Migration¶

We introduce some modifications in MMPretrain 1.x, and some of them are BC-breacking. To migrate your projects from MMClassification 0.x or MMSelfSup 0.x smoothly, please read this tutorial.

New dependencies¶

Warning

MMPretrain 1.x has new package dependencies, and a new environment should be created for MMPretrain 1.x even if you already have a well-rounded MMClassification 0.x or MMSelfSup 0.x environment. Please refer to the installation tutorial for the required package installation or install the packages manually.

MMEngine: MMEngine is the core the OpenMMLab 2.0 architecture, and we have split many compentents unrelated to computer vision from MMCV to MMEngine.
MMCV: The computer vision package of OpenMMLab. This is not a new dependency, but it should be upgraded to version 2.0.0rc1 or above.
rich: A terminal formatting package, and we use it to enhance some outputs in the terminal.
einops: Operators for Einstein notations.

General change of config¶

In this section, we introduce the general difference between old version(MMClassification 0.x or MMSelfSup 0.x) and MMPretrain 1.x.

Schedule settings¶

MMCls or MMSelfSup 0.x	MMPretrain 1.x	Remark
optimizer_config	/	It has been removed.
/	optim_wrapper	The `optim_wrapper` provides a common interface for updating parameters.
lr_config	param_scheduler	The `param_scheduler` is a list to set learning rate or other parameters, which is more flexible.
runner	train_cfg	The loop setting (`EpochBasedTrainLoop`, `IterBasedTrainLoop`) in `train_cfg` controls the work flow of the algorithm training.

Changes in optimizer and optimizer_config:

Now we use optim_wrapper field to specify all configurations related to optimization process. The optimizer has become a subfield of optim_wrapper.
The paramwise_cfg field is also a subfield of optim_wrapper, instead of optimizer.
The optimizer_config field has been removed, and all configurations has been moved to optim_wrapper.
The grad_clip field has been renamed to clip_grad.

Original

optimizer = dict(
    type='AdamW',
    lr=0.0015,
    weight_decay=0.3,
    paramwise_cfg = dict(
        norm_decay_mult=0.0,
        bias_decay_mult=0.0,
    ))

optimizer_config = dict(grad_clip=dict(max_norm=1.0))

New

optim_wrapper = dict(
    optimizer=dict(type='AdamW', lr=0.0015, weight_decay=0.3),
    paramwise_cfg = dict(
        norm_decay_mult=0.0,
        bias_decay_mult=0.0,
    ),
    clip_grad=dict(max_norm=1.0),
)

Changes in lr_config:

The lr_config field has been removed and replaced by the new param_scheduler.
The warmup related arguments have also been removed since we use a combination of schedulers to implement this functionality.

The new scheduler combination mechanism is highly flexible and enables the design of various learning rate/momentum curves. For more details, see the parameter schedulers tutorial.

Original

lr_config = dict(
    policy='CosineAnnealing',
    min_lr=0,
    warmup='linear',
    warmup_iters=5,
    warmup_ratio=0.01,
    warmup_by_epoch=True)

New

param_scheduler = [
    # warmup
    dict(
        type='LinearLR',
        start_factor=0.01,
        by_epoch=True,
        end=5,
        # Update the learning rate after every iters.
        convert_to_iter_based=True),
    # main learning rate scheduler
    dict(type='CosineAnnealingLR', by_epoch=True, begin=5),
]

Changes in runner:

Most of the configurations that were originally in the runner field have been moved to train_cfg, val_cfg, and test_cfg. These fields are used to configure the loop for training, validation, and testing.

Original

runner = dict(type='EpochBasedRunner', max_epochs=100)

New

# The `val_interval` is the original `evaluation.interval`.
train_cfg = dict(by_epoch=True, max_epochs=100, val_interval=1)
val_cfg = dict()   # Use the default validation loop.
test_cfg = dict()  # Use the default test loop.

In OpenMMLab 2.0, we introduced Loop to control the behaviors in training, validation and testing. As a result, the functionalities of Runner have also been changed. More details can be found in the MMEngine tutorials.

