def get_resnet_experiments(exp):
nb_branch = 2
if exp == 1:
params = [[64, 128, 256, 512]]
# params = [[100]]
# params = [[90]]
# params = [[70]]
# params = [[50]]
# params = [[30]]
cov_outputs = [512]
cov_mode = 'pmean'
else:
return
batch_size = 128
mode_list = [1]
cov_branch = 'o2t_no_wv'
early_stop = True
cov_regularizer = None
last_config_feature_maps = [128]
robust = False
regroup = False
cov_alpha = 0.2
# cov_beta = 0.1
cov_beta = 0.1
title = 'cifar10_resnet_{}'.format(cov_branch)
config = DCovConfig(params, mode_list, cov_outputs, cov_branch, cov_mode, early_stop, cov_regularizer,
nb_branch=nb_branch, last_conv_feature_maps=last_config_feature_maps, batch_size=batch_size,
exp=exp, epsilon=1e-5, title=title, robust=robust, cov_alpha=cov_alpha, regroup=regroup,
cov_beta=cov_beta)
return config
def get_resnet_batch_norm(exp):
nb_branch = 1
if exp == 1:
params = [[]]
# params = [[100]]
# params = [[90]]
# params = [[70]]
# params = [[50]]
# params = [[30]]
cov_outputs = [200]
cov_mode = 'pmean'
else:
return
batch_size = 128
mode_list = [1]
cov_branch = 'o2transform'
early_stop = True
cov_regularizer = None
last_config_feature_maps = []
robust = False
regroup = False
cov_alpha = 1
# cov_beta = 0.1
cov_beta = 0.7
title = 'cifar10_cv_covBeta_{}_{}'.format(cov_beta, cov_branch)
config = DCovConfig(params, mode_list, cov_outputs, cov_branch, cov_mode, early_stop, cov_regularizer,
nb_branch=nb_branch, last_conv_feature_maps=last_config_feature_maps, batch_size=batch_size,
exp=exp, epsilon=1e-5, title=title, robust=robust, cov_alpha=cov_alpha, regroup=regroup,
cov_beta=cov_beta)
return config
def get_cov_beta_cv(exp):
if exp == 1:
nb_branch = 1
params = [[257, 128, 64], ]
cov_outputs = [params[0][2]]
else:
return
batch_size = 128
cov_mode = 'pmean'
mode_list = [1]
cov_branch = 'o2transform'
early_stop = True
cov_regularizer = None
last_config_feature_maps = [512]
robust = True
regroup = False
cov_alpha = 0.75
# cov_beta = 0.1
cov_beta = 0.5
title = 'minc2500_cv_covBeta_{}_{}'.format(cov_beta, cov_branch)
config = DCovConfig(params, mode_list, cov_outputs, cov_branch, cov_mode, early_stop, cov_regularizer,
nb_branch=nb_branch, last_conv_feature_maps=last_config_feature_maps, batch_size=batch_size,
exp=exp, epsilon=1e-5, title=title, robust=robust, cov_alpha=cov_alpha, regroup=regroup,
cov_beta=cov_beta)
return config
def get_experiment_settings(exp=1):
cov_regularizer = None
nb_branch = 1
dropout = False
early_stop = True
if exp == 1:
params = [[], [64], [128], [100, 50], [128, 64], [512, 256],
[128, 128, 128]]
mode_list = [1]
cov_outputs = [128, 64, 47, 32]
cov_branch = 'o2transform'
cov_mode = 'channel'
dropout = False
elif exp == 2:
params = [[], [64], [128], [100, 50], [128, 64], [512, 256],
[128, 128, 128]]
mode_list = [1]
cov_outputs = [128, 64, 47, 32]
cov_branch = 'o2transform'
cov_mode = 'mean'
dropout = False
elif exp == 3:
params = [[], [64], [128], [100, 50], [128, 64], [512, 256],
[128, 128, 128]]
mode_list = [1]
cov_outputs = [128, 64, 47, 32]
cov_branch = 'o2transform'
cov_mode = 'mean'
dropout = True
elif exp == 4:
"""Test VGG16 with DCov-2 """
params = [[]]
mode_list = [1]
cov_outputs = [256]
cov_mode = 'mean'
cov_branch = 'o2transform'
early_stop = True
# cov_regularizer = 'Fob'
elif exp == 5:
""" Test LogTransform layers"""
params = [[64, 64]]
mode_list = [3]
cov_outputs = [64]
cov_mode = 'mean'
cov_branch = 'o2transform'
early_stop = True
else:
return
config = DCovConfig(params,
mode_list,
cov_outputs,
cov_branch,
cov_mode,
early_stop,
cov_regularizer,
dropout=dropout,
nb_branch=nb_branch)
return config
def get_log_experiment(exp):
cov_regularizer = None
nb_branch = 1
last_config_feature_maps = []
batch_size = 4
if exp == 1:
""" log test """
nb_branch = 1
# params = [[513, 513, 513], [256, 256, 256]]
# params = [[513, 513, 513, 513, 513, 513], [256, 256, 256, 256, 256]]
params = [[513, 257, 129], [513, 513, 513]]
# params = [[1024, 512], [1024, 512, 256], [512, 256]]
mode_list = [1]
cov_outputs = [128]
cov_mode = 'mean'
cov_branch = 'log'
early_stop = False
# cov_regularizer = 'Fob'
last_config_feature_maps = []
last_config_feature_maps = [1024, 512]
batch_size = 32
vectorization = 'wv'
elif exp == 2:
""" aaai paper test """
nb_branch = 1
# params = [[513, 513, 513], [256, 256, 256]]
# params = [[513, 513, 513, 513, 513, 513], [256, 256, 256, 256, 256]]
params = [[256, 128, 64], [513, 257, 129, 64]]
# params = [[1024, 512], [1024, 512, 256], [512, 256]]
mode_list = [1]
cov_outputs = [64]
cov_mode = 'mean'
cov_branch = 'log'
early_stop = False
# cov_regularizer = 'Fob'
last_config_feature_maps = []
last_config_feature_maps = [1024, 512]
batch_size = 32
vectorization = 'dense'
else:
return
config = DCovConfig(params,
mode_list,
cov_outputs,
cov_branch,
cov_mode,
early_stop,
cov_regularizer,
nb_branch=nb_branch,
last_conv_feature_maps=last_config_feature_maps,
batch_size=batch_size,
exp=exp,
vectorization=vectorization,
epsilon=1e-5)
return config
def get_ResNet_testing_ideas(exp):
""" Test VGG dimension reduction """
cov_regularizer = None
if exp == 1:
""" Experiment 1, cross validate number of branches. """
nb_branch = 2
# params = [[128, 64, 32], ]
params = [[257, 128, 64], ]
# params = [[64, 32, 16], ]
mode_list = [1]
# cov_outputs = [16]
cov_outputs = [params[0][2]]
cov_branch = 'o2t_no_wv'
cov_regularizer = None
# last_config_feature_maps = [512]
last_config_feature_maps = [1024]
