MMDetection系列 | 5. MMDetection运行配置介绍

---

如有错误，恳请指出。

---

开门见山，基于mmdet的官方文档直接介绍如何进行我们的运行配置。个人觉得，继承于default_runtime.py这个文件之后，主要需要自己稍微更改下的配置主要有7个，分别是：优化器配置、学习率配置、工作流程配置、检查点配置、日志配置、评估配置、训练设置。具体的配置流程如下所示。

如果需要其他钩子函数的实现与配置，具体可以查看参考资料1.

1. 优化器配置

optimizer = dict(type='SGD', lr=0.02, momentum=0.9, weight_decay=0.0001)
optimizer_config = dict(grad_clip=None)
1
2

• 使用梯度剪辑来稳定训练

optimizer_config = dict(
    _delete_=True, grad_clip=dict(max_norm=35, norm_type=2))
1
2

其中，_delete_=True将用新键替换backbone字段中的所有旧键

---

2. 学习率配置

lr_config = dict(
    policy='step',
    warmup='linear',
    warmup_iters=500,
    warmup_ratio=0.001,
    step=[8, 11])       # 表示初始学习率在第8和11个epoch衰减10倍
1
2
3
4
5
6

还有其他的配置方案：

• Poly schedule

lr_config = dict(policy='poly', power=0.9, min_lr=1e-4, by_epoch=False)
1

• ConsineAnnealing schedule

lr_config = dict(
    policy='CosineAnnealing',
    warmup='linear',
    warmup_iters=1000,
    warmup_ratio=1.0 / 10,
    min_lr_ratio=1e-5)
1
2
3
4
5
6

• 使用动量调度加速模型收敛

支持动量调度器根据学习率修改模型的动量，这可以使模型以更快的方式收敛。Momentum 调度器通常与 LR 调度器一起使用

lr_config = dict(
    policy='cyclic',
    target_ratio=(10, 1e-4),
    cyclic_times=1,
    step_ratio_up=0.4,
)
momentum_config = dict(
    policy='cyclic',
    target_ratio=(0.85 / 0.95, 1),
    cyclic_times=1,
    step_ratio_up=0.4,
)
1
2
3
4
5
6
7
8
9
10
11
12

---

3. 工作流程配置

工作流是 (phase, epochs) 的列表，用于指定运行顺序和时期。默认情况下，它设置为：

workflow = [('train', 1)]
1

这意味着运行 1 个 epoch 进行训练。有时用户可能想要检查验证集上模型的一些指标（例如损失、准确性）。在这种情况下，我们可以将工作流设置为

[('train', 1), ('val', 1)]
1

这样 1 个 epoch 的训练和 1 个 epoch 的验证将被迭代运行。而验证集的损失同样会被计算出来。如果想先进行验证，再进行训练，还可以设置如下：

[('val', 1), ('train', n)]
1

这样设置表示先对验证集进行验证与损失计算，再进行n个epoch的计算。

---

4. 检查点配置

checkpoint_config = dict(interval=20)          # 20个epoch保存一次权重
1

参数说明见：https://mmcv.readthedocs.io/en/latest/api.html#mmcv.runner.CheckpointHook

CLASSmmcv.runner.CheckpointHook(interval: int = - 1, by_epoch: bool = True, save_optimizer: bool = True, out_dir: Optional[str] = None, max_keep_ckpts: int = - 1, save_last: bool = True, sync_buffer: bool = False, file_client_args: Optional[dict] = None, **kwargs)