Runtime settings¶

Changes in checkpoint_config and log_config:

The checkpoint_config has been moved to default_hooks.checkpoint, and log_config has been moved to default_hooks.logger. Additionally, many hook settings that were previously included in the script code have been moved to the default_hooks field in the runtime configuration.

default_hooks = dict(
    # record the time of every iterations.
    timer=dict(type='IterTimerHook'),

    # print log every 100 iterations.
    logger=dict(type='LoggerHook', interval=100),

    # enable the parameter scheduler.
    param_scheduler=dict(type='ParamSchedulerHook'),

    # save checkpoint per epoch, and automatically save the best checkpoint.
    checkpoint=dict(type='CheckpointHook', interval=1, save_best='auto'),

    # set sampler seed in distributed evrionment.
    sampler_seed=dict(type='DistSamplerSeedHook'),

    # validation results visualization, set True to enable it.
    visualization=dict(type='VisualizationHook', enable=False),
)

In OpenMMLab 2.0, we have split the original logger into logger and visualizer. The logger is used to record information, while the visualizer is used to display the logger in different backends such as terminal, TensorBoard, and Wandb.

Original

log_config = dict(
    interval=100,
    hooks=[
        dict(type='TextLoggerHook'),
        dict(type='TensorboardLoggerHook'),
    ])

New

default_hooks = dict(
    ...
    logger=dict(type='LoggerHook', interval=100),
)

visualizer = dict(
    type='UniversalVisualizer',
    vis_backends=[dict(type='LocalVisBackend'), dict(type='TensorboardVisBackend')],
)

Changes in load_from and resume_from:

The resume_from is removed. And we use resume and load_from to replace it.
- If resume=True and load_from is not None, resume training from the checkpoint in load_from.
- If resume=True and load_from is None, try to resume from the latest checkpoint in the work directory.
- If resume=False and load_from is not None, only load the checkpoint, not resume training.
- If resume=False and load_from is None, do not load nor resume.

the resume_from field has been removed, and we use resume and load_from instead.

If resume=True and load_from is not None, training is resumed from the checkpoint in load_from.
If resume=True and load_from is None, the latest checkpoint in the work directory is used for resuming.
If resume=False and load_from is not None, only the checkpoint is loaded without resuming training.
If resume=False and load_from is None, neither checkpoint is loaded nor is training resumed.

Changes in dist_params: The dist_params field has become a subfield of env_cfg now. Additionally, some new configurations have been added to env_cfg.

env_cfg = dict(
    # whether to enable cudnn benchmark
    cudnn_benchmark=False,

    # set multi process parameters
    mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0),

    # set distributed parameters
    dist_cfg=dict(backend='nccl'),
)

Changes in workflow: workflow related functionalities are removed.

New field visualizer: The visualizer is a new design in OpenMMLab 2.0 architecture. The runner uses an instance of the visualizer to handle result and log visualization, as well as to save to different backends. For more information, please refer to the MMEngine tutorial.

visualizer = dict(
    type='UniversalVisualizer',
    vis_backends=[
        dict(type='LocalVisBackend'),
        # Uncomment the below line to save the log and visualization results to TensorBoard.
        # dict(type='TensorboardVisBackend')
    ]
)

New field default_scope: The start point to search module for all registries. The default_scope in MMPretrain is mmpretrain. See the registry tutorial for more details.

Other changes¶

We moved the definition of all registries in different packages to the mmpretrain.registry package.

Migration from MMClassification 0.x¶

Config files¶

In MMPretrain 1.x, we refactored the structure of configuration files, and the original files are not usable.

In this section, we will introduce all changes of the configuration files. And we assume you already have ideas of the config files.

Model settings¶

No changes in model.backbone, model.neck and model.head fields.

Changes in model.train_cfg:

BatchMixup is renamed to Mixup.
BatchCutMix is renamed to CutMix.
BatchResizeMix is renamed to ResizeMix.
The prob argument is removed from all augments settings, and you can use the probs field in train_cfg to specify probabilities of every augemnts. If no probs field, randomly choose one by the same probability.

Original

model = dict(
    ...
    train_cfg=dict(augments=[
        dict(type='BatchMixup', alpha=0.8, num_classes=1000, prob=0.5),
        dict(type='BatchCutMix', alpha=1.0, num_classes=1000, prob=0.5)
    ]
)

New

model = dict(
    ...
    train_cfg=dict(augments=[
        dict(type='Mixup', alpha=0.8), dict(type='CutMix', alpha=1.0)]
)

Data settings¶

Changes in data:

The original data field is splited to train_dataloader, val_dataloader and test_dataloader. This allows us to configure them in fine-grained. For example, you can specify different sampler and batch size during training and test.
The samples_per_gpu is renamed to batch_size.
The workers_per_gpu is renamed to num_workers.

