def generate(self):
""" The __getitem__ method
so that the object can be iterable
Args:
index: the index to get filename from self.ids
Returns:
img: tensor of shape (300, 300, 3)
boxes: tensor of shape (num_gt, 4)
labels: tensor of shape (num_gt,)
"""
for index in range(len(self.ids)):
# img, orig_shape = self._get_image(index)
img = self._get_image(index)
w, h = img.size
boxes, labels = self._get_annotation(index, (h, w))
boxes = tf.constant(boxes, dtype=tf.float32)
labels = tf.constant(labels, dtype=tf.int64)
augmentation_method = np.random.choice(self.augmentation)
if augmentation_method == 'patch':
img, boxes, labels = random_patching(img, boxes, labels)
elif augmentation_method == 'flip':
img, boxes, labels = horizontal_flip(img, boxes, labels)
img = np.array(img.resize((self.new_size, self.new_size)),
dtype=np.float32)
img = (img / 127.0) - 1.0
img = tf.constant(img, dtype=tf.float32)
gt_confs, gt_locs = compute_target(self.default_boxes, boxes,
labels)
yield img, gt_confs, gt_locs
def generate(self, mode="train"):
""" The __getitem__ method
so that the object can be iterable
Args:
Returns:
img: tensor of shape (300, 300, 3)
boxes: tensor of shape (num_gt, 4)
labels: tensor of shape (num_gt,)
"""
while (True):
self.on_end_epoch()
for index in range(self.__len__()):
imgs, gt_confs_all, gt_locs_all = [], [], []
for i in range(self.batch_size):
if (mode == 'train'):
sample = next(
tfds.as_numpy(
self.train_dataset.take(self.batch_size)))
else:
sample = next(
tfds.as_numpy(
self.val_dataset.take(self.batch_size)))
img = sample['image']
boxes = sample['objects']['bbox']
labels = sample['objects']['label']
# orig_shape = img.shape
# # img, orig_shape = self._get_image(index)
# filename = indices[index]
# img = self._get_image(index)
# w, h = img.size
# boxes, labels = self._get_annotation(index, (h, w))
boxes = tf.constant(boxes, dtype=tf.float32)
labels = tf.constant(labels + 1,
dtype=tf.int64) # 0 is background
# augmentation_method = np.random.choice(self.augmentation)
# if augmentation_method == 'patch':
# img, boxes, labels = random_patching(img, boxes, labels)
# elif augmentation_method == 'flip':
# img, boxes, labels = horizontal_flip(img, boxes, labels)
# print((self.input_shape, self.input_shape))
img = cv2.resize(img, (self.input_shape, self.input_shape))
# img = np.array(img.resize(
# (self.input_shape, self.input_shape)), dtype=np.float32)
img = (img / 127.0) - 1.0
img = tf.constant(img, dtype=tf.float32)
# if(len(boxes) == 0):
gt_confs, gt_locs = compute_target(self.default_boxes,
boxes, labels)
imgs.append(img)
gt_confs_all.append(gt_confs)
gt_locs_all.append(gt_locs)
yield tf.stack(imgs), tf.stack(gt_confs_all), tf.stack(
gt_locs_all)
def generate(self, subset=None):
""" The __getitem__ method
so that the object can be iterable
Args:
index: the index to get filename from self.ids
Returns:
img: tensor of shape (300, 300, 3)
boxes: tensor of shape (num_gt, 4)
labels: tensor of shape (num_gt,)
"""
# 만약 Train인 경우 File적용 Test라면 Filp 적용 X
if subset == 'train':
indices = self.train_ids
# 3. Random하게 Original Image를 사용할지 Flip을 실행할 Image를 사용할지 결정한다.
augmentation_method = np.random.choice(self.augmentation)
if augmentation_method == 'flip':
img, boxes, labels = horizontal_flip(img, boxes, labels)
elif subset == 'val':
indices = self.val_ids
else:
indices = self.ids
for index in range(len(indices)):
# img, orig_shape = self._get_image(index)
filename = indices[index]
img = self._get_image(index)
# 1. Input Image의 Size를 받는다.
w, h = img.size
# 2. get_annotation()을 통하여 Label과 Bounding Box의 Location을 입력받는다.
boxes, labels = self._get_annotation(index, (h, w))
boxes = tf.constant(boxes, dtype=tf.float32)
labels = tf.constant(labels, dtype=tf.int64)
# 4. Image의 Size를 Model Input에 맞게 (300,300)으로 바꾼뒤 0 ~ 1 사이의 값으로서 정규화를 한다.
img = np.array(img.resize((self.new_size, self.new_size)),
dtype=np.float32)
img = (img / 127.0) - 1.0
img = tf.constant(img, dtype=tf.float32)
