def check_model_list(model_list, datasets): opts = default_opt_creator() for dataset_str in datasets: for model in model_list: set_dataset(opts, dataset=dataset_str) set_default_opts_based_on_model_dataset(opts) set_model_string(opts, model) get_model_from_db(model, opts)
def check_model_list(model_list, datasets): opts = default_opt_creator() for dataset_str in datasets: for model in model_list: set_dataset(opts, dataset=dataset_str) set_default_opts_based_on_model_dataset(opts) set_model_string(opts, model) model_dict= get_model_from_db(model, opts) model = model_dict['model'] model.summary()
print(100 * '*', 3 * '\n', model_str, '\n', dataset_str, 3 * '\n', 100 * '*')
opts = default_opt_creator()
opts['experiment_name'] = experiment_name
opts['experiment_tag'] = experiment_name + '/' + dataset_str + '/' + weight_model_name + 'loaded_to_' + model_str
set_dataset(opts, dataset=dataset_str)
opts = set_model_string(opts, model_str)
opts = set_default_opts_based_on_model_dataset(opts)
input_shape = opts['training_opts']['dataset']['input_shape']
nb_class = opts['training_opts']['dataset']['nb_classes']
# opts = set_expand_rate(opts, param_expand_sel)
# optimizer = optimizers.Nadam()
optimizer = optimizers.SGD(lr=opts['optimizer_opts']['lr'], momentum=opts['optimizer_opts']['momentum'],
decay=opts['optimizer_opts']['decay'], nesterov=opts['optimizer_opts']['nestrov'])
# optimizer = optimizers.Adadelta()
""" MODEL PREPARE """
model = get_model_from_db(model_str, opts)
weight_model = get_model_from_db(weight_model_name, opts)
model_modification_utils.load_weights_by_block_index_list(model, [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11], os.path.join(
global_constant_var.get_experimentcase_abs_path(weight_model_experiment_name, dataset_str, weight_model_name), 'checkpoint'),
model_constructor_utils.CONVSH_NAME)
model_modification_utils.load_weights_by_block_index_list(weight_model, [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11], os.path.join(
global_constant_var.get_experimentcase_abs_path(weight_model_experiment_name, dataset_str, weight_model_name), 'checkpoint'),
model_constructor_utils.CONVSH_NAME)
model.compile(loss=opt_utils.get_loss(opts), optimizer=optimizer, metrics=opt_utils.get_metrics(opts))
weight_model.compile(loss=opt_utils.get_loss(opts), optimizer=optimizer, metrics=opt_utils.get_metrics(opts))
method_names = find_key_value_to_str_recursive(opts, '', {'param_expand'})
opts['experiment_name'] = method_names
# LOAD DATA
(data_train, label_train), (data_test, label_test) = load_data(dataset_str, opts)
data_train, data_test = preprocess_data_phase(opts, data_train, data_test)
data_gen = data_augmentation_phase(opts)
opts = default_opt_creator()
opts['experiment_name'] = experiment_name
set_dataset(opts, dataset=dataset_str)
opts = set_model_string(opts, model_str)
opts = set_default_opts_based_on_model_dataset(opts)
input_shape = opts['training_opts']['dataset']['input_shape']
nb_class = opts['training_opts']['dataset']['nb_classes']
# opts = set_expand_rate(opts, param_expand_sel)
# optimizer = optimizers.Nadam()
optimizer = optimizers.SGD(lr=opts['optimizer_opts']['lr'], momentum=opts['optimizer_opts']['momentum'],
decay=opts['optimizer_opts']['decay'], nesterov=opts['optimizer_opts']['nestrov'])
# optimizer = optimizers.Adadelta()
""" MODEL PREPARE """
model_dict = get_model_from_db(model_str, opts)
model = model_dict['model']
model_total_params = (model.count_params() // 100000) / 10
if (not total_params == 0) and (not model_total_params== total_params):
set_expand_rate(opts, np.sqrt(total_params/model_total_params)*get_expand_rate(opts));
print('Expand Rate Changed')
model_dict = get_model_from_db(model_str, opts)
model = model_dict['model']
if total_params==0: total_params = model_total_params;
opts['experiment_tag'] = experiment_name + '/' + dataset_str + '/' + model_str + '/' + str((model.count_params()//100000)/10)+'M'
# out_tensor_list = model_dict['out']
# output_num = len(out_tensor_list)
model.summary()
# model_modification_utils.load_weights_by_block_index_list(model, [1, 2, 3, 4, 5, 6, 7, 8, 9], os.path.join(
# global_constant_var.get_experimentcase_abs_path(experiment_name, dataset_str, 'nin_tree_berp_1'), 'checkpoint'),
'experiment_tag'] = experiment_name + '/' + dataset_str + '/' + model_str
set_dataset(opts, dataset=dataset_str)
opts = set_model_string(opts, model_str)
opts = set_default_opts_based_on_model_dataset(opts)
input_shape = opts['training_opts']['dataset']['input_shape']
nb_class = opts['training_opts']['dataset']['nb_classes']
# opts = set_expand_rate(opts, param_expand_sel)
# optimizer = optimizers.Nadam()
optimizer = optimizers.SGD(
lr=opts['optimizer_opts']['lr'],
momentum=opts['optimizer_opts']['momentum'],
decay=opts['optimizer_opts']['decay'],
nesterov=opts['optimizer_opts']['nestrov'])
# optimizer = optimizers.Adadelta()
""" MODEL PREPARE """
model = get_model_from_db(model_str, opts)
model.summary()
# model_modification_utils.load_weights_by_block_index_list(model, [1, 2, 3, 4, 5, 6, 7, 8, 9], os.path.join(
# global_constant_var.get_experimentcase_abs_path(experiment_name, dataset_str, 'nin_tree_berp_1'), 'checkpoint'),
# model_constructor_utils.CONVSH_NAME)
model.compile(loss=opt_utils.get_loss(opts),
optimizer=optimizer,
metrics=opt_utils.get_metrics(opts))
method_names = find_key_value_to_str_recursive(
opts, '', {'param_expand'})
opts['experiment_name'] = method_names
# LOAD DATA
(data_train,
label_train), (data_test,
label_test) = load_data(dataset_str, opts)
data_train, data_test = preprocess_data_phase(