def test_initialize(self):
c = ABC(10, objective_function)
c.add_param(0, 10)
c.add_param(0, 10)
c.initialize()
self.assertEqual(len(c._bees), 20)
def test_kwargs(self):
c = ABC(10, objective_function_kwargs, {'my_kwarg': 2})
c.add_param(0, 0)
c.add_param(0, 0)
c.initialize()
self.assertEqual(c.best_ret_val, 2)
def test_multiprocessing(self):
c = ABC(20, objective_function, num_processes=4)
self.assertEqual(c._num_processes, 4)
c.add_param(0, 10)
c.add_param(0, 10)
c.initialize()
c.search()
def test_custom_param_name(self):
c = ABC(20, objective_function)
c.add_param(0, 10, name='int1')
c.add_param(0, 10, name='int2')
c.initialize()
for _ in range(50):
c.search()
self.assertEqual(c.best_params, {'int1': 0, 'int2': 0})
def test_search(self):
c = ABC(20, objective_function)
c.add_param(0, 10)
c.add_param(0, 10)
c.initialize()
for _ in range(50):
c.search()
self.assertEqual(c.best_fitness, 1)
self.assertEqual(c.best_ret_val, 0)
self.assertEqual(c.best_params, {'P0': 0, 'P1': 0})
def test_get_stats(self):
c = ABC(10, objective_function)
c.add_param(0, 0)
c.add_param(0, 0)
c.initialize()
self.assertEqual(c.best_fitness, 1)
self.assertEqual(c.best_ret_val, 0)
self.assertEqual(c.best_params, {'P0': 0, 'P1': 0})
self.assertEqual(c.average_fitness, 1)
self.assertEqual(c.average_ret_val, 0)
def tune_training_parameters(n_bees: int, n_iter: int, dataset_train: QSPRDataset,
dataset_eval: QSPRDataset, n_processes: int = 1,
**kwargs) -> dict:
"""
Tunes learning rate, learning rate decay; additional **kwargs can include any in:
[
# ECNet parameters
'epochs' (default 100),
'batch_size' (default 32),
'valid_size' (default 0.2),
'patience' (default 32),
'hidden_dim' (default 128),
'n_hidden' (default 2),
'dropout': (default 0.0),
# Adam optim. alg. arguments
'beta_1' (default 0.9),
'beta_2' (default 0.999),
'eps' (default 1e-8),
'weight_decay' (default 0.0),
'amsgrad' (default False)
]
Args:
n_bees (int): number of employer bees to use in ABC algorithm
n_iter (int): number of iterations, or "search cycles", for ABC algorithm
dataset_train (QSPRDataset): dataset used to train evaluation models
dataset_eval (QSPRDataset): dataset used for evaluation
n_processes (int, optional): if > 1, uses multiprocessing when evaluating at an iteration
**kwargs: additional arguments
Returns:
dict: {'lr': float, 'lr_decay': float}
"""
kwargs['train_ds'] = dataset_train
kwargs['eval_ds'] = dataset_eval
abc = ABC(n_bees, _cost_train_hp, num_processes=n_processes, obj_fn_args=kwargs)
abc.add_param(CONFIG['training_params_range']['lr'][0],
CONFIG['training_params_range']['lr'][1], name='lr')
abc.add_param(CONFIG['training_params_range']['lr_decay'][0],
CONFIG['training_params_range']['lr_decay'][1], name='lr_decay')
abc.initialize()
for _ in range(n_iter):
abc.search()
return {
'lr': abc.best_params['lr'],
'lr_decay': abc.best_params['lr_decay']
}
def main():
# Create the colony with 10 employer bees and the above objective function
abc = ABC(10, minimize_integers)
# Add three integers, randomly initialized between 0 and 10 for each bee
abc.add_param(0, 10, name='Int_1')
abc.add_param(0, 10, name='Int_2')
abc.add_param(0, 10, name='Int_3')
# Initialize 10 employer bees, 10 onlooker bees
abc.initialize()
# Run the search cycle 10 times
for _ in range(10):
abc.search()
print('Average fitness: {}'.format(abc.average_fitness))
print('Average obj. fn. return value: {}'.format(abc.average_ret_val))
print('Best fitness score: {}'.format(abc.best_fitness))
print('Best obj. fn. return value: {}'.format(abc.best_ret_val))
