner(data_loaders[i], **kwargs)
  File "/opt/conda/lib/python3.8/site-packages/mmcv/runner/epoch_based_runner.py", line 50, in train
    self.run_iter(data_batch, train_mode=True, **kwargs)
  File "/opt/conda/lib/python3.8/site-packages/mmcv/runner/epoch_based_runner.py", line 29, in run_iter
    outputs = self.model.train_step(data_batch, self.optimizer,
  File "/opt/conda/lib/python3.8/site-packages/mmcv/parallel/data_parallel.py", line 75, in train_step
    return self.module.train_step(*inputs[0], **kwargs[0])
  File "/root/project/xbrain-ai/ThirdParty/MMCV/mmrazor/mmrazor/models/algorithms/general_distill.py", line 49, in train_step
    distill_losses = self.distiller.compute_distill_loss(data)
  File "/root/project/xbrain-ai/ThirdParty/MMCV/mmrazor/mmrazor/models/distillers/single_teacher.py", line 240, in compute_distill_loss
    losses[loss_name] = loss_module(s_out, t_out)
  File "/opt/conda/lib/python3.8/site-packages/torch/nn/modules/module.py", line 889, in _call_impl
    result = self.forward(*input, **kwargs)
  File "/root/project/xbrain-ai/ThirdParty/MMCV/mmrazor/mmrazor/models/losses/cwd.py", line 41, in forward
    assert preds_S.shape[-2:] == preds_T.shape[-2:]
AssertionError

Correct situation, no matter preds_S or preds_T, its shape should be (batch_size, 3, 80, 80), (batch_size, 3, 40, 40) or (batch_size, 3, 20, 20). But when the above error occurs, in my debugging, various sizes have appeared.
@tanghy2016 tanghy2016 added the bug Something isn't working label Jun 20, 2022
@tanghy2016 tanghy2016 mentioned this issue Jun 20, 2022
Open
@pppppM
Copy link
Collaborator

pppppM commented Jun 21, 2022

This should not be a bug, the resize of yolox is in the forward, which will cause the teacher and student input picture size to be inconsistent.

https://github.com/open-mmlab/mmdetection/blob/master/mmdet/models/detectors/yolox.py#L93

This issue #17 may help you.

@pppppM pppppM added usage and removed bug Something isn't working labels Jun 21, 2022
@tanghy2016
Copy link
Author

tanghy2016 commented Jun 21, 2022 •
edited
Loading

reference to your reply, I added:

    align_methods=[
        dict(method='YOLOX._preprocess', import_module='mmdet.models')
    ],

but it didn't take effect.

There is a note in BaseDistiller.__init__:

        if align_methods is None:
            self.context_manager = None
        else:
            # To obtain the python function's outputs, there will build a
            # specific context manager. When enter the context manager, the
            # functions will be rewrite. The context manager could record
            # inputs or outputs of the functions , and pass from teachr to
            # student. When exit the context manager, the rewritten functions
            # will restore.
            self.context_manager = ConversionContext(align_methods)

but, in SingleTeacherDistiller.exec_student_forward, input of student is **data.
Is it the problem here?

