transrank_ucf_r3d_rotate_mlp_200e.py

# model settings
trans = ['1x', '2x', 'rev']
org_len = 48
tgt_len = 16
loss_cfg = dict(type='MultiMRLoss', num_trans=len(trans), replicas=len(trans), strategy='pair', margin=2.)

img_norm_cfg = dict(
    mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_bgr=False)
t_trans_pipeline = [
    dict(type='ColorJitter', brightness=(0.5,1.5), contrast=(0.5,1.5), saturation=(0.5,1.5), hue=(-0.1,0.1)),
    dict(type='Normalize', mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375])
]

sample_cfg = [
    dict(type='normal', clip_len=16, frame_interval=2),
    dict(type='normal', clip_len=16, frame_interval=4),
    dict(type='rev', clip_len=16, frame_interval=2),
]
aux_loss_cfg = dict(type='CrossEntropyLoss', loss_weight=0.5)

model = dict(
    type='TransReg',
    backbone=dict(type='r3d_18'),
    aux_cfg=dict(rotate=dict(num_classes=4, loss_cls=aux_loss_cfg)),
    t_trans=dict(
        type='T_Trans',
        tgt_len=tgt_len,
        trans=trans,
        pipeline=t_trans_pipeline,
        skip=True,
        t_aug=False),
    cls_head=dict(
        type='I3DHead',
        num_classes=len(trans),
        in_channels=512,
        spatial_type='avg',
        mlp=[128],
        mlp_bn=True,
        dropout_ratio=0.,
        loss_cls=loss_cfg,
        init_std=0.01))

# dataset settings
dataset_type = 'VideoDataset'

# data_root is the folder that contains ucf101 videos
data_root = '/home/dhd/data/ucf101'
# ann_file is a list of training / testing samples, in the format of: 
# path/to/video1.mp4 label_index
# path/to/video2.mp4 label_index
# ...
ann_file_train = f'data/ucf101/ucf101_train_1.txt'
ann_file_val = f'data/ucf101/ucf101_test_1.txt'

train_pipeline = [
    dict(type='DecordInit'),
    dict(type='CustomSampleFrames', sample_cfg=sample_cfg, with_taug=(0.8, 1.2)),
    dict(type='DecordDecode'),
    dict(type='RandomResizedCrop', clip_wise=True),
    dict(type='Resize', scale=(112, 112), keep_ratio=False),
    dict(type='Flip', flip_ratio=0.5),
    dict(type='RandomRotate', clip_wise=False),
    dict(type='FormatShape', input_format='NCTHW'),
    dict(type='ToTensor', keys=['imgs', 'label', 'rotate_label']),
    dict(type='Collect', keys=['imgs', 'label'], meta_keys=['rotate_label']),
]
val_pipeline = [
    dict(type='DecordInit'),
    dict(type='CustomSampleFrames', sample_cfg=sample_cfg, with_taug=(0.8, 1.2), test_mode=True),
    dict(type='DecordDecode'),
    dict(type='Resize', scale=(-1, 112)),
    dict(type='CenterCrop', crop_size=112),
    dict(type='RandomRotate', clip_wise=False),
    dict(type='Normalize', **img_norm_cfg),
    dict(type='FormatShape', input_format='NCTHW'),
    dict(type='ToTensor', keys=['imgs', 'label', 'rotate_label']),
    dict(type='Collect', keys=['imgs', 'label'], meta_keys=['rotate_label']),
]
test_pipeline = val_pipeline

data = dict(
    videos_per_gpu=8,
    workers_per_gpu=2,
    val_dataloader=dict(videos_per_gpu=1),
    test_dataloader=dict(videos_per_gpu=1),
    train=dict(
        type='RepeatDataset',
        times=10,
        dataset=dict(
            type=dataset_type,
            ann_file=ann_file_train,
            data_prefix=data_root,
            pipeline=train_pipeline)),
    val=dict(
        type='EvalTTrans',
        num_trans=len(trans),
        dataset=dict(
            type=dataset_type,
            ann_file=ann_file_val,
            data_prefix=data_root,
            pipeline=val_pipeline)))
data['test'] = data['val']
# optimizer
optimizer = dict(
    type='SGD', lr=0.1, momentum=0.9,
    weight_decay=0.0001)  # 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='CosineAnnealing', min_lr=0, by_epoch=False)
total_epochs = 20
checkpoint_config = dict(interval=1)
evaluation = dict(
    interval=1, metrics=['top_k_accuracy'], save_eval=True)
log_config = dict(
    interval=20,
    hooks=[
        dict(type='TextLoggerHook'),

    ])
# runtime settings
dist_params = dict(backend='nccl')
log_level = 'INFO'
work_dir = f'./work_dirs/transrank_ucf_r3d_rotate_mlp_20e'
load_from = None
resume_from = None
workflow = [('train', 1)]