concat = 'concat'
last_avg = False
robust = True
elif exp == 2:
""" Experiment 2, no robust """
nb_branch = 2
# params = [[128, 64, 32], ]
params = [[257, 128, 64], ]
# params = [[64, 32, 16], ]
mode_list = [1]
# cov_outputs = [16]
cov_outputs = [params[0][2]]
cov_branch = 'o2t_no_wv'
cov_regularizer = None
# last_config_feature_maps = [512]
last_config_feature_maps = [1024]
concat = 'concat'
last_avg = False
robust = False
else:
return
cov_mode = 'pmean'
early_stop = True
batch_size = 32
regroup = False
cov_alpha = 0.75
if robust:
rb = 'robost'
else:
rb = ''
title = 'minc2500_RsNTEST_{}_{}_LC{}_exp_{}_{}'.format(cov_branch, rb, last_config_feature_maps, exp, concat)
config = DCovConfig(params, mode_list, cov_outputs, cov_branch, cov_mode, early_stop, cov_regularizer,
nb_branch=nb_branch, last_conv_feature_maps=last_config_feature_maps, batch_size=batch_size,
exp=exp, epsilon=1e-5, title=title, robust=robust, cov_alpha=cov_alpha, regroup=regroup,
concat=concat,
)
return config
def get_von_with_regroup(exp=1):
cov_regularizer = None
nb_branch = 1
last_config_feature_maps = []
batch_size = 4
cov_alpha = 0.01
if exp == 1:
""" Test Multi_branch Resnet 50 with residual learning """
nb_branch = 4
params = [[100, 50, 25], ]
# params = [[512, 256, 128, 64], ]
mode_list = [1]
cov_outputs = [25]
cov_mode = 'mean'
cov_branch = 'o2transform'
early_stop = False
# cov_regularizer = 'vN'
cov_regularizer = None
last_config_feature_maps = []
# last_config_feature_maps = [1024]
batch_size = 32
robust = True
regroup = False
cov_alpha = 0.75
elif exp == 2:
""" Test Multi_branch Resnet 50 with residual learning """
nb_branch = 2
params = [[257, 128, 64], ]
mode_list = [1]
cov_outputs = [64]
cov_mode = 'mean'
cov_branch = 'o2transform'
early_stop = False
# cov_regularizer = 'vN'
cov_regularizer = None
last_config_feature_maps = []
# last_config_feature_maps = [1024]
batch_size = 32
robust = True
regroup = False
cov_alpha = 0.75
else:
return
if robust:
rb = 'robost'
else:
rb = ''
title = 'cifar10_von_{}_{}_{}_{}'.format(cov_mode, cov_branch, rb, cov_regularizer)
config = DCovConfig(params, mode_list, cov_outputs, cov_branch, cov_mode, early_stop, cov_regularizer,
nb_branch=nb_branch, last_conv_feature_maps=last_config_feature_maps, batch_size=batch_size,
exp=exp, epsilon=1e-5, title=title, robust=robust, cov_alpha=cov_alpha, regroup=regroup)
return config
def get_new_experiment(exp):
batch_size = 128
cov_mode = 'pmean'
mode_list = [1]
cov_branch = 'o2t_no_wv'
early_stop = False
cov_regularizer = None
last_config_feature_maps = [512]
robust = True
regroup = False
cov_alpha = 0.3
# cov_beta = 0.1
cov_beta = 0.1
pooling = 'max'
if exp == 1:
nb_branch = 2
params = [[257, 128, 64], ]
cov_outputs = [params[0][2]]
elif exp == 2:
nb_branch = 2
params = [[256, 512, 512], ]
cov_outputs = [params[0][2]]
elif exp == 3:
nb_branch = 2
params = [[256,128,64,32]]
cov_outputs = [params[0][2]]
elif exp == 4:
nb_branch = 2
params = [[512, 256, 128, 64,]]
cov_outputs = [params[0][2]]
last_config_feature_maps = [1024]
elif exp == 5:
nb_branch = 4
params = [[256, 128, 64, 32]]
cov_outputs = [params[0][2]]
elif exp == 6:
nb_branch = 2
params = [[256, 128, 64, 32]]
cov_outputs = [params[0][2]]
pooling = 'avg'
else:
return
title = 'mit_indoor_newexperiment_{}_{}_{}'.format(cov_mode, cov_branch, pooling)
# weight_path = '/home/kyu/.keras/models/mit_indoor_baseline_resnet50.weights'
weight_path = 'imagenet'
config = DCovConfig(params, mode_list, cov_outputs, cov_branch, cov_mode, early_stop, cov_regularizer,
nb_branch=nb_branch, last_conv_feature_maps=last_config_feature_maps, batch_size=batch_size,
exp=exp, epsilon=1e-5, title=title, robust=robust, cov_alpha=cov_alpha, regroup=regroup,
cov_beta=cov_beta,
weight_path=weight_path,
pooling=pooling)
return config
def get_constraints_settings(exp=1):
cov_regularizer = None
nb_branch = 1
last_config_feature_maps = []
batch_size = 4
cov_alpha = 0.01
if exp == 1:
""" Test Multi branch ResNet 50 """
nb_branch = 1
params = [
[256, 128, 64],
]
# params = [[1024, 512], [1024, 512, 256], [512, 256]]
mode_list = [3]
cov_outputs = [128]
cov_mode = 'mean'
cov_branch = 'o2transform'
early_stop = True
# cov_regularizer = None
cov_regularizer = 'vN'
# last_config_feature_maps = []
last_config_feature_maps = [1024, 512, 256]
batch_size = 32
robust = True
# cov_alpha = 0.75
else:
return
if robust:
rb = 'robost'
else:
rb = ''
title = 'dtd_cUN_{}_{}_{}_{}'.format(cov_mode, cov_branch, rb,
cov_regularizer)
config = DCovConfig(params,
mode_list,
cov_outputs,
cov_branch,
cov_mode,
early_stop,
cov_regularizer,
nb_branch=nb_branch,
last_conv_feature_maps=last_config_feature_maps,
batch_size=batch_size,
exp=exp,
epsilon=1e-5,
title=title,
robust=robust,
cov_alpha=cov_alpha)
return config
def get_tensorboard_test_setting(exp=1):
cov_regularizer = None
nb_branch = 1
last_config_feature_maps = []
batch_size = 4
if exp == 1:
""" Test Multi_branch Resnet 50 with residual learning """
nb_branch = 4
params = [
[257, 128, 64],
[257, 128, 128, 64],
]
mode_list = [1]
cov_outputs = [64, 32]
cov_mode = 'mean'
cov_branch = 'o2transform'
early_stop = True
# cov_regularizer = 'vN'
# last_config_feature_maps = []
last_config_feature_maps = [1024]
batch_size = 32
robust = True
cov_alpha = 0.75
else:
return
if robust:
rb = 'robost'
else:
rb = ''
title = 'dtd_von_{}_{}_{}_{}'.format(cov_mode, cov_branch, rb,