print('Best parameters: {}\n'.format(abc.best_params))
def test_add_parameter(self):
c = ABC(10, objective_function)
c.add_param(0, 1)
c.add_param(2, 3)
c.add_param(4, 5)
self.assertEqual(len(c._params), 3)
self.assertEqual(c._params[0]._min_val, 0)
self.assertEqual(c._params[0]._max_val, 1)
self.assertEqual(c._params[1]._min_val, 2)
self.assertEqual(c._params[1]._max_val, 3)
self.assertEqual(c._params[2]._min_val, 4)
self.assertEqual(c._params[2]._max_val, 5)
self.assertEqual(c._params[0]._dtype, int)
c.add_param(0.0, 1.0)
self.assertEqual(c._params[3]._dtype, float)
self.assertTrue(c._params[0]._restrict)
c.add_param(0, 1, False)
self.assertFalse(c._params[4]._restrict)
def tune_hyperparameters(df: DataFrame,
vars: dict,
num_employers: int,
num_iterations: int,
num_processes: int = 1,
shuffle: str = None,
split: list = None,
validate: bool = True,
eval_set: str = None,
eval_fn: str = 'rmse',
epochs: int = 500) -> dict:
'''Tunes neural network learning/architecture hyperparameters
Args:
df (ecnet.utils.data_utils.DataFrame): currently loaded data
vars (dict): ecnet.Server._vars variables
num_employers (int): number of employer bees
num_iterations (int): number of search cycles for the colony
num_processes (int): number of parallel processes to utilize
shuffle (str): shuffles `train` or `all` sets if not None
split (list): if shuffle is True, [learn%, valid%, test%]
validate (bool): if True, uses periodic validation; otherwise, no
eval_set (str): set used to evaluate bee performance; `learn`, `valid`,
`train`, `test`, None (all sets)
eval_fn (str): error function used to evaluate bee performance; `rmse`,
`mean_abs_error`, `med_abs_error`
epochs (int): number of training epochs per bee ANN (def: 500)
Returns:
dict: tuned hyperparameters
'''
if name != 'nt':
set_start_method('spawn', force=True)
logger.log('info',
'Tuning architecture/learning hyperparameters',
call_loc='TUNE')
logger.log('debug',
'Arguments:\n\t| num_employers:\t{}\n\t| '
'num_iterations:\t{}\n\t| shuffle:\t\t{}\n\t| split:'
'\t\t{}\n\t| validate:\t\t{}\n\t| eval_set:\t\t{}\n\t'
'| eval_fn:\t\t{}'.format(num_employers, num_iterations,
shuffle, split, validate, eval_set,
eval_fn),
call_loc='TUNE')
fit_fn_args = {
'df': df,
'shuffle': shuffle,
'num_processes': num_processes,
'split': split,
'validate': validate,
'eval_set': eval_set,
'eval_fn': eval_fn,
'hidden_layers': vars['hidden_layers'],
'epochs': epochs
}
to_tune = [(1e-9, 1e-4, 'decay'), (1e-5, 0.1, 'learning_rate'),
(1, len(df.learn_set), 'batch_size'), (64, 1024, 'patience')]
for hl in range(len(vars['hidden_layers'])):
to_tune.append((1, 2 * len(df._input_names), 'hl{}'.format(hl)))
abc = ABC(num_employers, tune_fitness_function, fit_fn_args, num_processes)
for param in to_tune:
abc.add_param(param[0], param[1], name=param[2])
abc.initialize()
best_ret_val = abc.best_ret_val
best_params = abc.best_params
for i in range(num_iterations):
logger.log('info', 'Iteration {}'.format(i + 1), call_loc='TUNE')
abc.search()
new_best_ret = abc.best_ret_val
new_best_params = abc.best_params
logger.log('info',
'Best Performer: {}, {}'.format(new_best_ret, new_best_ret),
call_loc='TUNE')
if new_best_ret < best_ret_val:
best_ret_val = new_best_ret
best_params = new_best_params
vars['decay'] = best_params['decay']
vars['learning_rate'] = best_params['learning_rate']
vars['batch_size'] = best_params['batch_size']
vars['patience'] = best_params['patience']
for l_idx in range(len(vars['hidden_layers'])):
vars['hidden_layers'][l_idx][0] = best_params['hl{}'.format(l_idx)]
return vars