def get_grid_param_list():
dictlistprod = cartesian_dict_of_lists_product
default_params = [{
'save_folder': './data/models/sacred/paper/csp/online/remove-cz/',
'only_return_exp': False,
}]
subject_id_params = dictlistprod({
'subject_id': ['elkh'] #'anla', 'hawe', 'lufi', 'sama',
})
break_params = dictlistprod({
'with_breaks': [True, False],
})
preproc_params = dictlistprod({
'min_freq': [1],
})
grid_params = product_of_list_of_lists_of_dicts([
default_params,
subject_id_params,
break_params,
preproc_params,
])
return grid_params
def get_grid_param_list():
dictlistprod = cartesian_dict_of_lists_product
default_params = [{
'save_folder': './data/models/sacred/paper/bcic-iv-2a-cross/start/',
'only_return_exp': False,
'n_chans': 22
}]
subject_folder_params = dictlistprod({
'subject_id': range(1,10),
'data_folder': ['/home/schirrmr/data/bci-competition-iv/2a/',]
})
loss_params = dictlistprod({
'loss_expression': ['$tied_loss']})
preproc_params = dictlistprod({
'filt_order': [3,],#10
'clean_train': [False],
'low_cut_hz': [0,4],
'train_start_ms': [1500]})
model_params = dictlistprod({
'network': ['shallow'] #deep
})
eval_params = dictlistprod({
'kappa_mode': ['max']})
grid_params = product_of_list_of_lists_of_dicts([
default_params,
subject_folder_params,
preproc_params,
loss_params,
model_params,
eval_params,
])
return grid_params
def get_grid_param_list():
dictlistprod = cartesian_dict_of_lists_product
default_params = [{
"save_folder":
"/data/schirrmr/schirrmr/reversible/experiments/dropout-weight-decay",
"only_return_exp": False,
"debug": False,
}]
subject_id_params = dictlistprod({"subject_id": range(4, 5)})
data_params = dictlistprod({"n_sensors": [22], "final_hz": [256]})
preproc_params = dictlistprod({"half_before": [True]})
ival_params = [{"start_ms": 500, "stop_ms": 1500}]
training_params = dictlistprod({"max_epochs": [600, 1200]})
implementation_params = dictlistprod({"constant_memory": [False]})
init_params = dictlistprod({"set_distribution_to_empirical": [True]})
optim_params = dictlistprod({
"uni_noise_factor": [
0,
1e-2,
5e-2,
], #
"gauss_noise_factor": [0.1], #
"weight_decay": [
1e-4,
], #
})
network_params = dictlistprod({
"final_fft": [True, False],
'drop_p': [
0,
0.1,
]
}) #
save_params = [{"save_model": True}]
grid_params = product_of_list_of_lists_of_dicts([
default_params,
subject_id_params,
data_params,
implementation_params,
training_params,
init_params,
preproc_params,
ival_params,
optim_params,
network_params,
save_params,
])
return grid_params
def get_grid_param_list():
dictlistprod = cartesian_dict_of_lists_product
save_params = [
{
'save_folder': '/home/schirrmr/data/exps/invertible-eeg/images/',
},
]
data_params = [{
'first_n': None,
}]
debug_params = [{
'debug': False,
}]
train_params = dictlistprod({
'n_epochs': [200],
'batch_size': [32],
})
random_params= dictlistprod({
'np_th_seed': range(0,2),
})
optim_params = dictlistprod({
'lr': [5e-4],
'weight_decay': [5e-5],
'cross_ent_weight': [0,],#1,10,100,1000
'scale_2_cross_ent': [False, True],
'mask_for_cross_ent': [False],
'nll_weight': [1],
'linear_classifier': [False],
'flow_gmm': [True],
'flow_coupling': ['affine'],
})
model_params = dictlistprod({
'n_mixes': [32],
})
grid_params = product_of_list_of_lists_of_dicts([
save_params,
train_params,
debug_params,
random_params,
optim_params,