cov_regularizer)
config = DCovConfig(params,
mode_list,
cov_outputs,
cov_branch,
cov_mode,
early_stop,
cov_regularizer,
nb_branch=nb_branch,
last_conv_feature_maps=last_config_feature_maps,
batch_size=batch_size,
exp=exp,
epsilon=1e-5,
title=title,
robust=robust,
cov_alpha=cov_alpha)
return config
def get_aaai_experiment(exp):
cov_regularizer = None
nb_branch = 1
last_config_feature_maps = []
batch_size = 4
if exp == 1:
""" aaai paper test """
nb_branch = 1
# params = [[513, 513, 513], [256, 256, 256]]
# params = [[513, 513, 513, 513, 513, 513], [256, 256, 256, 256, 256]]
params = [[200, 100, 50]]
# params = [[1024, 512], [1024, 512, 256], [512, 256]]
mode_list = [1]
cov_outputs = [50]
cov_mode = 'pmean'
cov_branch = 'aaai'
early_stop = False
robust = False
# cov_regularizer = 'Fob'
# last_config_feature_maps = []
last_config_feature_maps = []
# last_config_feature_maps = [1024, 512, 256]
batch_size = 32
vectorization = 'dense'
elif exp == 2:
""" aaai paper test """
nb_branch = 2
# params = [[513, 513, 513], [256, 256, 256]]
# params = [[513, 513, 513, 513, 513, 513], [256, 256, 256, 256, 256]]
params = [[200, 100, 50]]
# params = [[1024, 512], [1024, 512, 256], [512, 256]]
mode_list = [1]
cov_outputs = [50]
cov_mode = 'mean'
cov_branch = 'aaai'
early_stop = False
robust = False
# cov_regularizer = 'Fob'
# last_config_feature_maps = []
last_config_feature_maps = [1024, 512]
batch_size = 32
vectorization = 'flatten'
else:
return
title = 'aaai_baseline'
config = DCovConfig(params, mode_list, cov_outputs, cov_branch, cov_mode, early_stop, cov_regularizer,
nb_branch=nb_branch, last_conv_feature_maps=last_config_feature_maps, batch_size=batch_size,
exp=exp, vectorization=vectorization, epsilon=1e-5, title=title, robust=robust)
return config
def get_residual_cov_experiment(exp):
cov_regularizer = None
nb_branch = 1
last_config_feature_maps = []
batch_size = 4
if exp == 1:
""" Test Residual learning of covariance branch """
nb_branch = 1
# params = [[513, 513, 513], [256, 256, 256]]
# params = [[513, 513, 513, 513, 513, 513], [256, 256, 256, 256, 256]]
params = [[65, 65, 65, 65]]
# params = [[1024, 512], [1024, 512, 256], [512, 256]]
mode_list = [1]
cov_outputs = [128]
cov_mode = 'pmean'
cov_branch = 'residual'
early_stop = True
# cov_regularizer = 'Fob'
last_config_feature_maps = []
# last_config_feature_maps = [1024, 512]
robust = False
elif exp == 2:
""" Test Multi branch ResNet 50 """
nb_branch = 4
params = [[513, 513, 513, 513, 513, 513], [513, 513, 513], [256, 256, 256]]
# params = [[1024, 512], [1024, 512, 256], [512, 256]]
mode_list = [1]
cov_outputs = [128]
cov_mode = 'pmean'
cov_branch = 'residual'
early_stop = True
# cov_regularizer = 'Fob'
last_config_feature_maps = []
# last_config_feature_maps = [1024]
robust = True
else:
return
if robust:
rb = 'robost'
else:
rb = ''
batch_size = 128
cov_alpha = 0.3
title = 'cifar10_{}_{}_{}_{}'.format(cov_mode, cov_branch, rb, cov_regularizer)
config = DCovConfig(params, mode_list, cov_outputs, cov_branch, cov_mode, early_stop, cov_regularizer,
nb_branch=nb_branch, last_conv_feature_maps=last_config_feature_maps, batch_size=batch_size,
exp=exp, title=title, robust=robust, cov_alpha=cov_alpha)
return config
def get_resnet_with_power(exp):
nb_branch = 1
if exp == 1:
params = [[64,128,256]]
# params = [[100]]
# params = [[90]]
# params = [[70]]
# params = [[50]]
# params = [[30]]
cov_outputs = [200]
cov_mode = 'pmean'
vectorization = 'wv'
elif exp == 2:
params = [[]]
# params = [[100]]
# params = [[90]]
# params = [[70]]
# params = [[50]]
# params = [[30]]
cov_outputs = [64]
cov_mode = 'pmean'
vectorization = 'mat_flatten'
else:
return
batch_size = 128
mode_list = [1]
cov_branch = 'pow_o2t'
early_stop = True
cov_regularizer = None
last_config_feature_maps = []
robust = False
regroup = False
cov_alpha = 0.2
# cov_beta = 0.1
cov_beta = 0.1
title = 'cifar10_cv_pow-o2t_{}_{}'.format(cov_beta, cov_branch)
config = DCovConfig(params, mode_list, cov_outputs, cov_branch, cov_mode, early_stop, cov_regularizer,
nb_branch=nb_branch, last_conv_feature_maps=last_config_feature_maps, batch_size=batch_size,
exp=exp, epsilon=1e-5, title=title, robust=robust, cov_alpha=cov_alpha, regroup=regroup,
cov_beta=cov_beta,
vectorization=vectorization)
return config
def get_iccv_experiment(exp):
cov_regularizer = None
if exp == 1:
""" aaai paper test """
nb_branch = 1
params = [
[],
]
mode_list = [1]
cov_outputs = [256]
cov_mode = 'channel'
cov_branch = 'pow_o2t'
early_stop = False
robust = False
# cov_regular izer = 'Frob'
# last_config_feature_maps = []
last_config_feature_maps = [256]
batch_size = 32
vectorization = 'mat_flatten'
else:
return
title = 'iccv_issecond_baseline'
config = DCovConfig(params,
mode_list,
cov_outputs,
cov_branch,
cov_mode,
early_stop,
cov_regularizer,
nb_branch=nb_branch,
last_conv_feature_maps=last_config_feature_maps,
batch_size=batch_size,
exp=exp,
vectorization=vectorization,
epsilon=1e-5,
title=title,
robust=robust)
return config
def get_matrix_bp(exp=1):
if exp == 1:
""" Test get matrix bp learning """
nb_branch = 1
mode_list = [1]
elif exp == 2:
""" Test get matrix back prop with multi branch """
nb_branch = 1
mode_list = [1]
elif exp == 3:
""" Test Multi_branch Resnet 50 with residual learning """
nb_branch = 4
mode_list = [3]
else:
return
params = [[]]
cov_outputs = [64]
cov_mode = 'channel'
cov_branch = 'matbp'
early_stop = True