• interval (int) – The saving period. If by_epoch=True, interval indicates epochs, otherwise it indicates iterations. Default: -1, which means “never”.
• by_epoch (bool) – Saving checkpoints by epoch or by iteration. Default: True.
• save_optimizer (bool) – Whether to save optimizer state_dict in the checkpoint. It is usually used for resuming experiments. Default: True.
• out_dir (str, optional) – The root directory to save checkpoints. If not specified, runner.work_dir will be used by default. If specified, the out_dir will be the concatenation of out_dir and the last level directory of runner.work_dir. Changed in version 1.3.16.
• max_keep_ckpts (int, optional) – The maximum checkpoints to keep. In some cases we want only the latest few checkpoints and would like to delete old ones to save the disk space. Default: -1, which means unlimited.
• save_last (bool, optional) – Whether to force the last checkpoint to be saved regardless of interval. Default: True.
• sync_buffer (bool, optional) – Whether to synchronize buffers in different gpus. Default: False.
• file_client_args (dict, optional) – Arguments to instantiate a FileClient. See mmcv.fileio.FileClient for details. Default: None. New in version 1.3.16.

---

5. 日志配置

包装多个记录器log_config挂钩并允许设置间隔。现在 MMCV 支持WandbLoggerHook、MlflowLoggerHook和TensorboardLoggerHook.

log_config = dict(
    interval=50,    # 每500个迭代就打印一次训练信息
    hooks=[
        dict(type='TextLoggerHook'),
        # dict(type='TensorboardLoggerHook')
    ])
1
2
3
4
5
6

参数说明见：https://mmcv.readthedocs.io/en/latest/api.html#mmcv.runner.EvalHook

CLASSmmcv.runner.LoggerHook(interval: int = 10, ignore_last: bool = True, reset_flag: bool = False, by_epoch: bool = True)[SOURCE]

• interval (int) – Logging interval (every k iterations). Default 10.
• ignore_last (bool) – Ignore the log of last iterations in each epoch if less than interval. Default True.
• reset_flag (bool) – Whether to clear the output buffer after logging. Default False.
• by_epoch (bool) – Whether EpochBasedRunner is used. Default True.

---

6. 评估配置

配置的evaluation将用于初始化EvalHook. 除了 key interval，其他参数如metric将传递给dataset.evaluate()
evaluation = dict(interval=1, metric=‘bbox’)

参数说明：https://mmcv.readthedocs.io/en/latest/api.html?highlight=EpochBasedRunner#mmcv.runner.EpochBasedRunner

mmcv.runner.EvalHook(dataloader: torch.utils.data.dataloader.DataLoader, start: Optional[int] = None, interval: int = 1, by_epoch: bool = True, save_best: Optional[str] = None, rule: Optional[str] = None, test_fn: Optional[Callable] = None, greater_keys: Optional[List[str]] = None, less_keys: Optional[List[str]] = None, out_dir: Optional[str] = None, file_client_args: Optional[dict] = None, **eval_kwargs)

• dataloader (DataLoader) – A PyTorch dataloader, whose dataset has implemented evaluate function.
• start (int | None, optional) – Evaluation starting epoch. It enables evaluation before the training starts if start <= the resuming epoch. If None, whether to evaluate is merely decided by interval. Default: None.
• interval (int) – Evaluation interval. Default: 1.
• by_epoch (bool) – Determine perform evaluation by epoch or by iteration. If set to True, it will perform by epoch. Otherwise, by iteration. Default: True.
• save_best (str, optional) – If a metric is specified, it would measure the best checkpoint during evaluation. The information about best checkpoint would be saved in runner.meta[‘hook_msgs’] to keep best score value and best checkpoint path, which will be also loaded when resume checkpoint. Options are the evaluation metrics on the test dataset. e.g., bbox_mAP, segm_mAP for bbox detection and instance segmentation. AR@100 for proposal recall. If save_best is auto, the first key of the returned OrderedDict result will be used. Default: None.
• rule (str | None, optional) – Comparison rule for best score. If set to None, it will infer a reasonable rule. Keys such as ‘acc’, ‘top’ .etc will be inferred by ‘greater’ rule. Keys contain ‘loss’ will be inferred by ‘less’ rule. Options are ‘greater’, ‘less’, None. Default: None.
• test_fn (callable, optional) – test a model with samples from a dataloader, and return the test results. If None, the default test function mmcv.engine.single_gpu_test will be used. (default: None)
• greater_keys (List[str] | None, optional) – Metric keys that will be inferred by ‘greater’ comparison rule. If None, _default_greater_keys will be used. (default: None)
• less_keys (List[str] | None, optional) – Metric keys that will be inferred by ‘less’ comparison rule. If None, _default_less_keys will be used. (default: None)
• out_dir (str, optional) – The root directory to save checkpoints. If not specified, runner.work_dir will be used by default. If specified, the out_dir will be the concatenation of out_dir and the last level directory of runner.work_dir. New in version 1.3.16.
• file_client_args (dict) – Arguments to instantiate a FileClient. See mmcv.fileio.FileClient for details. Default: None. New in version 1.3.16.
• **eval_kwargs – Evaluation arguments fed into the evaluate function of the dataset.