model_params,
data_params,
])
return grid_params
def get_grid_param_list():
dictlistprod = cartesian_dict_of_lists_product
default_params = [{
"save_folder":
"/data/schirrmr/schirrmr/reversible/experiments/new-deep-invertible",
"only_return_exp": False,
"debug": False,
}]
subject_id_params = dictlistprod({"subject_id": range(4, 10)})
data_params = dictlistprod({"n_sensors": [22], "final_hz": [256]})
preproc_params = dictlistprod({"half_before": [True]})
ival_params = [{"start_ms": 500, "stop_ms": 1500}]
training_params = dictlistprod({"max_epochs": [1000, 4000]})
model_params = dictlistprod({"model_name": ["deep_invertible"]})
implementation_params = dictlistprod({"constant_memory": [True]})
init_params = dictlistprod({
"data_zero_init": [False],
"set_distribution_to_empirical": [True]
})
network_params = dictlistprod({"final_fft": [False]}) # True
clf_params = dictlistprod({"clf_loss": [None,
'sliced']}) # "likelihood", None
dist_params = dictlistprod({
"ot_on_class_dims": [False], # , True
"independent_class_dists": [True],
})
save_params = [{"save_model": True}]
grid_params = product_of_list_of_lists_of_dicts([
default_params,
subject_id_params,
data_params,
implementation_params,
training_params,
init_params,
model_params,
preproc_params,
dist_params,
ival_params,
network_params,
clf_params,
save_params,
])
return grid_params
def get_grid_param_list():
dictlistprod = cartesian_dict_of_lists_product
debug_params = [{
'debug': False,
}]
data_params = dictlistprod({
'n_subjects': [None], #'#'
'n_seconds': [None], # 4 is one window
'subject_id': [None],
})
train_params = dictlistprod({
'n_epochs': [200],
})
random_params = dictlistprod({
'np_th_seed': range(0, 3),
})
optim_params = dictlistprod({
'lr': [5e-4],
'weight_decay': [5e-5],
})
model_params = dictlistprod({
'hidden_channels': [
512,
], #512
'n_virtual_chans': [
0,
], #1,2
'n_blocks_up': [4, 8],
'n_blocks_down': [4, 8],
'n_mixes': [128],
'splitter_last': ['haar', 'subsample'],
'init_perm_to_identity': [
True,
], #False#True
})
grid_params = product_of_list_of_lists_of_dicts([
data_params,
train_params,
debug_params,
random_params,
optim_params,
model_params,
])
return grid_params
def get_grid_param_list():
dictlistprod = cartesian_dict_of_lists_product
default_params = [{
'save_folder': './data/models/fbcsp/6-class/avg-cov/',
}]
filename_params = [
{
'filename':
'/data/schirrmr/schirrmr/offline-6-class-cabin/AnTiCUO1_1-4_250Hz.BBCI.mat',
},
{
'filename':
'/data/schirrmr/schirrmr/offline-6-class-cabin/FeHeCUO1_1-8_250Hz.BBCI.mat',
},
]
filterbank_params = [
{
'min_freq': 1,
'max_freq': 38,
'low_width': 6,
'high_width': 8,
'high_overlap': 4,
'last_low_freq': 10,
'low_overlap': 3,
'n_top_bottom_csp_filters': 5,
'n_selected_features': 20
},
{
'min_freq': 1,
'max_freq': 118,
'low_width': 6,
'high_width': 8,
'high_overlap': 4,
'last_low_freq': 10,
'low_overlap': 3,
'n_top_bottom_csp_filters': 5,
'n_selected_features': 20
},
]
sensor_params = [{
'sensors': 'all',
}, {
'sensors': 'C_sensors'
}]
grid_params = product_of_list_of_lists_of_dicts(
[default_params, filename_params, filterbank_params, sensor_params])
return grid_params
def get_grid_param_list():
dictlistprod = cartesian_dict_of_lists_product
default_params = [{
'save_folder': './data/models/pytorch/online/niri-repl/',
'only_return_exp': False,
}]
stop_params = [{