cov_regularizer = None
last_config_feature_maps = []
batch_size = 128
robust = False
regroup = False
cov_alpha = 0.75
concat = 'concat'
if robust:
rb = 'robost'
else:
rb = ''
title = 'cifar10_matbp_basline_{}_{}_{}_{}'.format(cov_mode, cov_branch, rb, cov_regularizer)
config = DCovConfig(params, mode_list, cov_outputs, cov_branch, cov_mode, early_stop, cov_regularizer,
nb_branch=nb_branch, last_conv_feature_maps=last_config_feature_maps, batch_size=batch_size,
exp=exp, epsilon=1e-5, title=title, robust=robust, cov_alpha=cov_alpha, regroup=regroup,
concat=concat, normalization=None)
return config
def get_residual_cov_experiment(exp):
cov_regularizer = None
nb_branch = 1
last_config_feature_maps = []
batch_size = 4
if exp == 1:
""" Test Multi branch ResNet 50 """
nb_branch = 1
params = [
[
257,
257,
257,
],
[257, 257, 257, 257, 257, 257],
]
# params = [[1024, 512], [1024, 512, 256], [512, 256]]
mode_list = [1]
cov_outputs = [128]
cov_mode = 'mean'
cov_branch = 'residual'
early_stop = True
# cov_regularizer = 'Fob'
last_config_feature_maps = []
last_config_feature_maps = [1024, 512, 256]
batch_size = 32
robust = False
elif exp == 2:
""" Test Multi_branch Resnet 50 with residual learning """
nb_branch = 2
params = [
[257, 257, 257],
]
mode_list = [1]
cov_outputs = [64]
cov_mode = 'mean'
cov_branch = 'residual'
early_stop = True
# cov_regularizer = 'Fob'
last_config_feature_maps = []
last_config_feature_maps = [1024]
batch_size = 32
robust = True
cov_alpha = 0.75
else:
return
if robust:
rb = 'robost'
else:
rb = ''
title = 'dtd_von_{}_{}_{}_{}'.format(cov_mode, cov_branch, rb,
cov_regularizer)
config = DCovConfig(params,
mode_list,
cov_outputs,
cov_branch,
cov_mode,
early_stop,
cov_regularizer,
nb_branch=nb_branch,
last_conv_feature_maps=last_config_feature_maps,
batch_size=batch_size,
exp=exp,
robust=robust,
cov_alpha=cov_alpha,
title=title)
return config
def get_ResNet_testing_ideas(exp):
""" Test VGG dimension reduction """
cov_regularizer = None
normalization = 'mean'
cov_mode = 'pmean'
if exp == 1:
""" Experiment 1, cross validate number of branches. """
nb_branch = 2
# params = [[128, 64, 32], ]
params = [
[257, 128, 64, 32],
]
# params = [[64, 32, 16], ]
mode_list = [1]
# cov_outputs = [16]
cov_outputs = [params[0][3]]
cov_branch = 'o2t_no_wv'
cov_regularizer = None
# last_config_feature_maps = [512]
last_config_feature_maps = [1024]
concat = 'concat'
last_avg = False
robust = True
elif exp == 2:
""" Experiment 2, robust """
nb_branch = 2
# params = [[128, 64, 32], ]
params = [
[257, 128, 64],
]
# params = [[64, 32, 16], ]
mode_list = [1]
# cov_outputs = [16]
cov_outputs = [params[0][2]]
cov_branch = 'o2t_no_wv'
cov_regularizer = None
# last_config_feature_maps = [512]
last_config_feature_maps = [1024]
concat = 'concat'
last_avg = False
robust = True
elif exp == 3:
""" Experiment 3 June 27 2017, no 1x1 and train from scratch """
nb_branch = 2
# params = [[128, 64, 32], ]
params = [
[257, 128, 64],
]
# params = [[64, 32, 16], ]
mode_list = [1]
# cov_outputs = [16]
cov_outputs = [params[0][2]]
cov_branch = 'multiple_o2t'
cov_regularizer = None
# last_config_feature_maps = [512]
last_config_feature_maps = [1024]
concat = 'concat'
last_avg = False
robust = False
elif exp == 4:
nb_branch = 2
params = [
[],
]
# params = [[256,]]
mode_list = [1]
cov_outputs = [128]
last_config_feature_maps = [1024]
cov_branch = 'o2transform'
normalization = 'mean'
# cov_branch = 'o2t_no_wv'
cov_regularizer = None
concat = 'concat'
robust = True
elif exp == 5:
""" 2017.8.15 New PV test with l2 normalization and sqrt after PV """
nb_branch = 1
params = [
[128],
]
mode_list = [1]
cov_outputs = [128]
last_config_feature_maps = [1024]
cov_branch = 'new_wv'
normalization = 'mean'
cov_regularizer = None
concat = 'concat'
robust = False
cov_mode = 'channel'
else:
return
early_stop = False
batch_size = 32
finetune_batch_size = 128
regroup = False
cov_alpha = 0.3
cov_beta = 0.1
if robust:
rb = 'robost'
else:
rb = ''
title = 'sun397_ResNetTest_{}_{}_LC{}_exp_{}_{}'.format(
cov_branch, rb, last_config_feature_maps, exp, concat)
config = DCovConfig(
params,
mode_list,
cov_outputs,
cov_branch,
cov_mode,
early_stop,
cov_regularizer,
nb_branch=nb_branch,
last_conv_feature_maps=last_config_feature_maps,
batch_size=batch_size,
exp=exp,
epsilon=1e-5,
title=title,
robust=robust,
cov_alpha=cov_alpha,
cov_beta=cov_beta,
regroup=regroup,
concat=concat,
weight_path='imagenet',
finetune_batch_size=finetune_batch_size,
normalization=normalization,
)
return config
def get_von_settings(exp=1):
cov_regularizer = None
nb_branch = 1
last_config_feature_maps = []
batch_size = 4
cov_alpha = 0.01
if exp == 1:
""" Test Multi branch ResNet 50 """
nb_branch = 1
params = [[256, 128, 64],]
# params = [[1024, 512], [1024, 512, 256], [512, 256]]
mode_list = [1]
cov_outputs = [128]
cov_mode = 'mean'
cov_branch = 'o2transform'
early_stop = False
# cov_regularizer = None
# cov_regularizer = 'vN'
# last_config_feature_maps = []
last_config_feature_maps = [1024, 512, 256]
batch_size = 32
robust = True
cov_alpha = 0.75
elif exp == 2:
""" Test Multi_branch Resnet 50 with residual learning """
nb_branch = 4
params = [[257, 128, 64], ]
mode_list = [1]
cov_outputs = [64]
cov_mode = 'mean'
cov_branch = 'o2transform'
early_stop = False
cov_regularizer = 'vN'
# last_config_feature_maps = []
last_config_feature_maps = [1024]
batch_size = 32
robust = False