Original

data = dict(
    samples_per_gpu=32,
    workers_per_gpu=2,
    train=dict(...),
    val=dict(...),
    test=dict(...),
)

New

train_dataloader = dict(
    batch_size=32,
    num_workers=2,
    dataset=dict(...),
    sampler=dict(type='DefaultSampler', shuffle=True)  # necessary
)

val_dataloader = dict(
    batch_size=32,
    num_workers=2,
    dataset=dict(...),
    sampler=dict(type='DefaultSampler', shuffle=False)  # necessary
)

test_dataloader = val_dataloader

Changes in pipeline:

The original formatting transforms ToTensor, ImageToTensor and Collect are combined as PackInputs.
We don’t recommend to do Normalize in the dataset pipeline. Please remove it from pipelines and set it in the data_preprocessor field.
The argument flip_prob in RandomFlip is renamed to prob.
The argument size in RandomCrop is renamed to crop_size.
The argument size in RandomResizedCrop is renamed to scale.
The argument size in Resize is renamed to scale. And Resize won’t support size like (256, -1), please use ResizeEdge to replace it.
The argument policies in AutoAugment and RandAugment supports using string to specify preset policies. AutoAugment supports “imagenet” and RandAugment supports “timm_increasing”.
RandomResizedCrop and CenterCrop won’t supports efficientnet_style, and please use EfficientNetRandomCrop and EfficientNetCenterCrop to replace them.

Note

We move some work of data transforms to the data preprocessor, like normalization, see the documentation for more details.

Original

img_norm_cfg = dict(
    mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)

train_pipeline = [
    dict(type='LoadImageFromFile'),
    dict(type='RandomResizedCrop', size=224),
    dict(type='RandomFlip', flip_prob=0.5, direction='horizontal'),
    dict(type='Normalize', **img_norm_cfg),
    dict(type='ImageToTensor', keys=['img']),
    dict(type='ToTensor', keys=['gt_label']),
    dict(type='Collect', keys=['img', 'gt_label'])
]

New

data_preprocessor = dict(
    mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)

train_pipeline = [
    dict(type='LoadImageFromFile'),
    dict(type='RandomResizedCrop', scale=224),
    dict(type='RandomFlip', prob=0.5, direction='horizontal'),
    dict(type='PackInputs'),
]

Changes in evaluation:

The evaluation field is splited to val_evaluator and test_evaluator. And it won’t supports interval and save_best arguments. The interval is moved to train_cfg.val_interval, see the schedule settings and the save_best is moved to default_hooks.checkpoint.save_best, see the runtime settings.
The ‘accuracy’ metric is renamed to Accuracy.
The ‘precision’, ‘recall’, ‘f1-score’ and ‘support’ are combined as SingleLabelMetric, and use items argument to specify to calculate which metric.
The ‘mAP’ is renamed to AveragePrecision.
The ‘CP’, ‘CR’, ‘CF1’, ‘OP’, ‘OR’, ‘OF1’ are combined as MultiLabelMetric, and use items and average arguments to specify to calculate which metric.

Original	evaluation = dict( interval=1, metric='accuracy', metric_options=dict(topk=(1, 5)) )
New	val_evaluator = dict(type='Accuracy', topk=(1, 5)) test_evaluator = val_evaluator
Original	evaluation = dict( interval=1, metric=['mAP', 'CP', 'OP', 'CR', 'OR', 'CF1', 'OF1'], metric_options=dict(thr=0.5), )
New	val_evaluator = [ dict(type='AveragePrecision'), dict(type='MultiLabelMetric', items=['precision', 'recall', 'f1-score'], average='both', thr=0.5), ] test_evaluator = val_evaluator

Packages¶

`mmpretrain.apis`¶

The documentation can be found here.

Function	Changes
`init_model`	No changes
`inference_model`	No changes. But we recommend to use `mmpretrain.ImageClassificationInferencer` instead.
`train_model`	Removed, use `runner.train` to train.
`multi_gpu_test`	Removed, use `runner.test` to test.
`single_gpu_test`	Removed, use `runner.test` to test.
`show_result_pyplot`	Removed, use `mmpretrain.ImageClassificationInferencer` to inference model and show the result.
`set_random_seed`	Removed, use `mmengine.runner.set_random_seed`.
`init_random_seed`	Removed, use `mmengine.dist.sync_random_seed`.

`mmpretrain.core`¶

The mmpretrain.core package is renamed to mmpretrain.engine.