'max_epochs': 200,
}]
grid_params = product_of_list_of_lists_of_dicts([
default_params,
stop_params,
])
return grid_params
def get_grid_param_list():
dictlistprod = cartesian_dict_of_lists_product
default_params = [{
'save_folder': './data/models/sacred/paper/bcic-iv-2b/deepshallow/',
'only_return_exp': False,
'n_chans': 3,
'run_after_early_stop': True,
}]
subject_folder_params = dictlistprod({
'subject_id':
range(1, 10),
'data_folder': ['/home/schirrmr/data/bci-competition-iv/2b/'],
'train_inds': [[1, 2, 3]],
'test_inds': [[4, 5]],
'sets_like_fbcsp_paper': [True],
})
model_params = dictlistprod({
'network': ['deep', 'shallow'], #'deep'#'shallow'
})
stop_params = dictlistprod({'stop_chan': ['misclass']}) #'misclass',
loss_params = dictlistprod({'loss_expression':
['$tied_loss']}) #'misclass',
preproc_params = dictlistprod({
'clean_train': [
False,
], #False
'filt_order': [
3,
], #10
'low_cut_hz': [0, 4]
})
grid_params = product_of_list_of_lists_of_dicts([
default_params,
subject_folder_params,
model_params,
stop_params,
preproc_params,
loss_params,
])
return grid_params
def get_grid_param_list():
dictlistprod = cartesian_dict_of_lists_product
default_params = [{
'save_folder': './data/models/sacred/paper/bcic-iv-2a/repl/',
'only_return_exp': False,
'n_chans': 22
}]
subject_folder_params = dictlistprod({
'subject_id':
range(1, 10),
'data_folder': [
'/home/schirrmr/data/bci-competition-iv/2a/',
]
})
grid_params = product_of_list_of_lists_of_dicts([
default_params,
subject_folder_params,
])
return grid_params
def get_grid_param_list():
dictlistprod = cartesian_dict_of_lists_product
default_params = [{
'save_folder': './data/models/sacred/paper/bcic-iv-2b/cv-proper-sets/',
'only_return_exp': False,
'n_chans': 3,
}]
subject_folder_params = dictlistprod({
'subject_id':
range(1, 10),
'data_folder': ['/home/schirrmr/data/bci-competition-iv/2b/'],
})
exp_params = dictlistprod({
'run_after_early_stop': [
True,
],
})
stop_params = dictlistprod({'stop_chan': ['misclass']}) #,
loss_params = dictlistprod({'loss_expression':
['$tied_loss']}) #'misclass',
preproc_params = dictlistprod({
'filt_order': [
3,
], #10
'low_cut_hz': [4],
'sets_like_fbcsp_paper': [False, True]
})
grid_params = product_of_list_of_lists_of_dicts([
default_params,
exp_params,
subject_folder_params,
stop_params,
preproc_params,
loss_params,
])
return grid_params
def get_grid_param_list():
dictlistprod = cartesian_dict_of_lists_product
save_params = [
{
'save_folder':
'/home/schirrmr/data/exps/invertible-eeg/hgd-21-ch-32-hz/',
},
]
data_params = dictlistprod({
'subject_id': list(range(1, 15)),
'dataset_name': ['hgd'], # 4 is one window
})
model_params = dictlistprod({
'hidden_channels': [
128,
], #512
'n_virtual_chans': [
0,
], #1,2
'n_blocks_up': [8],
'n_blocks_down': [8],
'n_mixes': [128],
'splitter_last': [
'haar',
],
'init_perm_to_identity': [
True,
], #False#True
})
grid_params = product_of_list_of_lists_of_dicts([
save_params,
data_params,
model_params,
])
return grid_params
def get_grid_param_list():
dictlistprod = cartesian_dict_of_lists_product
default_params = [
{
"save_folder": "/data/schirrmr/schirrmr/reversible/experiments/deepshallow",
"only_return_exp": False,
"debug": False,
}
]
subject_id_params = dictlistprod({"subject_id": range(4, 10)})
data_params = dictlistprod({"n_sensors": [22], "final_hz": [256]})