elif exp == 3:
""" Test Multi branch ResNet 50 """
nb_branch = 1
params = [[256, 128, 64], ]
# params = [[1024, 512], [1024, 512, 256], [512, 256]]
mode_list = [1]
cov_outputs = [128]
cov_mode = 'mean'
cov_branch = 'o2transform'
early_stop = False
cov_regularizer = None
# cov_regularizer = 'vN'
# last_config_feature_maps = []
last_config_feature_maps = [1024, 512, 256]
batch_size = 32
robust = True
cov_alpha = 0.75
elif exp == 4:
""" Test Multi_branch Resnet 50 with residual learning """
nb_branch = 2
params = [[257, 128, 64], ]
mode_list = [1]
cov_outputs = [64]
cov_mode = 'pmean'
cov_branch = 'o2transform'
early_stop = False
cov_regularizer = None
# cov_regularizer = 'vN'
# last_config_feature_maps = []
last_config_feature_maps = [1024]
batch_size = 32
robust = True
cov_alpha = 0.75
else:
return
if robust:
rb = 'robost'
else:
rb = ''
title = 'minc_von_{}_{}_{}_{}'.format(cov_mode, cov_branch, rb, cov_regularizer)
config = DCovConfig(params, mode_list, cov_outputs, cov_branch, cov_mode, early_stop, cov_regularizer,
nb_branch=nb_branch, last_conv_feature_maps=last_config_feature_maps, batch_size=batch_size,
exp=exp, epsilon=1e-5, title=title, robust=robust, cov_alpha=cov_alpha)
return config
def get_VGG_dimension_reduction(exp=1):
""" Test VGG dimension reduction """
cov_regularizer = None
if exp == 1:
nb_branch = 2
params = [[257, 128, 64], ]
mode_list = [1]
cov_outputs = [64]
cov_branch = 'o2transform'
cov_regularizer = None
last_config_feature_maps = [1024]
concat = 'concat'
elif exp == 2:
nb_branch = 2
params = [[257, 128, 64], ]
mode_list = [1]
cov_outputs = [64]
cov_branch = 'o2transform'
# last_config_feature_maps = []
last_config_feature_maps = [1024]
concat = 'concat'
elif exp == 3:
nb_branch = 4
params = [[257, 128, 64], ]
mode_list = [1]
cov_outputs = [64]
cov_branch = 'o2transform'
# last_config_feature_maps = []
last_config_feature_maps = [1024]
concat = 'concat'
elif exp == 4:
nb_branch = 2
params = [[257, 128, 64], ]
mode_list = [1]
cov_outputs = [64]
cov_branch = 'o2transform'
cov_regularizer = None
last_config_feature_maps = [1024]
concat = 'sum'
elif exp == 5:
nb_branch = 2
params = [[257, 128, 64], ]
mode_list = [1]
cov_outputs = [64]
cov_branch = 'o2t_no_wv'
# last_config_feature_maps = []
last_config_feature_maps = [1024]
concat = 'sum'
elif exp == 6:
nb_branch = 2
params = [[257, 128, 64], ]
mode_list = [1]
cov_outputs = [64]
cov_branch = 'o2t_no_wv'
# last_config_feature_maps = []
last_config_feature_maps = [1024]
concat = 'avg'
else:
return
cov_mode = 'pmean'
early_stop = True
batch_size = 32
robust = True
regroup = False
cov_alpha = 0.75
if robust:
rb = 'robost'
else:
rb = ''
title = 'minc2500_DR_{}_{}_LC{}'.format(cov_mode, rb, last_config_feature_maps)
config = DCovConfig(params, mode_list, cov_outputs, cov_branch, cov_mode, early_stop, cov_regularizer,
nb_branch=nb_branch, last_conv_feature_maps=last_config_feature_maps, batch_size=batch_size,
exp=exp, epsilon=1e-5, title=title, robust=robust, cov_alpha=cov_alpha, regroup=regroup,
concat=concat)
return config
def get_matrix_bp(exp=1):
cov_regularizer = None
nb_branch = 1
last_config_feature_maps = []
batch_size = 4
cov_alpha = 0.01
if exp == 1:
""" Test get matrix bp learning """
nb_branch = 1
params = [[]]
mode_list = [1]
cov_outputs = [64]
cov_mode = 'channel'
cov_branch = 'matbp'
early_stop = False
# cov_regularizer = 'vN'
cov_regularizer = None
# last_config_feature_maps = []
last_config_feature_maps = [256]
batch_size = 32
robust = False
regroup = False
cov_alpha = 0.75
concat = 'concat'
elif exp == 2:
""" Test get matrix back prop with multi branch """
nb_branch = 2
params = [[]]
mode_list = [1]
cov_outputs = [64]
cov_mode = 'mean'
cov_branch = 'matbp'
early_stop = False
# cov_regularizer = 'vN'
cov_regularizer = None
# last_config_feature_maps = []
last_config_feature_maps = [1024]
batch_size = 32
robust = True
regroup = False
cov_alpha = 0.75
concat = 'concat'
elif exp == 3:
""" Test Multi_branch Resnet 50 with residual learning """
nb_branch = 4
params = [
[257, 128, 64],
]
mode_list = [3]
cov_outputs = [64]
cov_mode = 'mean'
cov_branch = 'matbp'
early_stop = False
# cov_regularizer = 'vN'
cov_regularizer = None
last_config_feature_maps = []
# last_config_feature_maps = [1024]
batch_size = 32
robust = True
regroup = False
cov_alpha = 0.75
concat = 'concat'
else:
return
if robust:
rb = 'robost'
else:
rb = ''
title = 'dtd_baseline_matbp_von_{}_{}_{}_{}'.format(
cov_mode, cov_branch, rb, cov_regularizer)
weight_path = '/home/kyu/.keras/models/tmp/VGG16_o2_para-mode_1_matbp_784_finetune.weights'
config = DCovConfig(
params,
mode_list,
cov_outputs,
cov_branch,
cov_mode,
early_stop,
cov_regularizer,
nb_branch=nb_branch,
last_conv_feature_maps=last_config_feature_maps,
batch_size=batch_size,
exp=exp,
epsilon=1e-5,
title=title,
robust=robust,
cov_alpha=cov_alpha,
regroup=regroup,
concat=concat,
weight_path=weight_path,
)
return config
def get_experiment_settings(exp=1):
cov_regularizer = None
nb_branch = 1
batch_size = 32
last_config_feature_maps = []
if exp == 1:
params = [[], [1024], [512], [1024, 512], [512, 256], [2048, 1444],
[2048, 1024, 512]]
mode_list = [1]
cov_outputs = [512, 256, 128, 64]
cov_branch = 'o2transform'
cov_mode = 'channel'
early_stop = True
elif exp == 2:
params = [[1024], [512], [1024, 512], [512, 256], [2048, 1444],
[2048, 1024, 512]]
mode_list = [1]
cov_outputs = [512, 256, 128, 64]
