config.py.example

data_source = {
    'MS1M': {
        'params': {
            'train_db': 'data/faces_emore/train_clean_reshuffled',
        },
        'num_classes': 85742,
        'train_samples_num': 5822653
    }
}

data_params = {
    'mean': 0.0,
    'std': 1.0,
    'shuffle': True,
    'rand_crop': True,
    'rand_mirror': True,
    'max_rotate_angle': 5,
    'contrast': 0.1,
    'brightness': 0.1,
    'saturation': 0.1,
    'hue': 0.1
}

test_params = {
    'dbs_root': 'data/faces_emore',
    'dbs': ['lfw', 'cfp_fp', 'agedb_30'],
}

teacher_params = {
    'insightface-resnet': {
        'type': 'insightface-resnet',
        'depth': 50,
        'input_type': 1,
        'output_type': 'G',
        'activation': 'prelu',
        'embedding_dim': 512,
        'init': ('experiments/MS1M__r50G/CosFace-1.0/2020-03-03--11-19-56/student-converted', 18)
    }
}

student_params = {
    'insightface-resnet': {
        'type': 'insightface-resnet',
        'depth': 18,
        'input_type': 1,
        'output_type': 'G',
        'activation': 'prelu',
        'embedding_dim': 512,
        'init': ()
    }
}

loss_params = {
    'classification': {
        'weight': 0.7,
        # combined-margin params (also used for knowledge distillation)
        'type': 'combined-margin',
        'margin_1': 1.0,
        'margin_2': 0.0,
        'margin_3': 0.35,
        'scale': 64.0,
        'student_init': (),
    },
    'HKD': {
        'weight': 0.3,
        'temperature': 4.0,
        'teacher_init': ('experiments/MS1M__r50G/CosFace-1.0/2020-03-03--11-19-56/student-clf-converted', 18)
    },
    'DarkRank': {
        'weight': 0.0,
        'metric_func': 'l2',
        'type': 'hard',     # 'hard' | 'soft'
        'alpha': 3.0,
        'beta': 3.0,
        'list_length': 4,
    },
    'RKD-D': {
        'weight': 0.0,
        'metric_func': 'l2',    # 'cs' | 'l2'
        'huber_delta': 1.0,
        'batch_normalize': False,
    },
    'RKD-A': {
        'weight': 0.0,
        'huber_delta': 1.0,
    },
    'PWR': {
        'weight': 100.0,
        'metric_func': 'l2',    # 'cs' | 'l2'
        'loss_type': 'diff_exp',    # 'diff_power' | 'diff_exp' | 'ranknet'
        # loss params
        'diff_margin_type': 'teacher_diff',    # 'const' | 'teacher_diff' | 'teacher_std'
        'diff_margin': 0.0,
        'diff_power': 1.0,
        'diff_exp_beta': 1.0,
        'ranknet_beta': 1.0
    }
}

db_name = 'MS1M'
teacher_type = 'insightface-resnet'
student_type = 'insightface-resnet'
config = {
    'data_source': data_source[db_name],
    'data_params': data_params,
    'test_params': test_params,

    'teacher_params': teacher_params[teacher_type],
    'student_params': student_params[student_type],
    'loss_params': loss_params,

    'experiment_dir': "experiments/MS1M__r50_to_t18G",

    'input_shape': (3, 112, 112),
    'batch_size': 92,       # batch size per gpu

    # Optimizer params
    'opt_params': {
        'lr_base': 0.1,
        'lr_factor': 0.1,
        'lr_epoch_step': 2,
        'num_epoch': 13,
        'momentum': 0.9,
        'wd': 0.0005
    },

    # Display and testing frequency params
    'display_period': 50,
    'test_period': 5000,

    # GPU devices
    'devices_id': [0, 1, 2, 3, 4, 5],
}