preproc_params = dictlistprod({"half_before": [True]})
ival_params = [{"start_ms": 500, "stop_ms": 1500}]
training_params = dictlistprod({"max_epochs": [100]})
model_params = dictlistprod({"model": ["deep_invertible",],
"final_fft": [True],
"add_bnorm": [False],}) # , True
optim_params = dictlistprod({"weight_decay": [0.5 * 0.001, 0.5*0.01],
"act_norm": [True, False]})
save_params = [{"save_model": False}]
grid_params = product_of_list_of_lists_of_dicts(
[
default_params,
subject_id_params,
data_params,
training_params,
preproc_params,
ival_params,
training_params,
model_params,
save_params,
optim_params,
]
)
return grid_params
def get_grid_param_list():
dictlistprod = cartesian_dict_of_lists_product
save_params = [
{
'save_folder':
'/home/schirrmr/data/exps/invertible-eeg/tuh-all-chans/',
},
]
data_params = dictlistprod({
'n_subjects': [2076], #'#'
'n_seconds': [15 * 4], # 4 is one window
'dataset_name': ['tuh'], # 4 is one window
})
grid_params = product_of_list_of_lists_of_dicts([
save_params,
data_params,
])
return grid_params
def get_grid_param_list():
dictlistprod = cartesian_dict_of_lists_product
default_params = [{
'save_folder': './data/models/pytorch/auto-diag/smac-test/',
'only_return_exp': False,
}]
load_params = [{
'max_recording_mins': 35,
'n_recordings': 5,
}]
clean_defaults = {
'max_min_threshold': None,
'shrink_val': None,
'max_min_expected': None,
'max_abs_val': None,
'batch_set_zero_val': None,
'batch_set_zero_test': None,
'max_min_remove': None,
}
clean_variants = [
#{},
#{'batch_set_zero_val': 500, 'batch_set_zero_test': True},
{
'max_abs_val': 800
},
#{'max_abs_val' : 500},
#{'shrink_val': 200},
#{'shrink_val': 500},
# {'shrink_val': 800},
]
clean_params = product_of_list_of_lists_of_dicts([[clean_defaults],
clean_variants])
preproc_params = dictlistprod({
'sec_to_cut': [60],
'duration_recording_mins': [3],
'sampling_freq': [100],
'low_cut_hz': [
None,
],
'high_cut_hz': [
None,
],
'divisor': [10],
})
standardizing_defaults = {
'exp_demean': False,
'exp_standardize': False,
'moving_demean': False,
'moving_standardize': False,
'channel_demean': False,
'channel_standardize': False,
}
standardizing_variants = [
{},
]
standardizing_params = product_of_list_of_lists_of_dicts(
[[standardizing_defaults], standardizing_variants])
split_params = dictlistprod({
'n_folds': [5],
'i_test_fold': [0],
})
model_params = [
{
'input_time_length': 1200,
'final_conv_length': 40,
'model_name': 'shallow',
},
]
model_constraint_params = dictlistprod({
'model_constraint': ['defaultnorm', None],
})
iterator_params = [{'batch_size': 64}]
stop_params = [{
'max_epochs': 3,
}]
grid_params = product_of_list_of_lists_of_dicts([
default_params, load_params, clean_params, preproc_params,
split_params, model_params, iterator_params, standardizing_params,
stop_params, model_constraint_params
])
return grid_params
def get_grid_param_list():
dictlistprod = cartesian_dict_of_lists_product
default_params = [{
'save_folder': './data/models/pytorch/auto-diag/threepath-10fold/',
'only_return_exp': False,
}]
load_params = [{
'max_recording_mins': 35,
'n_recordings': 1500,
}]
clean_defaults = {
'max_min_threshold': None,
'shrink_val': None,
'max_min_expected': None,
'max_abs_val': None,
'batch_set_zero_val': None,
'batch_set_zero_test': None,
'max_min_remove': None,
}