cov_branch = 'o2transform'
cov_mode = 'mean'
early_stop = True
elif exp == 3:
"""Test the Regularizer """
params = [[]]
mode_list = [1]
cov_outputs = [512, 256, 128]
cov_mode = 'channel'
cov_branch = 'o2transform'
early_stop = True
cov_regularizer = 'Fob'
elif exp == 4:
"""Test VGG16 with DCov-2 """
params = [[]]
mode_list = [1]
cov_outputs = [256]
cov_mode = 'mean'
cov_branch = 'o2transform'
early_stop = True
# cov_regularizer = 'Fob'
elif exp == 5:
""" Test LogTransform layers"""
params = [[]]
mode_list = [3]
cov_outputs = [256]
cov_mode = 'mean'
cov_branch = 'o2transform'
early_stop = True
elif exp == 6:
""" Test ResNet 50 """
params = [[256, 128]]
mode_list = [1]
cov_outputs = [128]
cov_mode = 'mean'
cov_branch = 'o2transform'
early_stop = True
# cov_regularizer = 'Fob'
last_config_feature_maps = [1024, 512]
# last_config_feature_maps = []
elif exp == 7:
""" Test ResNet 50 with different branches """
nb_branch = 4
params = [[256, 128]]
mode_list = [1]
cov_outputs = [128]
cov_mode = 'pmean'
# cov_mode = 'mean'
cov_branch = 'o2transform'
early_stop = True
# cov_regularizer = 'Fob'
last_config_feature_maps = []
# last_config_feature_maps = []
batch_size = 16
else:
return
config = DCovConfig(params,
mode_list,
cov_outputs,
cov_branch,
cov_mode,
early_stop,
cov_regularizer,
nb_branch=nb_branch,
last_conv_feature_maps=last_config_feature_maps,
batch_size=batch_size)
return config
def get_experiment_settings(exp=1):
cov_regularizer = None
nb_branch = 1
last_config_feature_maps = []
batch_size = 4
if exp == 1:
params = [[], [1024], [512], [1024, 512], [512, 256], [2048, 1444], [2048, 1024, 512]]
mode_list = [1]
cov_outputs = [512, 256, 128, 64]
cov_branch = 'o2transform'
cov_mode = 'channel'
early_stop = True
elif exp == 2:
params = [[1024], [512], [1024, 512], [512, 256], [2048, 1444], [2048, 1024, 512]]
mode_list = [1]
cov_outputs = [512, 256, 128, 64]
cov_branch = 'o2transform'
cov_mode = 'mean'
early_stop = True
elif exp == 3:
"""Test the Regularizer """
params = [[]]
mode_list = [1]
cov_outputs = [512, 256, 128]
cov_mode = 'channel'
cov_branch = 'o2transform'
early_stop = True
cov_regularizer = 'Fob'
elif exp == 4:
"""Test VGG16 with DCov-2 """
params = [[256, 128]]
mode_list = [1]
cov_outputs = [128]
cov_mode = 'mean'
cov_branch = 'o2transform'
early_stop = True
# cov_regularizer = 'Fob'
# last_config_feature_maps = [256]
batch_size = 32
elif exp == 5:
""" Test ResNet 50 """
params = [[256, 128], [256, 128, 64], [256], [128], [256, 256, 128],]
# params = [[1024, 512], [1024, 512, 256], [512, 256]]
mode_list = [1]
cov_outputs = [128]
cov_mode = 'mean'
cov_branch = 'o2transform'
early_stop = True
# cov_regularizer = 'Fob'
# last_config_feature_maps = []
last_config_feature_maps = [1024, 512]
batch_size = 32
elif exp == 6:
""" Test Multi branch ResNet 50 """
# nb_branch = 4
nb_branch = 8
params = [[256, 128], [256, 128, 64],]
# params = [[1024, 512], [1024, 512, 256], [512, 256]]
mode_list = [1]
cov_outputs = [128, 64]
cov_mode = 'mean'
cov_branch = 'o2transform'
early_stop = True
# cov_regularizer = 'Fob'
last_config_feature_maps = []
# last_config_feature_maps = [1024, 512]
batch_size = 32
else:
return
config = DCovConfig(params, mode_list, cov_outputs, cov_branch, cov_mode, early_stop, cov_regularizer,
nb_branch=nb_branch, last_conv_feature_maps=last_config_feature_maps, batch_size=batch_size,
exp=exp)
return config
def get_matrix_bp_vgg(exp=1):
if exp == 1:
""" Test get matrix bp learning """
nb_branch = 1
params = [[]]
mode_list = [1]
cov_outputs = [64]
cov_mode = 'mean'
cov_branch = 'matbp'
early_stop = False
# cov_regularizer = 'vN'
cov_regularizer = None
last_config_feature_maps = [512]
# last_config_feature_maps = [1024, 512]
batch_size = 32
robust = False
regroup = False
cov_alpha = 0.75
concat = 'concat'
elif exp == 2:
""" Test get matrix back prop with multi branch """
nb_branch = 2
params = [[]]
mode_list = [1]
cov_outputs = [64]
cov_mode = 'mean'
cov_branch = 'matbp'
early_stop = False
# cov_regularizer = 'vN'
cov_regularizer = None
# last_config_feature_maps = []
last_config_feature_maps = [1024]
batch_size = 32
robust = True
regroup = False
cov_alpha = 0.75
concat = 'concat'
elif exp == 3:
""" Test Multi_branch Resnet 50 with residual learning """
nb_branch = 4
params = [[257, 128, 64], ]
mode_list = [3]
cov_outputs = [64]
cov_mode = 'mean'
cov_branch = 'matbp'
early_stop = False
# cov_regularizer = 'vN'
cov_regularizer = None
last_config_feature_maps = []
# last_config_feature_maps = [1024]
batch_size = 32
robust = True
regroup = False
cov_alpha = 0.75
concat = 'concat'
else:
return
if robust:
rb = 'robost'
else:
rb = ''
title = 'minc_baseline_{}_{}_{}_{}'.format(cov_mode, cov_branch, rb, cov_regularizer)
config = DCovConfig(params, mode_list, cov_outputs, cov_branch, cov_mode, early_stop, cov_regularizer,
nb_branch=nb_branch, last_conv_feature_maps=last_config_feature_maps, batch_size=batch_size,
exp=exp, epsilon=1e-5, title=title, robust=robust, cov_alpha=cov_alpha, regroup=regroup,
concat=concat, normalization=None)
return config
def get_different_concat(exp=1):
cov_regularizer = None
nb_branch = 1
last_config_feature_maps = []
batch_size = 4
cov_alpha = 0.01
if exp == 1:
""" Test Multi_branch Resnet 50 with residual learning """
nb_branch = 4
params = [
[257, 128, 64],
]
mode_list = [1]
cov_outputs = [64]
cov_mode = 'pmean'
cov_branch = 'o2t_no_wv'