---

7. 训练设置

runner = dict(type='EpochBasedRunner', max_epochs=150)   # 设置模型训练多少次
1

参数说明：https://mmcv.readthedocs.io/en/latest/api.html#mmcv.runner.EpochBasedRunner

mmcv.runner.EpochBasedRunner(model: torch.nn.modules.module.Module, batch_processor: Optional[Callable] = None, optimizer: Optional[Union[Dict, torch.optim.optimizer.Optimizer]] = None, work_dir: Optional[str] = None, logger: Optional[logging.Logger] = None, meta: Optional[Dict] = None, max_iters: Optional[int] = None, max_epochs: Optional[int] = None)

---

总结：

一般来说，我们写配置文件都会继承default_runtime.py这个文件

_base_ = [
    '../_base_/default_runtime.py'
]
1
2
3

这个文件的内容如下所示：

checkpoint_config = dict(interval=5)    # 每5个epoch保存一次权重
# yapf:disable
log_config = dict(
    interval=50,    # 每500个迭代就打印一次训练信息
    hooks=[
        dict(type='TextLoggerHook'),
        # dict(type='TensorboardLoggerHook')
    ])
# yapf:enable
custom_hooks = [dict(type='NumClassCheckHook')]

dist_params = dict(backend='nccl')
log_level = 'INFO'
load_from = None        # 加载权重文件
resume_from = None
workflow = [('train', 1)]

# disable opencv multithreading to avoid system being overloaded
opencv_num_threads = 0
# set multi-process start method as `fork` to speed up the training
mp_start_method = 'fork'

# Default setting for scaling LR automatically
#   - `enable` means enable scaling LR automatically
#       or not by default.
#   - `base_batch_size` = (8 GPUs) x (2 samples per GPU).
auto_scale_lr = dict(enable=False, base_batch_size=16)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27

一般不需要更改太多的内容，可以时代的更改log_config进行合理的打印训练信息，还有设置checkpoint_config进行合理的保存权重文件，其他的设置按默认即可。

下面展示我继承了default_runtime.py后更改的内容，其实就是更改了以上我所介绍的七点内容：

_base_ = [
    '../_base_/default_runtime.py'
]

......
# optimizer
optimizer = dict(   # 设置使用AdamW优化器(默认使用的是SGD)
    type='AdamW',
    lr=0.0001,
    weight_decay=0.0001,
    paramwise_cfg=dict(custom_keys={'backbone': dict(lr_mult=0.1, decay_mult=1.0)}))

evaluation = dict(interval=5, metric='bbox')   # 5个epoch验证一次
optimizer_config = dict(grad_clip=dict(max_norm=0.1, norm_type=2))  # 设置梯度裁剪(default_runtime.py中默认为None)
checkpoint_config = dict(interval=20)          # 20个epoch保存一次权重
log_config = dict(interval=50,     # 每50次迭代训练就打印一次信息(注意是迭代而不是epoch)
                  hooks=[dict(type='TextLoggerHook')])

# learning policy
lr_config = dict(policy='step', step=[100])              # 学习率在100个epoch进行衰减
runner = dict(type='EpochBasedRunner', max_epochs=150)   # 训练150个epoch
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21

---

参考资料：

1. Customize Runtime Settings

2. mmcv官方文档