clean_variants = [
{'max_abs_val' : 800},
]
clean_params = product_of_list_of_lists_of_dicts(
[[clean_defaults], clean_variants])
preproc_params = dictlistprod({
'sec_to_cut': [60],
'duration_recording_mins': [3],
'sampling_freq': [100],
'low_cut_hz': [None,],
'high_cut_hz': [None,],
'divisor': [10],
})
standardizing_defaults = {
'exp_demean': False,
'exp_standardize': False,
'moving_demean': False,
'moving_standardize': False,
'channel_demean': False,
'channel_standardize': False,
}
standardizing_variants = [
{},
]
standardizing_params = product_of_list_of_lists_of_dicts(
[[standardizing_defaults], standardizing_variants])
split_params = dictlistprod({
'n_folds': [10],
'i_test_fold': [0,1,2,3,4,5,6,7,8,9],
})
model_params = dictlistprod({
'virtual_chan_1x1_conv': [True],
'mean_across_features': [False],
'n_classifier_filters': [100,],
'n_start_filters': [8,10],
'drop_prob': [0.5],
'early_bnorm': [False],
'extra_conv_stride': [2,4,],
'later_kernel_len': [5,9],
})
model_params = product_of_list_of_lists_of_dicts(
[model_params,
[
{
'n_preds_per_input': 3000,
'input_time_length': 6000,
},
]
]
)
final_layer_params = dictlistprod({
'sigmoid': [False ],
})
model_constraint_params = dictlistprod({
'model_constraint': ['defaultnorm', ],
})
iterator_params = [{
'batch_size': 64
}]
stop_params = [{
'max_epochs': 35,
}]
grid_params = product_of_list_of_lists_of_dicts([
default_params,
load_params,
clean_params,
preproc_params,
split_params,
model_params,
final_layer_params,
iterator_params,
standardizing_params,
stop_params,
model_constraint_params
])
return grid_params
def get_grid_param_list():
dictlistprod = cartesian_dict_of_lists_product
default_params = [{
'save_folder': './data/models/pytorch/auto-diag/more-layers/',
'only_return_exp': False,
}]
load_params = [{
'max_recording_mins': 35,
'n_recordings': 1500,
}]
clean_defaults = {
'max_min_threshold': None,
'shrink_val': None,
'max_min_expected': None,
'max_abs_val': None,
'batch_set_zero_val': None,
'batch_set_zero_test': None,
'max_min_remove': None,
}
clean_variants = [
{
'max_abs_val': 800
},
]
clean_params = product_of_list_of_lists_of_dicts([[clean_defaults],
clean_variants])
preproc_params = dictlistprod({
'sec_to_cut': [60],
'duration_recording_mins': [3],
'sampling_freq': [100],
'low_cut_hz': [
None,
],
'high_cut_hz': [
None,
],
'divisor': [10],
})
standardizing_defaults = {
'exp_demean': False,
'exp_standardize': False,
'moving_demean': False,
'moving_standardize': False,
'channel_demean': False,
'channel_standardize': False,
}
standardizing_variants = [
{},
]
standardizing_params = product_of_list_of_lists_of_dicts(
[[standardizing_defaults], standardizing_variants])
split_params = dictlistprod({
'n_folds': [10],
'i_test_fold': [0, 1, 2, 3, 4, 5, 6, 7, 8, 9],
})
model_params = [
{
'input_time_length': 18000,
'final_conv_length': 1,
},
]
model_params = product_of_list_of_lists_of_dicts([
model_params,
dictlistprod({
'pool_stride': [3],
'n_blocks_to_add': [0, 1, 2]
}) + dictlistprod({
'pool_stride': [4],
'n_blocks_to_add': [0, 1]
})
])
final_layer_params = dictlistprod({
'sigmoid': [False],
})
model_constraint_params = dictlistprod({
'model_constraint': [
'defaultnorm',
],
})
iterator_params = [{'batch_size': 32}]
stop_params = [{
'max_epochs': 35,
}]
grid_params = product_of_list_of_lists_of_dicts([