early_stop = False
# cov_regularizer = 'vN'
cov_regularizer = None
# last_config_feature_maps = []
last_config_feature_maps = []
batch_size = 10
robust = True
regroup = False
cov_alpha = 0.75
concat = 'matrix_diag'
elif exp == 2:
""" Test Multi_branch Resnet 50 with residual learning """
nb_branch = 2
params = [
[257, 128, 64],
]
mode_list = [1]
cov_outputs = [64]
cov_mode = 'pmean'
cov_branch = 'o2t_no_wv'
early_stop = False
# cov_regularizer = 'vN'
cov_regularizer = None
# last_config_feature_maps = []
last_config_feature_maps = [1024]
batch_size = 32
robust = True
regroup = False
cov_alpha = 0.75
concat = 'matrix_diag'
elif exp == 3:
""" Test Multi_branch Resnet 50 with residual learning """
nb_branch = 2
params = [
[257, 128, 64],
]
mode_list = [1]
cov_outputs = [64]
cov_mode = 'pmean'
cov_branch = 'o2t_no_wv'
early_stop = False
# cov_regularizer = 'vN'
cov_regularizer = None
# last_config_feature_maps = []
last_config_feature_maps = [1024]
batch_size = 32
robust = True
regroup = False
cov_alpha = 0.75
concat = 'sum'
else:
return
if robust:
rb = 'robost'
else:
rb = ''
title = 'dtd_diagcc_{}_{}_{}_{}'.format(cov_mode, cov_branch, rb,
cov_regularizer)
config = DCovConfig(params,
mode_list,
cov_outputs,
cov_branch,
cov_mode,
early_stop,
cov_regularizer,
nb_branch=nb_branch,
last_conv_feature_maps=last_config_feature_maps,
batch_size=batch_size,
exp=exp,
epsilon=1e-5,
title=title,
robust=robust,
cov_alpha=cov_alpha,
regroup=regroup,
concat=concat)
return config
def get_VGG_correlation(exp):
cov_regularizer = None
if exp == 1:
""" Test Multiple o2t with so bn"""
nb_branch = 2
params = [
[257, 128, 64, 64],
]
mode_list = [1]
cov_outputs = [64]
cov_branch = 'o2transform'
cov_regularizer = None
last_config_feature_maps = [512]
# last_config_feature_maps = [1024]
concat = 'concat'
last_avg = False
robust = True
so_mode = 2
elif exp == 2:
""" Test Multiple o2t with so bn"""
nb_branch = 2
# params = [[64, 64, 32], ]
# params = [[257, 128, 64], ]
params = [
[257, 128, 64, 64],
]
mode_list = [1]
cov_outputs = [64]
cov_branch = 'corr'
cov_regularizer = None
last_config_feature_maps = [512]
# last_config_feature_maps = [1024]
concat = 'concat'
last_avg = False
robust = True
so_mode = 2
else:
return
cov_mode = 'pmean'
early_stop = False
batch_size = 32
regroup = False
cov_alpha = 0.75
if robust:
rb = 'robost'
else:
rb = ''
title = 'sun397_VGG_Corr_{}_{}_LC{}_exp_{}_{}'.format(
cov_branch, rb, last_config_feature_maps, exp, concat)
config = DCovConfig(params,
mode_list,
cov_outputs,
cov_branch,
cov_mode,
early_stop,
cov_regularizer,
nb_branch=nb_branch,
last_conv_feature_maps=last_config_feature_maps,
batch_size=batch_size,
exp=exp,
epsilon=1e-5,
title=title,
robust=robust,
cov_alpha=cov_alpha,
regroup=regroup,
concat=concat,
vectorization=None,
load=False,
so_mode=so_mode)
return config
def get_VGG_testing_ideas(exp):
""" Test VGG dimension reduction """
cov_regularizer = None
if exp == 1:
""" Experiment 1, cross validate number of branches. """
nb_branch = 2
# params = [[128, 64, 32], ]
params = [
[257, 128, 64],
]
# params = [[64, 32, 16], ]
mode_list = [1]
# cov_outputs = [16]
cov_outputs = [params[0][2]]
cov_branch = 'o2t_no_wv'
cov_regularizer = None
last_config_feature_maps = [512]
# last_config_feature_maps = [1024]
concat = 'concat'
last_avg = False
robust = True
elif exp == 2:
""" Cross Validate on the summing methods """
nb_branch = 2
params = [
[257, 128, 64],
]
mode_list = [1]
cov_outputs = [64]
cov_branch = 'o2transform'
cov_regularizer = None
last_config_feature_maps = [512]
# last_config_feature_maps = [1024]
concat = 'ave'
last_avg = False
robust = False
elif exp == 3:
""" Test with robust single branch"""
nb_branch = 1
params = [
[257, 128, 64],
]
mode_list = [1]
cov_outputs = [64]
cov_branch = 'o2transform'
cov_regularizer = None
# last_config_feature_maps = []
last_config_feature_maps = [512]
concat = 'concat'
last_avg = False
robust = True
elif exp == 4:
""" Cross Validate on the summing methods for Riemannian"""
nb_branch = 2
# params = [[128, 64, 32],]
params = [
[257, 128, 64],
]
mode_list = [1]
cov_outputs = [64]
cov_branch = 'o2t_no_wv'
cov_regularizer = None
last_config_feature_maps = [512]
# last_config_feature_maps = [1024]
concat = 'concat'
last_avg = False
robust = True
elif exp == 5:
""" Cross Validate on the summing methods for Riemannian"""
nb_branch = 2
params = [
[64, 64, 32],
]
# params = [[257, 128, 64], ]
mode_list = [1]
cov_outputs = [64]
cov_branch = 'multiple_o2t'
cov_regularizer = None
last_config_feature_maps = [512]
# last_config_feature_maps = [1024]
concat = 'concat'
last_avg = False
robust = True
elif exp == 6:
""" Test Multiple o2t with so bn"""
nb_branch = 2
params = [
[64, 64, 32],
]
# params = [[257, 128, 64], ]
mode_list = [1]
cov_outputs = [64]
cov_branch = 'sobn_multiple_o2t'
cov_regularizer = None
last_config_feature_maps = [512]
# last_config_feature_maps = [1024]
concat = 'concat'
last_avg = False
robust = True
elif exp == 7:
""" Test the Multi-branch with FPN network (Kaiming CVPR) """
nb_branch = 2
params = [[256, 128, 64, 64]]
mode_list = [4]
cov_outputs = [64]
cov_branch = 'o2t_no_wv'
cov_regularizer = None
last_config_feature_maps = [512]
concat = 'concat'
robust = True
else:
return
cov_mode = 'pmean'
early_stop = False
batch_size = 32
regroup = False
cov_alpha = 0.3
if robust:
rb = 'robost'
else:
rb = ''
