mmaction2: RuntimeError while training slowfast for AVA custom classes

Hi @irvingzhang0512 , @ImSau and I are working on this as well. Please let us know if we can contribute to this in any way. We would be glad to do so. For a faster turnaround, we wanted to do an end to end training with a smaller dataset (like a couple of actions only), if possible

Thanks alot, That’s really helpful.

What if training with Custom Dataset like AVA with limited classes?

Some extra work is needed to training custom dataset, e.g. creating/modifying label_file/exclude_file/proposal_file, taking care of timestamp_start/timestamp_end/_FPS in AVADataset. I plan to create a related pr after my team finishes creating our own ava-like dataset(At least two months).

I was hoping that training custom classes could to help to train custom dataset along with ava dataset.

BTW, training custom classes will NOT always get better results. Slowonly training results could refer to #555

I happened to do some tests about your custom classes [11, 12, 14, 15, 17, 79, 80]. Training custom classes lead to WORSE results.

Can you please add config file for Slowfast likewise slowonly for custom classes?

because I think I am missing something in backbone.

No need to modify model codes. Modify num_classes and custom_classes like slowonly config.

# define custom_classes and num_classes
custom_classes = [11, 12, 14, 15, 17, 79, 80]
num_classes = len(custom_classes) + 1

# model setting
model = dict(
    type='FastRCNN',
    backbone=dict(
        type='ResNet3dSlowFast',
        pretrained=None,
        resample_rate=8,
        speed_ratio=8,
        channel_ratio=8,
        slow_pathway=dict(
            type='resnet3d',
            depth=50,
            pretrained=None,
            lateral=True,
            conv1_kernel=(1, 7, 7),
            dilations=(1, 1, 1, 1),
            conv1_stride_t=1,
            pool1_stride_t=1,
            inflate=(0, 0, 1, 1),
            spatial_strides=(1, 2, 2, 1)),
        fast_pathway=dict(
            type='resnet3d',
            depth=50,
            pretrained=None,
            lateral=False,
            base_channels=8,
            conv1_kernel=(5, 7, 7),
            conv1_stride_t=1,
            pool1_stride_t=1,
            spatial_strides=(1, 2, 2, 1))),
    roi_head=dict(
        type='AVARoIHead',
        bbox_roi_extractor=dict(
            type='SingleRoIExtractor3D',
            roi_layer_type='RoIAlign',
            output_size=8,
            with_temporal_pool=True),
        bbox_head=dict(
            type='BBoxHeadAVA',
            in_channels=2304,
            num_classes=num_classes,
            multilabel=True,
            dropout_ratio=0.5)))

train_cfg = dict(
    rcnn=dict(
        assigner=dict(
            type='MaxIoUAssignerAVA',
            pos_iou_thr=0.9,
            neg_iou_thr=0.9,
            min_pos_iou=0.9),
        sampler=dict(
            type='RandomSampler',
            num=32,
            pos_fraction=1,
            neg_pos_ub=-1,
            add_gt_as_proposals=True),
        pos_weight=1.0,
        debug=False))
test_cfg = dict(rcnn=dict(action_thr=0.00))

dataset_type = 'AVADataset'
data_root = 'data/ava/rawframes'
anno_root = 'data/ava/annotations'

ann_file_train = f'{anno_root}/ava_train_v2.1.csv'
ann_file_val = f'{anno_root}/ava_val_v2.1.csv'

exclude_file_train = f'{anno_root}/ava_train_excluded_timestamps_v2.1.csv'
exclude_file_val = f'{anno_root}/ava_val_excluded_timestamps_v2.1.csv'

label_file = f'{anno_root}/ava_action_list_v2.1.pbtxt'

proposal_file_train = (f'{anno_root}/ava_dense_proposals_train.FAIR.'
                       'recall_93.9.pkl')
proposal_file_val = f'{anno_root}/ava_dense_proposals_val.FAIR.recall_93.9.pkl'

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

train_pipeline = [
    dict(type='SampleAVAFrames', clip_len=32, frame_interval=2),
    dict(type='RawFrameDecode'),
    dict(type='RandomRescale', scale_range=(256, 320)),
    dict(type='RandomCrop', size=256),
    dict(type='Flip', flip_ratio=0.5),
    dict(type='Normalize', **img_norm_cfg),
    dict(type='FormatShape', input_format='NCTHW', collapse=True),
    # Rename is needed to use mmdet detectors
    dict(type='Rename', mapping=dict(imgs='img')),
    dict(type='ToTensor', keys=['img', 'proposals', 'gt_bboxes', 'gt_labels']),
    dict(
        type='ToDataContainer',
        fields=[
            dict(key=['proposals', 'gt_bboxes', 'gt_labels'], stack=False)
        ]),
    dict(
        type='Collect',
        keys=['img', 'proposals', 'gt_bboxes', 'gt_labels'],
        meta_keys=['scores', 'entity_ids'])
]
# The testing is w/o. any cropping / flipping
val_pipeline = [
    dict(type='SampleAVAFrames', clip_len=32, frame_interval=2),
    dict(type='RawFrameDecode'),
    dict(type='Resize', scale=(-1, 256)),
    dict(type='Normalize', **img_norm_cfg),
    dict(type='FormatShape', input_format='NCTHW', collapse=True),
    # Rename is needed to use mmdet detectors
    dict(type='Rename', mapping=dict(imgs='img')),
    dict(type='ToTensor', keys=['img', 'proposals']),
    dict(type='ToDataContainer', fields=[dict(key='proposals', stack=False)]),
    dict(
        type='Collect',
        keys=['img', 'proposals'],
        meta_keys=['scores', 'img_shape'],
        nested=True)
]

data = dict(
    videos_per_gpu=9,
    workers_per_gpu=4,
    val_dataloader=dict(videos_per_gpu=1),
    test_dataloader=dict(videos_per_gpu=1),
    train=dict(
        type=dataset_type,
        ann_file=ann_file_train,
        exclude_file=exclude_file_train,
        pipeline=train_pipeline,
        label_file=label_file,
        proposal_file=proposal_file_train,
        person_det_score_thr=0.9,
        num_classes=num_classes,
        custom_classes=custom_classes,
        data_prefix=data_root),
    val=dict(
        type=dataset_type,
        ann_file=ann_file_val,
        exclude_file=exclude_file_val,
        pipeline=val_pipeline,
        label_file=label_file,
        proposal_file=proposal_file_val,
        person_det_score_thr=0.9,
        num_classes=num_classes,
        custom_classes=custom_classes,
        data_prefix=data_root))
data['test'] = data['val']

optimizer = dict(type='SGD', lr=0.05652, momentum=0.9, weight_decay=0.00001)
# optimizer = dict(type='SGD', lr=0.1125, momentum=0.9, weight_decay=0.00001)
# this lr is used for 8 gpus

optimizer_config = dict(grad_clip=dict(max_norm=40, norm_type=2))
# learning policy

lr_config = dict(
    policy='step',
    step=[10, 15],
    warmup='linear',
    warmup_by_epoch=True,
    warmup_iters=5,
    warmup_ratio=0.05)
total_epochs = 20
checkpoint_config = dict(interval=1)
workflow = [('train', 1)]
evaluation = dict(interval=1, save_best='mAP@0.5IOU')
log_config = dict(
    interval=20, hooks=[
        dict(type='TextLoggerHook'),
    ])
dist_params = dict(backend='nccl')
log_level = 'INFO'
work_dir = ('./work_dirs/ava/'
            'slowfast_kinetics_pretrained_r50_4x16x1_20e_ava_rgb_custom')
load_from = ('https://download.openmmlab.com/mmaction/recognition/slowfast/'
             'slowfast_r50_4x16x1_256e_kinetics400_rgb/'
             'slowfast_r50_4x16x1_256e_kinetics400_rgb_20200704-bcde7ed7.pth')
resume_from = None
find_unused_parameters = False

have closed it. I just misunderstood.

yeah… But you can decrease the resample_rate and speed_ratio, and then you can decrease the clip_len. The performance may be worse.

Could you provide a longer backtrace?