Sub package	Changes
`evaluation`	Removed, use the metrics in `mmpretrain.evaluation`.
`hook`	Moved to `mmpretrain.engine.hooks`
`optimizers`	Moved to `mmpretrain.engine.optimizers`
`utils`	Removed, the distributed environment related functions can be found in the `mmengine.dist` package.
`visualization`	Removed, the related functionalities are implemented in `mmengine.visualization.Visualizer`.

The MMClsWandbHook in hooks package is waiting for implementation.

The CosineAnnealingCooldownLrUpdaterHook in hooks package is removed, and we support this functionality by the combination of parameter schedulers, see the tutorial.

`mmpretrain.datasets`¶

The documentation can be found here.

Dataset class	Changes
`CustomDataset`	Add `data_root` argument as the common prefix of `data_prefix` and `ann_file` and support to load unlabeled data.
`ImageNet`	Same as `CustomDataset`.
`ImageNet21k`	Same as `CustomDataset`.
`CIFAR10` & `CIFAR100`	The `test_mode` argument is a required argument now.
`MNIST` & `FashionMNIST`	The `test_mode` argument is a required argument now.
`VOC`	Requires `data_root`, `image_set_path` and `test_mode` now.
`CUB`	Requires `data_root` and `test_mode` now.

The mmpretrain.datasets.pipelines is renamed to mmpretrain.datasets.transforms.

Transform class	Changes
`LoadImageFromFile`	Removed, use `mmcv.transforms.LoadImageFromFile`.
`RandomFlip`	Removed, use `mmcv.transforms.RandomFlip`. The argument `flip_prob` is renamed to `prob`.
`RandomCrop`	The argument `size` is renamed to `crop_size`.
`RandomResizedCrop`	The argument `size` is renamed to `scale`. The argument `scale` is renamed to `crop_ratio_range`. Won’t support `efficientnet_style`, use `EfficientNetRandomCrop`.
`CenterCrop`	Removed, use `mmcv.transforms.CenterCrop`. Won’t support `efficientnet_style`, use `EfficientNetCenterCrop`.
`Resize`	Removed, use `mmcv.transforms.Resize`. The argument `size` is renamed to `scale`. Won’t support size like `(256, -1)`, use `ResizeEdge`.
`AutoAugment` & `RandomAugment`	The argument `policies` supports using string to specify preset policies.
`Compose`	Removed, use `mmcv.transforms.Compose`.

`mmpretrain.models`¶

The documentation can be found here. The interface of all backbones, necks and losses didn’t change.

Changes in ImageClassifier:

Method of classifiers	Changes
`extract_feat`	No changes
`forward`	Now only accepts three arguments: `inputs`, `data_samples` and `mode`. See the documentation for more details.
`forward_train`	Replaced by `loss`.
`simple_test`	Replaced by `predict`.
`train_step`	The `optimizer` argument is replaced by `optim_wrapper` and it accepts `OptimWrapper`.
`val_step`	The original `val_step` is the same as `train_step`, now it calls `predict`.
`test_step`	New method, and it’s the same as `val_step`.

Changes in heads:

Method of heads	Changes
`pre_logits`	No changes
`forward_train`	Replaced by `loss`.
`simple_test`	Replaced by `predict`.
`loss`	It accepts `data_samples` instead of `gt_labels` to calculate loss. The `data_samples` is a list of ClsDataSample.
`forward`	New method, and it returns the output of the classification head without any post-processs like softmax or sigmoid.

`mmpretrain.utils`¶

Function	Changes
`collect_env`	No changes
`get_root_logger`	Removed, use `mmengine.logging.MMLogger.get_current_instance`
`load_json_log`	The output format changed.
`setup_multi_processes`	Removed, use `mmengine.utils.dl_utils.set_multi_processing`.
`wrap_non_distributed_model`	Removed, we auto wrap the model in the runner.
`wrap_distributed_model`	Removed, we auto wrap the model in the runner.
`auto_select_device`	Removed, we auto select the device in the runner.

Migration from MMSelfSup 0.x¶

Config¶

This section illustrates the changes of our config files in the _base_ folder, which includes three parts

Datasets: configs/_base_/datasets
Models: configs/_base_/models
Schedules: configs/_base_/schedules

Dataset settings¶

In MMSelfSup 0.x, we use key data to summarize all information, such as samples_per_gpu, train, val, etc.

In MMPretrain 1.x, we separate train_dataloader, val_dataloader to summarize information correspodingly and the key data has been removed.

