applying mean teacher to my own dataset

[liangzimei]

hi, i have already achieve ~94% with 4000 labels on cifar10. But for my own three classification task, i have 160k labelled data and unlabelled data. i can not get expected results( worse than using the labelled data to train directly ). is lr strategy sensitive to it? here is my setting (finetuning on mobilenet-v1 and using 4 gpus): thanks in advance

defaults = {

    # Technical details
    'workers': 20,
    'checkpoint_epochs': 20,
    'evaluation_epochs': 5,

    # Data
    'dataset': 'my dataset',
    'train_subdir': 'train',
    'eval_subdir': 'test',

    # Data sampling
    'base_batch_size': 100,
    'base_labeled_batch_size': 50,

    # Architecture
    'arch': 'mnet1',

    # Costs
    'consistency_type': 'mse',
    'consistency_rampup': 5,
    'consistency': 20.0,
    'logit_distance_cost': .01,
    'weight_decay': 2e-4,

    # Optimization
    'lr_rampup': 0,
    'base_lr': 0.001,
    'nesterov': True,
}

[YilinLiu97]

Same problem here.

[fahad7033]

Did you fix the problem?

[liangzimei]

@fahad7033 i haven't done it in a long time, you can try to train more iters.

[jetoHui520]

hi, i have already achieve ~94% with 4000 labels on cifar10. But for my own three classification task, i have 160k labelled data and unlabelled data. i can not get expected results( worse than using the labelled data to train directly ). is lr strategy sensitive to it? here is my setting (finetuning on mobilenet-v1 and using 4 gpus): thanks in advance

defaults = {

    # Technical details
    'workers': 20,
    'checkpoint_epochs': 20,
    'evaluation_epochs': 5,

    # Data
    'dataset': 'my dataset',
    'train_subdir': 'train',
    'eval_subdir': 'test',

    # Data sampling
    'base_batch_size': 100,
    'base_labeled_batch_size': 50,

    # Architecture
    'arch': 'mnet1',

    # Costs
    'consistency_type': 'mse',
    'consistency_rampup': 5,
    'consistency': 20.0,
    'logit_distance_cost': .01,
    'weight_decay': 2e-4,

    # Optimization
    'lr_rampup': 0,
    'base_lr': 0.001,
    'nesterov': True,
}

So I wonder how to create my own datasets.......

[paozhuanis1]

I have the same problem about how to apply the mean teacher to my own dataset(especially for audio) and model.