def __init__(self, config, out_dir=None):
super(DataValuation, self).__init__(config, out_dir)
self.datasets = ds.loader.load_dataset(**self.config['dataset_config'])
self.dataset_id = self.config['dataset_config']['dataset_id']
self.data_dir = self.config['dataset_config']['data_dir']
self.train = self.datasets.train
self.test = self.datasets.test
self.validation = self.datasets.validation
self.sample_weights = ds.loader.load_supplemental_info(self.dataset_id + '_weights',
data_dir=self.data_dir)\
if self.config['sample_weights'] else [np.ones(self.train.x.shape[0]),
np.ones(self.validation.x.shape[0]),
np.ones(self.test.x.shape[0])]
self.nonfires = ds.loader.load_supplemental_info(
self.dataset_id + '_nonfires', data_dir=self.data_dir)
model_dir = os.path.join(self.base_dir, 'models')
model_config = LogisticRegression.default_config()
model_config['arch'] = LogisticRegression.infer_arch(self.train)
model_config['arch']['fit_intercept'] = True
# Heuristic for determining maximum batch evaluation sizes without OOM
D = model_config['arch']['input_dim'] * model_config['arch'][
'num_classes']
model_config['grad_batch_size'] = max(1,
self.config['max_memory'] // D)
model_config['hessian_batch_size'] = max(
1, self.config['max_memory'] // (D * D))
self.model_dir = model_dir
self.model_config = model_config
# Convenience member variables
self.num_train = self.train.num_examples
self.num_classes = self.model_config['arch']['num_classes']
def get_model(self, dataset_id=None):
if not hasattr(self, 'model'):
dataset = self.get_dataset(dataset_id)
model_config = LogisticRegression.default_config()
model_config['arch'] = LogisticRegression.infer_arch(dataset.train)
model_dir = os.path.join(self.base_dir, 'models')
self.model = LogisticRegression(model_config, model_dir)
return self.model
def get_model(self):
if not hasattr(self, 'model'):
self.model = LogisticRegression(
self.model_config,
self.model_dir,
random_state=np.random.RandomState(2))
return self.model
def __init__(self, config, out_dir=None):
super(TestLogreg, self).__init__(config, out_dir)
self.datasets = ds.loader.load_dataset(**self.config['dataset_config'])
model_dir = os.path.join(self.base_dir, 'models')
model_config = LogisticRegression.default_config()
model_config['arch'] = LogisticRegression.infer_arch(
self.datasets.train)
model_config['arch']['fit_intercept'] = self.config['fit_intercept']
self.model_dir = model_dir
self.model_config = model_config
MAX_MEMORY = int(1e7)
D = model_config['arch']['input_dim'] * model_config['arch'][
'num_classes']
self.eval_args = {
'grad_batch_size': max(1, MAX_MEMORY // D),
'hess_batch_size': max(1, MAX_MEMORY // (D * D)),
}
def __init__(self, config, out_dir=None):
super(SubsetInfluenceLogreg, self).__init__(config, out_dir)
self.datasets = ds.loader.load_dataset(**self.config['dataset_config'])
self.train = self.datasets.train
self.test = self.datasets.test
self.validation = self.datasets.validation
model_dir = os.path.join(self.base_dir, 'models')
model_config = LogisticRegression.default_config()
model_config['arch'] = LogisticRegression.infer_arch(self.datasets.train)
model_config['arch']['fit_intercept'] = True
# Heuristic for determining maximum batch evaluation sizes without OOM
D = model_config['arch']['input_dim'] * model_config['arch']['num_classes']
model_config['grad_batch_size'] = max(1, self.config['max_memory'] // D)
model_config['hessian_batch_size'] = max(1, self.config['max_memory'] // (D * D))
# Set the method for computing inverse HVP
model_config['inverse_hvp_method'] = self.config['inverse_hvp_method']
self.model_dir = model_dir
self.model_config = model_config
# Convenience member variables
self.dataset_id = self.config['dataset_config']['dataset_id']
self.num_train = self.datasets.train.num_examples
self.num_classes = self.model_config['arch']['num_classes']
self.num_subsets = self.config['num_subsets']
if self.subset_choice_type == "types":
self.subset_size = int(self.num_train * self.config['subset_rel_size'])
elif self.subset_choice_type == "range":
self.subset_min_size = int(self.num_train * self.config['subset_min_rel_size'])
self.subset_max_size = int(self.num_train * self.config['subset_max_rel_size'])
tasks_dir = os.path.join(self.base_dir, 'tasks')
self.task_queue = TaskQueue(tasks_dir, master_only=self.config['master_only'])
self.task_queue.define_task('retrain_subsets', self.retrain_subsets)
self.task_queue.define_task('self_pred_infl', self.self_pred_infl)
self.task_queue.define_task('newton_batch', self.newton_batch)
def __init__(self, config, out_dir=None):
super(CreditAssignment, self).__init__(config, out_dir)
self.datasets = ds.loader.load_dataset(**self.config['dataset_config'])
self.dataset_id = self.config['dataset_config']['dataset_id']
self.data_dir = self.config['dataset_config']['data_dir']
self.train = self.datasets.train
print("Shape of training set: {}".format(self.train.x.shape))
self.test = self.datasets.test
self.validation = self.datasets.validation
self.sample_weights = ds.loader.load_supplemental_info(self.dataset_id + '_weights',
data_dir=self.data_dir)\
if self.config['sample_weights'] else [np.ones(self.train.x.shape[0]),
np.ones(self.validation.x.shape[0]),
np.ones(self.test.x.shape[0])]
self.num_train = self.train.num_examples
model_dir = os.path.join(self.base_dir, 'models')
model_config = LogisticRegression.default_config()
model_config['arch'] = LogisticRegression.infer_arch(self.train)
model_config['arch']['fit_intercept'] = True
# Heuristic for determining maximum batch evaluation sizes without OOM
D = model_config['arch']['input_dim'] * model_config['arch'][
'num_classes']
if 'grad_batch_size' in self.config and self.config[
'grad_batch_size'] is not None:
model_config['grad_batch_size'] = self.config['grad_batch_size']
else:
model_config['grad_batch_size'] = max(
1, self.config['max_memory'] // D)
if 'hessian_batch_size' in self.config and self.config[
'hessian_batch_size'] is not None:
model_config['hessian_batch_size'] = self.config[
'hessian_batch_size']
else:
model_config['hessian_batch_size'] = max(
1, self.config['max_memory'] // (D * D))
self.model_dir = model_dir
self.model_config = model_config
# Convenience member variables
self.num_classes = self.model_config['arch']['num_classes']
self.nonfires = ds.loader.load_supplemental_info(
self.dataset_id + '_nonfires', data_dir=self.data_dir)
def print_class_balance(ds, name):
print("Dataset {}:".format(name))
for i, val in enumerate(
np.bincount(ds.labels) / ds.labels.shape[0]):
print("Class {} is {} of the dataset.".format(i, val))
print_class_balance(self.train, 'train')
print_class_balance(self.test, 'test')
print_class_balance(self.nonfires, 'nonfires')
self.task_queue = TaskQueue(os.path.join(self.base_dir, 'tasks'),\
master_only=self.config['master_only'])
self.task_queue.define_task('compute_all_and_fixed_test_and_nonfire_influence',\
self.compute_all_and_fixed_test_and_nonfire_influence)
self.task_queue.define_task('retrain_subsets', self.retrain_subsets)
self.task_queue.define_task('self_pred_infl', self.self_pred_infl)
def get_model(self):
if not hasattr(self, 'model'):
self.model = LogisticRegression(self.model_config, self.model_dir)
return self.model