default_params, load_params, clean_params, preproc_params,
split_params, model_params, final_layer_params, iterator_params,
standardizing_params, stop_params, model_constraint_params
])
return grid_params
def get_grid_param_list():
dictlistprod = cartesian_dict_of_lists_product
train_test_filenames = [
{
'train_filename': 'BhNoMoSc1S001R01_ds10_1-12.BBCI.mat',
'test_filename': 'BhNoMoSc1S001R13_ds10_1-2BBCI.mat',
},
{
'train_filename': 'FaMaMoSc1S001R01_ds10_1-14.BBCI.mat',
'test_filename': 'FaMaMoSc1S001R15_ds10_1-2BBCI.mat',
},
{
'train_filename': 'FrThMoSc1S001R01_ds10_1-11.BBCI.mat',
'test_filename': 'FrThMoSc1S001R12_ds10_1-2BBCI.mat',
},
{
'train_filename': 'GuJoMoSc01S001R01_ds10_1-11.BBCI.mat',
'test_filename': 'GuJoMoSc01S001R12_ds10_1-2BBCI.mat'
},
{
'train_filename': 'KaUsMoSc1S001R01_ds10_1-11.BBCI.mat',
'test_filename': 'KaUsMoSc1S001R12_ds10_1-2BBCI.mat'
},
{
'train_filename': 'LaKaMoSc1S001R01_ds10_1-9.BBCI.mat',
'test_filename': 'LaKaMoSc1S001R10_ds10_1-2BBCI.mat'
},
{
'train_filename': 'LuFiMoSc3S001R01_ds10_1-11.BBCI.mat',
'test_filename': 'LuFiMoSc3S001R12_ds10_1-2BBCI.mat'
},
{
'train_filename': 'MaJaMoSc1S001R01_ds10_1-11.BBCI.mat',
'test_filename': 'MaJaMoSc1S001R12_ds10_1-2BBCI.mat'
},
{
'train_filename': 'MaKiMoSC01S001R01_ds10_1-4.BBCI.mat',
'test_filename': 'MaKiMoSC01S001R05_ds10_1-2BBCI.mat'
},
{
'train_filename': 'MaVoMoSc1S001R01_ds10_1-11.BBCI.mat',
'test_filename': 'MaVoMoSc1S001R12_ds10_1-2BBCI.mat'
},
# {
# 'train_filename': 'PiWiMoSc1S001R01_ds10_1-11.BBCI.mat',
# 'test_filename': 'PiWiMoSc1S001R12_ds10_1-2BBCI.mat'
# },
# {
# 'train_filename': 'RoBeMoSc03S001R01_ds10_1-9.BBCI.mat',
# 'test_filename': 'RoBeMoSc03S001R10_ds10_1-2BBCI.mat'
# },
# {
# 'train_filename': 'RoScMoSc1S001R01_ds10_1-11.BBCI.mat',
# 'test_filename': 'RoScMoSc1S001R12_ds10_1-2BBCI.mat'
# },
# {
# 'train_filename': 'StHeMoSc01S001R01_ds10_1-10.BBCI.mat',
# 'test_filename': 'StHeMoSc01S001R11_ds10_1-2BBCI.mat'
# },
]
data_split_train_params = dictlistprod({
'n_folds': [10],
'i_test_fold': list(range(9, 10)),
'test_on_eval_set': [False],
})
# data_split_test_params = [
# {
# 'n_folds': None,
# 'i_test_fold': None,
# 'test_on_eval_set': True,
# }]
no_early_stop_params = [{
'valid_set_fraction': None,
'use_validation_set': False,
'max_increase_epochs': None,
}]
adamw_adam_comparison_params = [{
'use_norm_constraint': False,
'optimizer_name': 'adam',
'schedule_weight_decay': False,
}, {
'use_norm_constraint': False,
'optimizer_name': 'adamw',
'schedule_weight_decay': True,
}]
scheduler_params = dictlistprod({
'scheduler_name': [
'cosine',
],
'restarts': [None],
})
lr_weight_decay_params = dictlistprod({
'model_name': ['deep'],
'init_lr':
np.array([1 / 8.0, 4.0, 8.0]) * 0.01,
'weight_decay':
np.array([
0, 1 / 64.0, 1 / 32.0, 1 / 16.0, 1 / 8.0, 1 / 4.0, 1 / 2.0, 1.0,
2.0, 4.0, 8.0
]) * 0.001,
}) + dictlistprod({
'model_name': ['deep'],
'init_lr':
np.array([1 / 8.0, 1 / 4.0, 1 / 2.0, 1.0, 2.0, 4.0, 8.0]) * 0.01,
'weight_decay':
np.array([1 / 64.0, 8.0]) * 0.001, #8.0 possibly removable
}) + dictlistprod({
'model_name': ['shallow'],
'init_lr':
np.array([
1 / 128.0,
1 / 64.0,
1 / 2.0,
]) * 0.01,
'weight_decay':
np.array([
0, 1 / 128.0, 1 / 64.0, 1 / 32.0, 1 / 16.0, 1 / 8.0, 1 / 4.0,
1 / 2.0, 1.0, 2.0, 4.0
]) * 0.001,
}) + dictlistprod({
'model_name': ['shallow'],
'init_lr':
