def prepare_dataset(self):
mode_settings = importlib.import_module(self.settings_type)
train_args = mode_settings.get_args('train')
# load dataset
train_mini_batch_loader = DatasetPreProcessor(train_args)
test_mini_batch_loader = DatasetPreProcessor(mode_settings.get_args('test'))
print("---set mini_batch----------")
if train_args.importance_sampling:
print("importance----------")
train_it = ImportantSerialIterator( \
train_mini_batch_loader, \
train_args.training_params.batch_size, \
shuffle=train_args.shuffle, \
p=np.loadtxt(train_args.weights_file_path))
else:
if train_args.training_params.iter_type=='multi':
iterator = chainer.iterators.MultiprocessIterator
else:
iterator = chainer.iterators.SerialIterator
train_it = iterator( \
train_mini_batch_loader, \
train_args.training_params.batch_size, \
shuffle=train_args.shuffle)
val_batch_size = 1
val_it = iterator( \
test_mini_batch_loader, \
val_batch_size, repeat=False, shuffle=False)
return train_it, val_it, train_mini_batch_loader.__len__()
def prepare_dataset():
# load dataset
test_mini_batch_loader = DatasetPreProcessor('test')
val_it = chainer.iterators.SerialIterator( \
test_mini_batch_loader, \
1, repeat=False, shuffle=False)
return val_it, test_mini_batch_loader.__len__(), test_mini_batch_loader.pairs
def prepare_dataset(data_set):
train_args = get_args('train')
# load dataset
if data_set == 'mnist':
train_mini_batch_loader, test_mini_batch_loader = \
chainer.datasets.get_mnist()
else:
train_mini_batch_loader = DatasetPreProcessor(train_args)
test_mini_batch_loader = DatasetPreProcessor(get_args('test'))
print("---set mini_batch----------")
if train_args.importance_sampling:
print("importance----------")
train_it = ImportantSerialIterator( \
train_mini_batch_loader, \
train_args.training_params.batch_size, \
shuffle=train_args.shuffle, \
p=np.loadtxt(train_args.weights_file_path))
else:
if train_args.training_params.iter_type == 'multi':
iterator = chainer.iterators.MultiprocessIterator
else:
iterator = chainer.iterators.SerialIterator
train_it = iterator( \
train_mini_batch_loader, \
train_args.training_params.batch_size, \
shuffle=train_args.shuffle)
val_it = iterator( \
test_mini_batch_loader, \
1, repeat=False, shuffle=False)
return train_it, val_it, train_mini_batch_loader.__len__()
def prepare_dataset():
# データセットのロード
train_mini_batch_loader = \
DatasetPreProcessor(chainer.global_config.user_train_args)
train_it = chainer.iterators.SerialIterator(
train_mini_batch_loader,
chainer.global_config.user_train_args.training_params.batch_size)
return train_mini_batch_loader, train_mini_batch_loader.__len__()
def prepare_dataset(self):
# load dataset
if chainer.config.data_set == 'mnist':
train_mini_batch_loader, test_mini_batch_loader = \
chainer.datasets.get_mnist()
else:
train_mini_batch_loader = DatasetPreProcessor('train')
test_mini_batch_loader = DatasetPreProcessor('test')
if config.training_params.iter_type == 'multi':
iterator = chainer.iterators.MultiprocessIterator
else:
iterator = chainer.iterators.SerialIterator
train_it = iterator( \
train_mini_batch_loader, \
config.training_params.batch_size, \
shuffle=config.train)
val_batch_size = 1
val_it = iterator( \
test_mini_batch_loader, \
val_batch_size, repeat=False, shuffle=False)
return train_it, val_it, train_mini_batch_loader.__len__()
def test(model):
# データセットのロード
test_mini_batch_loader = \
DatasetPreProcessor(chainer.global_config.user_test_args)
test_data_size = test_mini_batch_loader.__len__()
print("------test data size")
print(test_data_size)
sum_accuracy = 0
sum_iou = 0
for idx in range(test_data_size):
raw_x, raw_t = test_mini_batch_loader.load_data([idx])
x = chainer.Variable(raw_x)
t = chainer.Variable(raw_t)
model(x, t)
print("accuracy:", model.accuracy)
sum_accuracy += model.accuracy
sum_iou += np.array(model.iou)
# 出力ラベルを画像に変換
result_labels = model.h.data.argmax(axis=1)
color_data = pd.read_csv(chainer.global_config.user_test_args.label_path)
restoration(
result_labels,
color_data,
chainer.global_config.user_test_args.output_path,
idx,
chainer.global_config.user_test_args.n_class)
print("test mean accuracy {}".format(sum_accuracy/test_data_size))
print("test mean IoU {}, meanIU {}" \
.format(sum_iou/test_data_size, np.sum(sum_iou/test_data_size)/n_class))
h, w, _ = raw_image.shape
mask, target_class = \
self.compute_attention_mask(preprocessed_image, gt)
mask = mask[0].transpose(1, 2, 0).astype(np.uint8)
xp = cuda.get_array_module(mask)
if xp != np:
mask = self.xp.asnumpy(mask)
mask = cv2.resize(mask, (w, h))
return raw_image * mask, target_class
if __name__ == '__main__':
args = get_args()
mini_batch_loader = DatasetPreProcessor(args)
visualizer = AttentionVisualizer(args)
_, model_eval = EasyTrainer.prepare_model(args)
args.net = model_eval
for idx, (image_path, label) in enumerate(mini_batch_loader.pairs):
raw_image = cv2.imread(image_path)
preprocessed_image, _ = mini_batch_loader.get_example(idx)
attention_view, target_class = \
visualizer.visualize_attention(raw_image, preprocessed_image, label)
name, extension = os.path.basename(image_path).split('.')
image_fname = name + '_' + str(label) + '_' + str(
target_class) + '.' + extension
output_path = os.path.join(args.output_path,