Skip to content

Latest commit

 

History

History
169 lines (122 loc) · 5.98 KB

migration.md

File metadata and controls

169 lines (122 loc) · 5.98 KB

Migration from MMTracking 0.xx

Compared with the 0.xx versions of MMTracking, the latest 1.xx version of MMTracking has the following important modifications.

Overall Structures

The core in the old versions of MMTracking is splited into engine, evaluation, structures, visualization and model/task_moduls in the 1.xx version of MMTracking. Details can be seen in the user guides.

Configs

file names

old: deepsort_faster-rcnn_fpn_4e_mot17-private-half.py

new: deepsort_faster-rcnn-resnet50-fpn_8x2bs-4e_mot17halftrain_test-mot17halfval.py

keys of dataset loader

old

data = dict(
    samples_per_gpu=16,
    workers_per_gpu=4,
    persistent_workers=True,
    samples_per_epoch=60000,
    train=dict(
        type='GOT10kDataset',
        ann_file=data_root +
        'got10k/annotations/got10k_train_infos.txt',
        img_prefix=data_root + 'got10k',
        pipeline=train_pipeline,
        split='train',
        test_mode=False),
    val=dict(
        type='GOT10kDataset',
        ann_file=data_root + 'got10k/annotations/got10k_test_infos.txt',
        img_prefix=data_root + 'got10k',
        pipeline=test_pipeline,
        split='test',
        test_mode=True),
    test=dict(
        type='GOT10kDataset',
        ann_file=data_root + 'got10k/annotations/got10k_test_infos.txt',
        img_prefix=data_root + 'got10k',
        pipeline=test_pipeline,
        split='test',
        test_mode=True))

new

train_dataloader = dict(
    batch_size=16,
    num_workers=4,
    persistent_workers=True,
    sampler=dict(type='QuotaSampler', samples_per_epoch=60000),
    dataset=dict(
        type='GOT10kDataset',
        data_root=data_root,
        ann_file='GOT10k/annotations/got10k_train_infos.txt',
        data_prefix=dict(img_path='GOT10k'),
        pipeline=train_pipeline,
        test_mode=False))
val_dataloader = dict(
    batch_size=1,
    num_workers=4,
    persistent_workers=True,
    drop_last=False,
    sampler=dict(type='VideoSampler'),
    dataset=dict(
        type='GOT10kDataset',
        data_root='data/',
        ann_file='GOT10k/annotations/got10k_test_infos.txt',
        data_prefix=dict(img_path='GOT10k'),
        pipeline=test_pipeline,
        test_mode=True))
test_dataloader = val_dataloader

keys of optimizer

old

optimizer = dict(
    type='AdamW',
    lr=0.0001,
    weight_decay=0.0001,
    paramwise_cfg=dict(
        custom_keys=dict(backbone=dict(lr_mult=0.1, decay_mult=1.0))))
optimizer_config = dict(grad_clip=dict(max_norm=0.1, norm_type=2))

new

optim_wrapper = dict(
    type='OptimWrapper',
    optimizer=dict(type='AdamW', lr=0.0001, weight_decay=0.0001),
    clip_grad=dict(max_norm=0.1, norm_type=2),
    paramwise_cfg=dict(
        custom_keys=dict(backbone=dict(lr_mult=0.1, decay_mult=1.0))))

keys of learning scheduler

old

lr_config = dict(policy='step', step=[400])

new

param_scheduler = dict(type='MultiStepLR', milestones=[400], gamma=0.1)

Model

Data preprocessor

The 1.xx versions of MMtracking add TrackDataPreprocessor. The data out from the data pipeline is transformed by this module and then fed into the model.

Train

The training forward of models and heads is performed by calling loss function in their respective classes. The arguments of loss function in models contain a dict of Tensor and a list of TrackDataSample.

Test

The test forward of models and heads is performed by calling predict function in their respective classes. The arguments of predict function in models contain a dict of Tensor and a list of TrackDataSample.

Data

data structure

The 1.xx versions of MMtracking add two new data structure: TrackDataSample and ReIDDataSample. These data structures wrap the annotations and predictions from one image (sequence) and are used as interfaces between different components.

dataset class

The 1.xx versions of MMTracking add two base dataset classes which inherient from the BaseDataset in MMEngine: BaseSOTDataset and BaseVideoDataset. The former is only used in SOT and the latter is used for all other tasks.

data pipeline

  1. Most of the transforms on image sequences in the old MMTracking are refactored in the latest MMTracking. Specifically, we use TransformBroadcaster to wrap the transforms of single image.

Some transforms on image sequences, such as SeqCropLikeStark, are reserved since TransformBroadcaster doesn't support setting different arguments respectively for each image in the sequence.

  1. We pack the VideoCollect, ConcatSameTypeFrames and SeqDefaultFormatBundle in the old MMTracking into PackTrackInputs in the latest MMTracking.

  2. The normalizaion in the pipeline in the old MMTracking is removed and this operation is implemented in the model forward.

data sampler

The 1.xx versions of MMtracking add DATA_SAMPLERS registry. You can customize different dataset samplers in the configs. Details about the samplers can be seen here.

Evaluation

The old versions of MMTarcking implement evaluation in the dataset class. In the 1.xx versions of MMTracking, we add METRICS registry. All evaluation are implemented in the metric classes registered in METRICS. Details can be seen here.

Visualization

The 1.xx versions of MMTracking add TrackLocalVisualizer and DetLocalVisualizer which are registered in VISUALIZER. Compared with the 0.xx versions of MMTracking, we support the visualization of images and feature maps. Details can be seen here.

Engine

The runner, hook, logging and optimizer in the training, evaluation and test are refactored in the 1.xx versions of MMTracking. Details can be seen in MMEngine.