title = 'sun397_VGGTEST_{}_{}_LC{}_exp_{}_{}'.format(
cov_branch, rb, last_config_feature_maps, exp, concat)
config = DCovConfig(params,
mode_list,
cov_outputs,
cov_branch,
cov_mode,
early_stop,
cov_regularizer,
nb_branch=nb_branch,
last_conv_feature_maps=last_config_feature_maps,
batch_size=batch_size,
exp=exp,
epsilon=1e-5,
title=title,
robust=robust,
cov_alpha=cov_alpha,
regroup=regroup,
concat=concat,
vectorization=None,
load=False,
normalization=None)
return config
def get_experiment_settings(exp=1):
nb_branch = 1
batch_size = 128
if exp == 1:
params = [[64, 32, 16], [25, 25, 25],]
cov_outputs = [16]
cov_mode = 'pmean'
elif exp == 2:
""" Test fitnet with simple settings """
nb_branch = 2
params = [[100, 50, 25], ]
cov_outputs = [64]
cov_mode = 'pmean'
elif exp == 3:
# params = [[64, 64, 64], [100, 100, 100], [50, 50, 50],]
# params = [[64], [64, 128], [64,128,256],]
params = [[64], [64, 32], [64,32,16],]
cov_outputs = []
cov_mode = 'pmean'
title = 'cifar10_CrossV-WV_{}'.format(cov_mode, )
elif exp == 4:
params = [[30], [50], [70]]
# params = [[], [50,], [70], [64,64,64]]
# cov_outputs = range(10, 201, 20)
cov_outputs = range(20, 201, 20)
cov_mode = 'pmean'
title = 'cifar10_CrossV-WV_{}'.format(cov_mode,)
elif exp == 5:
# params = [[50, 40, 30], [50, 60, 70], [50,30,10], [50, 70, 90]]
# params = [[50, 100, 200, 400,],]
# params = [[50,50,50,50],]
# params = [[50,50,50,50, 50],]
params = [[50, 100, 200, 400,],]
# params = [[50, 100, 200, 400, 800],]
# params = [[800,400,200,100,50]]
# params = [[50, 50, 50]]
# params = [[50, 50,], [50], [50,50,50]]
# params = [[50, 50,], [50], [50,50,50]]
# params = [[50, 100, 200], [50, 100], [100,50], [200,100,50]]
cov_outputs = [params[0][-1]]
cov_mode = 'pmean'
title = 'cifar10-CrossV-O2T_{}_'.format(cov_mode)
elif exp == 6:
params = [[50, 25], [50, 75], ]
cov_outputs = [25, 50, 75]
cov_mode = 'pmean'
title = 'cifar10-CrossV-O2T_{}_'.format(cov_mode)
elif exp == 7:
# params = [[64,64,64]]
# params = [[190], [170],[110], [30 ]]
# params = [[90]]
params = [[70]]
# params = [[50]]
# params = [[30]]
# cov_outputs = [200, 160, 80, 40] # 90
# cov_outputs = [190, 170, ] # 70
# cov_outputs = [20] # 50
cov_outputs = [40,70,10] # 30
# cov_outputs = [50]
cov_mode = 'pmean'
title = 'cifar10-CrossV-O2T_{}_'.format(cov_mode)
elif exp == 8:
# params = [[]]
# params = [[150]]
params = [[100]]
# params = [[90]]
# params = [[70]]
# params = [[50]]
# params = [[30]]
# cov_outputs = range(20, 201, 20)
cov_outputs = range(220, 401, 20)
cov_mode = 'pmean'
title = 'cifar10_CrossV-WV_{}'.format(cov_mode,)
elif exp == 9:
params = [[150]]
# params = [[100]]
# params = [[90]]
# params = [[70]]
# params = [[50]]
# params = [[30]]
cov_outputs = range(10, 201, 20)
cov_mode = 'pmean'
title = 'cifar10_CrossV-WV_{}'.format(cov_mode,)
else:
return
mode_list = [1]
cov_branch = 'o2transform'
early_stop = True
cov_regularizer = None
last_config_feature_maps = []
robust = False
regroup = False
cov_alpha = 1
cov_beta = 0.1
if robust:
rb = 'robost'
else:
rb = ''
config = DCovConfig(params, mode_list, cov_outputs, cov_branch, cov_mode, early_stop, cov_regularizer,
nb_branch=nb_branch, last_conv_feature_maps=last_config_feature_maps, batch_size=batch_size,
exp=exp, epsilon=1e-5, title=title, robust=robust, cov_alpha=cov_alpha, regroup=regroup,
cov_beta=cov_beta)
return config
def get_VGG_testing_ideas(exp):
""" Test VGG dimension reduction """
cov_regularizer = None
if exp == 1:
""" Experiment 1, cross validate number of branches. """
nb_branch = 4
# params = [[128, 64, 32], ]
# params = [[257, 128, 64], ]
params = [
[257, 128, 128],
]
# params = [[64, 32, 16], ]
mode_list = [1]
# cov_outputs = [16]
cov_outputs = [params[0][2]]
cov_branch = 'o2t_no_wv'
cov_regularizer = None
# last_config_feature_maps = [512]
last_config_feature_maps = [1024]
concat = 'concat'
last_avg = False
robust = True
elif exp == 2:
""" Cross Validate on the summing methods """
nb_branch = 2
params = [
[257, 128, 64],
]
mode_list = [1]
cov_outputs = [64]
cov_branch = 'o2transform'
cov_regularizer = None
last_config_feature_maps = [512]
# last_config_feature_maps = [1024]
concat = 'ave'
last_avg = False
robust = False
elif exp == 3:
""" Test with robust single branch"""
nb_branch = 1
params = [
[257, 128, 64],
]
mode_list = [1]
cov_outputs = [64]
cov_branch = 'o2transform'
cov_regularizer = None
# last_config_feature_maps = []
last_config_feature_maps = [512]
concat = 'concat'
last_avg = False
robust = True
elif exp == 4:
""" Cross Validate on the summing methods for Riemannian"""
nb_branch = 2
# params = [[128, 64, 32],]
params = [
[257, 128, 64],
]
mode_list = [1]
cov_outputs = [64]
cov_branch = 'o2t_no_wv'
cov_regularizer = None
last_config_feature_maps = [512]
# last_config_feature_maps = [1024]
concat = 'concat'
last_avg = False
robust = True
else:
return
cov_mode = 'pmean'
early_stop = True
batch_size = 32
regroup = False
cov_alpha = 0.75
if robust:
rb = 'robost'
else:
rb = ''
cov_beta = 0.3
title = 'ImageNet_VGG16_{}_{}_LC{}_exp_{}_{}'.format(
cov_branch, rb, last_config_feature_maps, exp, concat)
config = DCovConfig(
params,
mode_list,
cov_outputs,
cov_branch,
cov_mode,
early_stop,
cov_regularizer,
nb_branch=nb_branch,
last_conv_feature_maps=last_config_feature_maps,
batch_size=batch_size,
exp=exp,
epsilon=1e-5,
title=title,
robust=robust,
cov_alpha=cov_alpha,
regroup=regroup,
concat=concat,
cov_beta=cov_beta,
)
return config