Original

data = dict(
    samples_per_gpu=32,  # total 32*8(gpu)=256
    workers_per_gpu=4,
    train=dict(
        type=dataset_type,
        data_source=dict(
            type=data_source,
            data_prefix='data/imagenet/train',
            ann_file='data/imagenet/meta/train.txt',
        ),
        num_views=[1, 1],
        pipelines=[train_pipeline1, train_pipeline2],
        prefetch=prefetch,
    ),
    val=...)

New

train_dataloader = dict(
    batch_size=32,
    num_workers=4,
    persistent_workers=True,
    sampler=dict(type='DefaultSampler', shuffle=True),
    collate_fn=dict(type='default_collate'),
    dataset=dict(
        type=dataset_type,
        data_root=data_root,
        ann_file='meta/train.txt',
        data_prefix=dict(img_path='train/'),
        pipeline=train_pipeline))
val_dataloader = ...

Besides, we remove the key of data_source to keep the pipeline format consistent with that in other OpenMMLab projects. Please refer to Config for more details.

Changes in pipeline:

Take MAE as an example of pipeline:

train_pipeline = [
    dict(type='LoadImageFromFile'),
    dict(
        type='RandomResizedCrop',
        scale=224,
        crop_ratio_range=(0.2, 1.0),
        backend='pillow',
        interpolation='bicubic'),
    dict(type='RandomFlip', prob=0.5),
    dict(type='PackInputs')
]

Model settings¶

In the config of models, there are two main different parts from MMSeflSup 0.x.

There is a new key called data_preprocessor, which is responsible for preprocessing the data, like normalization, channel conversion, etc. For example:

data_preprocessor=dict(
    mean=[123.675, 116.28, 103.53],
    std=[58.395, 57.12, 57.375],
    bgr_to_rgb=True)
model = dict(
    type='MAE',
    data_preprocessor=dict(
        mean=[127.5, 127.5, 127.5],
        std=[127.5, 127.5, 127.5],
        bgr_to_rgb=True),
    backbone=...,
    neck=...,
    head=...,
    init_cfg=...)

There is a new key loss in head in MMPretrain 1.x, to determine the loss function of the algorithm. For example:

model = dict(
    type='MAE',
    backbone=...,
    neck=...,
    head=dict(
        type='MAEPretrainHead',
        norm_pix=True,
        patch_size=16,
        loss=dict(type='MAEReconstructionLoss')),
    init_cfg=...)

Package¶

The table below records the general modification of the folders and files.

MMSelfSup 0.x	MMPretrain 1.x	Remark
apis	apis	The high level APIs are updated.
core	engine	The `core` folder has been renamed to `engine`, which includes `hooks`, `opimizers`. (API link)
datasets	datasets	The datasets is implemented according to different datasets, such as ImageNet, Places205. (API link)
datasets/data_sources	/	The `data_sources` has been removed and the directory of `datasets` now is consistent with other OpenMMLab projects.
datasets/pipelines	datasets/transforms	The `pipelines` folder has been renamed to `transforms`. (API link)
/	evaluation	The `evaluation` is created for some evaluation functions or classes. (API link)
models/algorithms	selfsup	The algorithms are moved to `selfsup` folder. (API link)
models/backbones	selfsup	The re-implemented backbones are moved to corresponding self-supervised learning algorithm `.py` files. (API link)
models/target_generators	selfsup	The target generators are moved to corresponding self-supervised learning algorithm `.py` files. (API link)
/	models/losses	The `losses` folder is created to provide different loss implementations, which is from `heads`. (API link)
/	structures	The `structures` folder is for the implementation of data structures. In MMPretrain, we implement a new data structure, `DataSample`, to pass and receive data throughout the training/val process. (API link)
/	visualization	The `visualization` folder contains the visualizer, which is responsible for some visualization tasks like visualizing data augmentation. (API link)

Migration¶

New dependencies¶

General change of config¶

Schedule settings¶

Runtime settings¶

Other changes¶

Migration from MMClassification 0.x¶

Config files¶

Model settings¶

Data settings¶

Packages¶

mmpretrain.apis¶

mmpretrain.core¶

mmpretrain.datasets¶

mmpretrain.models¶

mmpretrain.utils¶

Migration from MMSelfSup 0.x¶

Config¶

Dataset settings¶

Model settings¶

Package¶

`mmpretrain.apis`¶

`mmpretrain.core`¶

`mmpretrain.datasets`¶

`mmpretrain.models`¶

`mmpretrain.utils`¶