# 5. Utils -> box_utils -> compute_target를 통하여 실제 Label을 Model에 맞는 Label로서 변경한다.
gt_confs, gt_locs = compute_target(self.default_boxes, boxes,
labels)
# 6. Filename, Image, Ground Truth Label, Ground Truth Location을 반환한다
# Generator로서 특정 Index후 다음 Index로 반환하기 위하여 Return 값을 yield로서 선언
yield filename, img, gt_confs, gt_locs
def generate(self, subset=None):
""" The __getitem__ method
so that the object can be iterable
Args:
index: the index to get filename from self.ids
Returns:
img: tensor of shape (300, 300, 3)
boxes: tensor of shape (num_gt, 4)
labels: tensor of shape (num_gt,)
"""
if subset == 'train':
indices = self.train_ids
elif subset == 'val':
indices = self.val_ids
else:
indices = self.ids
for index in range(len(indices)):
# img, orig_shape = self._get_image(index)
filename = indices[index]
img = self._get_image(index)
w, h = img.size
boxes, labels = self._get_annotation(index, (h, w))
boxes = tf.constant(boxes, dtype=tf.float32)
labels = tf.constant(labels, dtype=tf.int64)
augmentation_method = np.random.choice(self.augmentation)
if augmentation_method == 'patch':
img, boxes, labels = random_patching(img, boxes, labels)
elif augmentation_method == 'flip':
img, boxes, labels = horizontal_flip(img, boxes, labels)
img = np.array(img.resize((self.new_size, self.new_size)),
dtype=np.float32)
img = np.repeat(img[:, :, np.newaxis], 3,
axis=2) # transform grayscale into "forced" rgb
img = (img / 127.0) - 1.0
img = tf.constant(img, dtype=tf.float32)
# print(tf.shape(img))
gt_confs, gt_locs = compute_target(self.default_boxes, boxes,
labels)
yield filename, img, gt_confs, gt_locs
def generate(self, split, coco, ids, num_examples, config):
"""
num_examples : The number of examples to be used.
It's used if you want to make model overfit a few examples
"""
for image_id in ids:
filename, image, original_size = self._get_image(image_id, coco)
boxes, labels = self._get_annotation(image_id, coco)
if len(labels) == 0:
continue
#assert image.shape[:2] == original_size[:2]
height, width = original_size
#normalize bounding box coord (range in 0~1.0)
boxes = list(map(lambda box : (box[0]/width, box[1]/height, box[2]/width, box[3]/height), boxes))
boxes = np.array(boxes, np.float32); labels = np.array(labels, np.float32);
gt_confs, gt_locs = box_utils.compute_target(self.default_boxes, boxes, labels)
image = cv2.resize(image, self.input_shape[:2], interpolation = cv2.INTER_AREA)
if (config['inference_mode'] == 'mAP'):
yield filename, image, labels, boxes
else:
yield filename, image, gt_confs, gt_locs
def __getitem__(self, index):
""" The __getitem__ method
so that the object can be iterable
Args:
index: the index to get filename from self.ids
Returns:
img: Torch tensor of shape (3, 300, 300)
boxes: Torch tensor of shape (num_gt, 4)
labels: Torch tensor of shape (num_gt,)
"""
img, orig_shape = self._get_image(index)
boxes, labels = self._get_annotation(index, orig_shape)
boxes = torch.from_numpy(boxes)
labels = torch.from_numpy(labels)
if self.augmentation:
if np.random.rand() > 0.5:
img, boxes = self._random_flip(img, boxes)
gt_confs, gt_locs = compute_target(self.default_boxes, boxes, labels)
return img, gt_confs, gt_locs
def generate(self, subset=None, num=20000):
""" The __getitem__ method
so that the object can be iterable
Args:
index: the index to get filename from self.ids
Returns:
img: tensor of shape (300, 300, 3)
boxes: tensor of shape (num_gt, 4)
labels: tensor of shape (num_gt,)
"""
if subset == 'train':
indices = self.train_ids
elif subset == 'val':
indices = self.val_ids
else:
indices = self.ids
indices = np.array(indices)
counter = 0
while (counter < num):
counter = counter + 1
itemp = np.random.randint(0, len(indices), size=[self.batch_size])
# indexes = indices[itemp]
# img, orig_shape = self._get_image(index)
# filename = indices[index]
imgs, gt_confs_all, gt_scores_all, gt_locs_all, o_imgs, o_boxes, o_labels = [], [],[], [], [], [], []
for index in itemp:
img = self._get_image(index)
w, h = img.size
boxes, labels = self._get_annotation(index, (h, w))
boxes = tf.constant(boxes, dtype=tf.float32)
