[Docs] Train spatio temporal stream on custom coco dataset

[gguzzy]

The doc issue

I wanted to know how we can actually use an annotated dataset in coco format, containing specific keypoints annotations, to train it on a custom 2s-agcn both bone-motion and/or joint-motion. The issue I do not get is why we should use train_test_split to create train, test and val datasets and how I would change this `dataset_type = 'CocoDataset' ann_file_train = 'train.json' ann_file_val = 'test.json' ann_file_test = 'val.json' classes = ('Drunk', 'Not drunk') train_pipeline = [ dict(type='PreNormalize2D'), dict(type='GenSkeFeat', dataset='coco', feats=['bm']), dict(type='UniformSampleFrames', clip_len=100), dict(type='PoseDecode'), dict(type='FormatGCNInput', num_person=2), dict(type='PackActionInputs') ] val_pipeline = [ dict(type='PreNormalize2D'), dict(type='GenSkeFeat', dataset='coco', feats=['bm']), dict( type='UniformSampleFrames', clip_len=100, num_clips=1, test_mode=True), dict(type='PoseDecode'), dict(type='FormatGCNInput', num_person=2), dict(type='PackActionInputs') ] test_pipeline = [ dict(type='PreNormalize2D'), dict(type='GenSkeFeat', dataset='coco', feats=['bm']), dict( type='UniformSampleFrames', clip_len=100, num_clips=10, test_mode=True), dict(type='PoseDecode'), dict(type='FormatGCNInput', num_person=2), dict(type='PackActionInputs') ]

train_dataloader = dict( classes=classes, batch_size=16, num_workers=2, persistent_workers=True, sampler=dict(type='DefaultSampler', shuffle=True), dataset=dict( type='RepeatDataset', times=5, dataset=dict( type=dataset_type, ann_file=ann_file_train, pipeline=train_pipeline, classes=classes, # split='xsub_train'`

Considering that this code is not working since mmaction2 cannot find:

1. CocoDataset raising: not in the mmaction::dataset registry. Please check whether the value of CocoDataset is correct or it was registered as expected.

Is there any available script to autoconvert my annotations or should I do anything else to accomplish a custom training? I do not want to lose my annotations.

Thanks

Suggest a potential alternative/fix

Write a real notebook showing how to train a real custom dataset!!! All documentation on website is obsolete, ambiguous and not working. Many key components of training are missing, and makes no sense to create a custom class for a well-known annotations as Coco are.

Thanks.