@pppppM pppppM self-assigned this Jun 22, 2022
@pppppM
Copy link
Collaborator

pppppM commented Jun 22, 2022

Same error as before adding align_method？

@tanghy2016
Copy link
Author

Yes. The config after adding align_method is as follows：

algorithm = dict(
    type='GeneralDistill',
    architecture=dict(
        type='MMDetArchitecture',
        model=dict(
            type='mmdet.YOLOX',
            input_size=(640, 640),
            random_size_range=(15, 25),
            random_size_interval=10,
            backbone=dict(
                type='ResNet',
                depth=18,
                num_stages=4,
                out_indices=(1, 2, 3),
                norm_cfg=dict(type='BN', requires_grad=True),
                norm_eval=True,
                style='pytorch',
                init_cfg=dict(
                    type='Pretrained', checkpoint='torchvision://resnet18')),
            neck=dict(
                type='YOLOXPAFPN',
                in_channels=[128, 256, 512],
                out_channels=128,
                num_csp_blocks=1),
            bbox_head=dict(
                type='YOLOXHead',
                num_classes=1,
                in_channels=128,
                feat_channels=128),
            train_cfg=dict(
                assigner=dict(type='SimOTAAssigner', center_radius=2.5)),
            test_cfg=dict(
                score_thr=0.01, nms=dict(type='nms', iou_threshold=0.65)))),
    with_student_loss=True,
    with_teacher_loss=False,
    distiller=dict(
        type='SingleTeacherDistiller',
        teacher=dict(
            type='mmdet.YOLOX',
            init_cfg=dict(
                type='Pretrained',
                checkpoint='/root/minio_model/1518897383465840642/epoch_19.pth'
            ),
            input_size=(640, 640),
            random_size_range=(15, 25),
            random_size_interval=10,
            backbone=dict(
                type='CSPDarknet', deepen_factor=0.33, widen_factor=0.5),
            neck=dict(
                type='YOLOXPAFPN',
                in_channels=[128, 256, 512],
                out_channels=128,
                num_csp_blocks=1),
            bbox_head=dict(
                type='YOLOXHead',
                num_classes=1,
                in_channels=128,
                feat_channels=128),
            train_cfg=dict(
                assigner=dict(type='SimOTAAssigner', center_radius=2.5)),
            test_cfg=dict(
                score_thr=0.01, nms=dict(type='nms', iou_threshold=0.65))),
        teacher_trainable=False,
        align_methods=[
            dict(method='YOLOX._preprocess', import_module='mmdet.models')
        ],
        components=[
            dict(
                student_module='neck.out_convs.0.conv',
                teacher_module='neck.out_convs.0.conv',
                losses=[
                    dict(
                        type='ChannelWiseDivergence',
                        name='loss_cwd_logits',
                        tau=1,
                        loss_weight=5)
                ]),
            dict(
                student_module='neck.out_convs.1.conv',
                teacher_module='neck.out_convs.1.conv',
                losses=[
                    dict(
                        type='ChannelWiseDivergence',
                        name='loss_cwd_logits',
                        tau=1,
                        loss_weight=5)
                ]),
            dict(
                student_module='neck.out_convs.2.conv',
                teacher_module='neck.out_convs.2.conv',
                losses=[
                    dict(
                        type='ChannelWiseDivergence',
                        name='loss_cwd_logits',
                        tau=1,
                        loss_weight=5)
                ])
        ]))

@pppppM pppppM linked a pull request Jun 27, 2022 that will close this issue
[Bug] Fix a bug in base distiller and add yolox distill example #192
Open
6 tasks
@pppppM pppppM added the Bug:P2 label Jun 27, 2022
@pppppM
Copy link
Collaborator

pppppM commented Jun 27, 2022

I just submitted a pr and added an example about yolox.

@tanghy2016
Copy link
Author

ths, solved

@tanghy2016
Copy link
Author

hi, I have a new problem. This problem is not encountered when the iteration total epoch setting is small (<10). It is encountered when it is large (20, 50, etc.). Is it related to lr_config.num_last_epochs?:

2022-09-26 15:09:49,822 - mmdet - INFO - Exp name: cwd_neck-yolox_s-csp-r18.py
2022-09-26 15:09:49,822 - mmdet - INFO - Epoch(val) [4][49]	AP50: 0.0000, mAP: 0.0004
2022-09-26 15:09:49,834 - mmdet - INFO - No mosaic and mixup aug now!
2022-09-26 15:09:49,940 - mmdet - INFO - Add additional L1 loss now!
Traceback (most recent call last):
  File "MMCV/train.py", line 55, in <module>
    main()
  File "MMCV/train.py", line 49, in main
    train_distill(args)
  File "/root/project/xbrain-ai/XbrainAiModels/MMCV/tools/train_distill.py", line 44, in train_distill
    train_mmdet(args, cfg)
  File "/root/project/xbrain-ai/XbrainAiModels/MMCV/tools/train_mmdet.py", line 43, in train_mmdet
    train_mmdet_model(
  File "/root/project/xbrain-ai/ThirdParty/MMCV/mmrazor/mmrazor/apis/mmdet/train.py", line 206, in train_mmdet_model
    runner.run(data_loader, cfg.workflow)
  File "/opt/conda/lib/python3.8/site-packages/mmcv/runner/epoch_based_runner.py", line 127, in run
    epoch_runner(data_loaders[i], **kwargs)
  File "/opt/conda/lib/python3.8/site-packages/mmcv/runner/epoch_based_runner.py", line 45, in train
    self.call_hook('before_train_epoch')
  File "/opt/conda/lib/python3.8/site-packages/mmcv/runner/base_runner.py", line 309, in call_hook
    getattr(hook, fn_name)(self)
  File "/root/project/xbrain-ai/ThirdParty/MMCV/mmdetection/mmdet/core/hook/yolox_mode_switch_hook.py", line 47, in before_train_epoch
    model.bbox_head.use_l1 = True
  File "/opt/conda/lib/python3.8/site-packages/torch/nn/modules/module.py", line 947, in __getattr__
    raise AttributeError("'{}' object has no attribute '{}'".format(
AttributeError: 'AlignMethodDistill' object has no attribute 'bbox_head'

lr_config as follow:

max_epochs = 20
num_last_epochs = 15
resume_from = None
interval = 10

# learning policy
lr_config = dict(
    _delete_=True,
    policy='YOLOX',
    warmup='exp',
    by_epoch=False,
    warmup_by_epoch=True,
    warmup_ratio=1,
    warmup_iters=3,  # 5 epoch
    num_last_epochs=num_last_epochs,
    min_lr_ratio=0.05)