np.array([
1 / 128.0,
1 / 64.0,
1 / 32.0,
1 / 16.0,
1 / 8.0,
1 / 4.0,
1 / 2.0,
]) * 0.01,
'weight_decay':
np.array([1 / 128.0, 1 / 64.0, 2.0, 4.0]) * 0.001,
}) + dictlistprod({
'model_name': ['resnet-xavier-uniform'],
'init_lr':
np.array([1 / 128.0, 1 / 64.0, 1 / 2.0, 1]) * 0.01,
'weight_decay':
np.array([
0, 1 / 32.0, 1 / 16.0, 1 / 8.0, 1 / 4.0, 1 / 2.0, 1.0, 2.0, 4.0,
8.0, 16
]) * 0.001,
}) + dictlistprod({
'model_name': ['resnet-xavier-uniform'],
'init_lr':
np.array([
1 / 128.0, 1 / 64.0, 1 / 32.0, 1 / 16.0, 1 / 8.0, 1 / 4.0, 1 / 2.0,
1
]) * 0.01,
'weight_decay':
np.array([16.0]) * 0.001,
})
# final_settings_params = [{
# 'max_epochs': 320,
# 'save_folder': '/home/schirrmr/data/models/adameegeval/final-no-restart/',
# 'use_norm_constraint': False,
# 'restarts': None,
# }]
# final_settings_variants = [
# {
# 'model_name': 'deep',
# 'optimizer_name': 'adam',
# 'scheduler_name': None,
# 'schedule_weight_decay': False,
# 'init_lr': 0.5 * 0.01,
# 'weight_decay': (1/32.0) * 0.001,
# },
# {
# 'model_name': 'deep',
# 'optimizer_name': 'adam',
# 'scheduler_name': 'cosine',
# 'schedule_weight_decay': False,
# 'init_lr': 2.0 * 0.01,
# 'weight_decay': (1/8.0) * 0.001,
# },
# {
# 'model_name': 'deep',
# 'optimizer_name': 'adamw',
# 'scheduler_name': None,
# 'schedule_weight_decay': True,
# 'init_lr': 2.0 * 0.01,
# 'weight_decay': 0 * 0.001,
# },
# {
# 'model_name': 'deep',
# 'optimizer_name': 'adamw',
# 'scheduler_name': 'cosine',
# 'schedule_weight_decay': True,
# 'init_lr': 1.0 * 0.01,
# 'weight_decay': 0.5 * 0.001,
# },
# {
# 'model_name': 'shallow',
# 'optimizer_name': 'adam',
# 'scheduler_name': None,
# 'schedule_weight_decay': False,
# 'init_lr': (1/32.0) * 0.01,
# 'weight_decay': (1/32.0) * 0.001,
# },
# {
# 'model_name': 'shallow',
# 'optimizer_name': 'adam',
# 'scheduler_name': 'cosine',
# 'schedule_weight_decay': False,
# 'init_lr': (1/8.0) * 0.01,
# 'weight_decay': 1.0 * 0.001,
# },
# {
# 'model_name': 'shallow',
# 'optimizer_name': 'adamw',
# 'scheduler_name': None,
# 'schedule_weight_decay': True,
# 'init_lr': (1/32.0) * 0.01,
# 'weight_decay': (1/8.0) * 0.001,
# },
# {
# 'model_name': 'shallow',
# 'optimizer_name': 'adamw',
# 'scheduler_name': 'cosine',
# 'schedule_weight_decay': True,
# 'init_lr': (1/16.0) * 0.01,
# 'weight_decay': 0 * 0.001,
# },
# {
# 'model_name': 'resnet-xavier-uniform',
# 'optimizer_name': 'adam',
# 'scheduler_name': None,
# 'schedule_weight_decay': False,
# 'init_lr': (1/32.0) * 0.01,
# 'weight_decay': 0 * 0.001,
# },
# {
# 'model_name': 'resnet-xavier-uniform',
# 'optimizer_name': 'adam',
# 'scheduler_name': 'cosine',
# 'schedule_weight_decay': False,
# 'init_lr': (1/8.0) * 0.01,
# 'weight_decay': 2.0 * 0.001,
# },
# {
# 'model_name': 'resnet-xavier-uniform',
# 'optimizer_name': 'adamw',
# 'scheduler_name': None,
# 'schedule_weight_decay': True,
# 'init_lr': (1/16.0) * 0.01,
# 'weight_decay': (1/32.0) * 0.001,
# },
# {
# 'model_name': 'resnet-xavier-uniform',
# 'optimizer_name': 'adamw',
# 'scheduler_name': 'cosine',
# 'schedule_weight_decay': True,
# 'init_lr': (1/32.0) * 0.01,
# 'weight_decay': 2.0 * 0.001,
# },
# ]
seed_params = dictlistprod({
'np_th_seed': [0] #0,1,2,3,4
})
preproc_params = dictlistprod({
'low_cut_hz': [4] #0
})
stop_params = [
# # {'max_epochs': None},
{
'max_epochs':
20,
'save_folder':