labels = tf.constant(labels, dtype=tf.int64)
augmentation_method = np.random.choice(self.augmentation)
if augmentation_method == 'patch':
img, boxes, labels = random_patching(img, boxes, labels)
img = np.float32(img)
elif augmentation_method == 'flip':
img, boxes, labels = horizontal_flip(
np.float32(img), boxes, labels)
else:
img = np.array(img, dtype=np.float32)
# print(img.shape, img.dtype)
img_o = cv2.resize(img, (self.new_size, self.new_size))
img = (img_o / 127.0) - 1.0
img = tf.constant(img, dtype=tf.float32)
gt_confs, gt_locs, gt_scores = compute_target(
self.default_boxes, boxes, labels)
imgs.append(img)
gt_confs_all.append(gt_confs)
gt_locs_all.append(gt_locs)
gt_scores_all.append(gt_scores)
o_imgs.append(img_o)
o_boxes.append(boxes)
o_labels.append(labels)
# print('1', img.dtype, gt_confs.dtype, gt_locs.dtype)
imgs, gt_confs_all, gt_locs_all = tf.stack(imgs), tf.stack(
gt_confs_all), tf.stack(gt_locs_all)
# print('----', np.max(imgs.numpy()), np.max(gt_confs_all.numpy()), np.max(gt_locs_all))
yield tf.stack(imgs), tf.stack(gt_confs_all), tf.stack(
gt_locs_all), tf.stack(
gt_scores_all), o_imgs, o_boxes, o_labels
try:
config = cfg['SSD300'] #[args.arch.upper()]
except AttributeError:
pass
default_boxes = generate_default_boxes(config)
print(default_boxes)
import tensorflow_datasets as tfds
builder = tfds.builder('coco/' + '2017')
# 1. Create the tfrecord files (no-op if already exists)
builder.download_and_prepare()
# 2. Load the `tf.data.Dataset`
train_dataset = builder.as_dataset(split='train', shuffle_files=True)
val_dataset = builder.as_dataset(split='validation', shuffle_files=False)
info = builder.info
train_dataset = train_dataset
for i in range(10):
sample = next(tfds.as_numpy(train_dataset.take(1)))
print(sample['image/filename'])
img = sample['image']
boxes = sample['objects']['bbox']
labels = sample['objects']['label']
boxes = tf.constant(boxes, dtype=tf.float32)
labels = tf.constant(labels, dtype=tf.int64)
gt_confs, gt_locs = compute_target(default_boxes, boxes, labels)
def generate(self, num=20000):
""" The __getitem__ method
so that the object can be iterable
Args:
Returns:
img: tensor of shape (300, 300, 3)
boxes: tensor of shape (num_gt, 4)
labels: tensor of shape (num_gt,)
"""
counter = 0
while (counter < num):
counter = counter + 1
imgs, gt_confs_all, gt_locs_all,gt_scores_all, o_imgs, o_boxes, o_labels = [], [],[], [], [], [], []
for i in range(self.batch_size):
img, boxes, labels, shape = None, None, None, None
while (True):
self.global_index = self.global_index + 1
if (self.global_index == self.__len__()):
self.on_end_epoch()
self.global_index = 0
img, boxes, labels, shape = self.read_img(
self.global_index)
if (img is not None):
if (len(labels) > 0):
break
# orig_shape = img.shape
# # img, orig_shape = self._get_image(index)
# filename = indices[index]
# img = self._get_image(index)
# w, h = img.size
# boxes, labels = self._get_annotation(index, (h, w))
boxes = tf.constant(boxes, dtype=tf.float32)
labels = tf.constant(labels, dtype=tf.int64)
augmentation_method = np.random.choice(self.augmentation)
if augmentation_method == 'patch':
# print('patching')
img, boxes, labels = random_patching(img, boxes, labels)
elif augmentation_method == 'flip':
# print('flipping')
img, boxes, labels = horizontal_flip(img, boxes, labels)
# print((self.input_shape, self.input_shape))
img = cv2.resize(img,
(self.input_shape[0], self.input_shape[1]))
# img = np.array(img.resize(
# (self.input_shape, self.input_shape)), dtype=np.float32)
o_imgs.append(img)
img = img / 127.0 - 1
img = tf.constant(img, dtype=tf.float32)
# if(len(boxes) == 0):
gt_confs, gt_locs, gt_scores = compute_target(
self.default_boxes, boxes, labels)
# gt_confs, gt_locs, gt_scores = tf_ssd_bboxes_encode(labels, boxes, self.default_boxes, len(self.cat_names) + 1, None)
imgs.append(img)
gt_confs_all.append(gt_confs)
gt_locs_all.append(gt_locs)
gt_scores_all.append(gt_scores)
o_boxes.append(boxes)
o_labels.append(labels)
yield tf.stack(imgs), tf.stack(gt_confs_all), tf.stack(
gt_locs_all), tf.stack(
gt_scores_all), o_imgs, o_boxes, o_labels