@Linker-Stars
Copy link

hi, I have a new problem. This problem is not encountered when the iteration total epoch setting is small (<10). It is encountered when it is large (20, 50, etc.). Is it related to lr_config.num_last_epochs?:

2022-09-26 15:09:49,822 - mmdet - INFO - Exp name: cwd_neck-yolox_s-csp-r18.py
2022-09-26 15:09:49,822 - mmdet - INFO - Epoch(val) [4][49]	AP50: 0.0000, mAP: 0.0004
2022-09-26 15:09:49,834 - mmdet - INFO - No mosaic and mixup aug now!
2022-09-26 15:09:49,940 - mmdet - INFO - Add additional L1 loss now!
Traceback (most recent call last):
  File "MMCV/train.py", line 55, in <module>
    main()
  File "MMCV/train.py", line 49, in main
    train_distill(args)
  File "/root/project/xbrain-ai/XbrainAiModels/MMCV/tools/train_distill.py", line 44, in train_distill
    train_mmdet(args, cfg)
  File "/root/project/xbrain-ai/XbrainAiModels/MMCV/tools/train_mmdet.py", line 43, in train_mmdet
    train_mmdet_model(
  File "/root/project/xbrain-ai/ThirdParty/MMCV/mmrazor/mmrazor/apis/mmdet/train.py", line 206, in train_mmdet_model
    runner.run(data_loader, cfg.workflow)
  File "/opt/conda/lib/python3.8/site-packages/mmcv/runner/epoch_based_runner.py", line 127, in run
    epoch_runner(data_loaders[i], **kwargs)
  File "/opt/conda/lib/python3.8/site-packages/mmcv/runner/epoch_based_runner.py", line 45, in train
    self.call_hook('before_train_epoch')
  File "/opt/conda/lib/python3.8/site-packages/mmcv/runner/base_runner.py", line 309, in call_hook
    getattr(hook, fn_name)(self)
  File "/root/project/xbrain-ai/ThirdParty/MMCV/mmdetection/mmdet/core/hook/yolox_mode_switch_hook.py", line 47, in before_train_epoch
    model.bbox_head.use_l1 = True
  File "/opt/conda/lib/python3.8/site-packages/torch/nn/modules/module.py", line 947, in __getattr__
    raise AttributeError("'{}' object has no attribute '{}'".format(
AttributeError: 'AlignMethodDistill' object has no attribute 'bbox_head'

lr_config as follow:

max_epochs = 20
num_last_epochs = 15
resume_from = None
interval = 10

# learning policy
lr_config = dict(
    _delete_=True,
    policy='YOLOX',
    warmup='exp',
    by_epoch=False,
    warmup_by_epoch=True,
    warmup_ratio=1,
    warmup_iters=3,  # 5 epoch
    num_last_epochs=num_last_epochs,
    min_lr_ratio=0.05)

I think this is caused by the YOLOXModeSwitchHook of the yolox model. You can rewrite this hook and add model.bbox_ head.use_ L1=True is changed to model.architecture.model.bbox_ head.use_ l1 = True

humu789 pushed a commit to humu789/mmrazor that referenced this issue Feb 13, 2023
@VVsssssk
[Docs]Fix docs about cfg (open-mmlab#184)
fa626a5 
* add docs about config comment

* fix blank

* fix comment

* fix comment

* fix comment

* fix comment
humu789 pushed a commit to humu789/mmrazor that referenced this issue Feb 13, 2023
@lvhan028
[enhancement] Reorganizing OpenMMLab projects in readme (open-mmlab#184)
e89becd 
* [enhancement] Reorganizing OpenMMLab projects in readme

* add MMPose in supported codebase list

* add MMPose in supported codebase list
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Assignees

@pppppM pppppM

Labels
Bug:P2 usage
Projects
None yet
Milestone
No milestone
Development

Successfully merging a pull request may close this issue.

[Bug] Fix a bug in base distiller and add yolox distill example pppppM/mmrazor
3 participants
@tanghy2016 @Linker-Stars @pppppM