'/home/schirrmr/data/models/adameegeval/4sec-cv-lr-wd-20-epoch-2/',
},
{
'max_epochs':
40,
'save_folder':
'/home/schirrmr/data/models/adameegeval/4sec-cv-lr-wd-40-epoch-2/',
},
{
'max_epochs':
80,
'save_folder':
'/home/schirrmr/data/models/adameegeval/4sec-cv-lr-wd-80-epoch/',
},
{
'max_epochs':
160,
'save_folder':
'/home/schirrmr/data/models/adameegeval/4sec-cv-lr-wd-160-epoch/',
},
{
'max_epochs':
320,
'save_folder':
'/home/schirrmr/data/models/adameegeval/4sec-cv-lr-wd-320-epoch/',
},
]
debug_params = [{
'debug': False,
}]
grid_params = product_of_list_of_lists_of_dicts([
train_test_filenames,
data_split_train_params,
preproc_params,
no_early_stop_params,
adamw_adam_comparison_params,
scheduler_params,
lr_weight_decay_params,
stop_params,
seed_params,
debug_params,
])
return grid_params
def get_grid_param_list():
dictlistprod = cartesian_dict_of_lists_product
default_params = [{
'save_folder':
'/data/schirrmr/schirrmr/models/auto-diag/final-eval-from-smalldata-config/',
'only_return_exp': False,
}]
seed_params = dictlistprod({
'np_th_seed': list(range(5, 10)) #[0,1,2,3,4]
})
save_params = [{
'save_predictions': False,
'save_crop_predictions': False,
}]
load_params = [{
'max_recording_mins': 35,
'n_recordings': None,
}]
clean_params = [{
'max_abs_val': 800,
}]
sensor_params = [
{
'n_chans': 21,
'sensor_types': ['EEG'],
},
]
preproc_params = dictlistprod({
'sec_to_cut_at_start': [60],
'sec_to_cut_at_end': [0],
'duration_recording_mins': [20],
'test_recording_mins': [None],
'sampling_freq': [100],
'divisor': [
None,
], # 10 before
'clip_before_resample': [False], #False,
})
# this differentiates train/test also.
split_params = dictlistprod({
'test_on_eval': [True],
'n_folds': [5],
'i_test_fold': [4],
'shuffle': [False],
})
model_params = [
# {
# 'input_time_length': 6000,
# 'final_conv_length': 35,
# 'model_name': 'shallow',
# 'n_start_chans': 40,
# 'n_chan_factor': None,
# 'model_constraint': 'defaultnorm',
# 'stride_before_pool': None,
# 'scheduler': None,
# 'optimizer': 'adam',
# 'learning_rate': 1e-3,
# 'weight_decay': 0,
# 'merge_train_valid': False,
# },
# {
# 'input_time_length': 6000,
# 'final_conv_length': 1,
# 'model_name': 'deep',
# 'n_start_chans': 25,
# 'n_chan_factor': 2,
# 'model_constraint': 'defaultnorm',
# 'stride_before_pool': True,
# 'scheduler': None,
# 'optimizer': 'adam',
# 'learning_rate': 1e-3,
# 'weight_decay': 0,
# 'merge_train_valid': False,
# },
# {
# 'input_time_length': 6000,
# 'final_conv_length': 35,
# 'model_name': 'shallow',
# 'n_start_chans': 40,
# 'n_chan_factor': None,
# 'model_constraint': None,
# 'stride_before_pool': None,
# 'scheduler': 'cosine',
# 'optimizer': 'adamw',
# 'learning_rate': 0.0625 * 0.01,
# 'weight_decay': 0,
# 'merge_train_valid': True,
# },
{
'input_time_length': 6000,
'final_conv_length': 1,
'model_name': 'deep',
'n_start_chans': 25,
'n_chan_factor': 2,
'model_constraint': None,
'stride_before_pool': True,
'scheduler': 'cosine',
'optimizer': 'adamw',
'learning_rate': 1 * 0.01,
'weight_decay': 0.5 * 0.001,
'merge_train_valid': True,
},
]
iterator_params = [{'batch_size': 64}]
stop_params = [{
'max_epochs': 35,
}]
grid_params = product_of_list_of_lists_of_dicts([
default_params,
seed_params,
save_params,
load_params,
clean_params,
preproc_params,
sensor_params,
split_params,
model_params,
iterator_params,
stop_params,
])
return grid_params
