def get_saliency_for_classification(data,
label,
saliency_function,
batch_size=50,
balance_data=False,
class_func='sum',
reduce_func='sum',
generator=None,
generator_kwargs=None,
generator_epochs=50,
horizontal_flip=False):
label_ids = np.unique(label) if label.ndim == 1 else list(
range(label.shape[-1]))
saliency_function = []
if isinstance(class_func, str):
class_func = getattr(np, class_func)
data_, label_ = utils.balance_data(data,
label) if balance_data else data, label
for label_id in label_ids:
index = np.where(
label_ == label_id)[0] if label_.ndim == 1 else np.where(
label_[:, label_id] == 1)[0]
label_saliency = get_saliency_func(data_[index], label_[index],
saliency_function, batch_size,
reduce_func, generator,
generator_kwargs, generator_epochs)
if horizontal_flip and len(data_.shape) > 2:
label_saliency += get_saliency_func(data_[index][..., ::-1, :],
label_[index],
saliency_function, batch_size,
reduce_func)
label_saliency /= np.maximum(1e-6, np.sum(np.abs(label_saliency)))
saliency_function.append(label_saliency)
saliency = class_func(saliency_function, axis=0)
# factor = np.sign(np.max(saliency_function, axis=0))
# print('bad values :', (factor < 0).sum())
saliency /= np.maximum(1e-6, np.abs(saliency).sum())
del data_, label_, saliency_function
return saliency
dataset_name, max_features = dataset_stats
print('loading dataset', dataset_name)
dataset = load_dataset(dataset_name,
directory=datasets_directory,
normalize=dataset_name
not in ['dexter', 'dorothea'])
print('data loaded. labels =', dataset['train']['data'].shape)
batch_size = min(len(dataset['train']['data']), b_size)
input_shape = dataset['train']['data'].shape[-1:]
nclasses = len(np.unique(dataset['train']['label']))
data = dataset['train']['data']
label = dataset['train']['label']
data, label = balance_data(data, label)
valid_data = dataset['validation']['data']
valid_label = dataset['validation']['label']
label = kutils.to_categorical(label, 2)
valid_label = kutils.to_categorical(valid_label, 2)
total_features = data.shape[-1]
fs_filename = root_directory + dataset_name + '_gamma_' + str(
gamma) + '_ranks.json'
output_filename = root_directory + dataset_name + '_gamma_' + str(
gamma) + '_results.json'
try:
with open(fs_filename) as outfile: