def __call__(self, src, label):
"""Apply transform to training image/label."""
# random color jittering
img = experimental.image.random_color_distort(src)
# random expansion with prob 0.5
if np.random.uniform(0, 1) > 0.5:
img, expand = timage.random_expand(
img, fill=[m * 255 for m in self._mean])
bbox = tbbox.translate(label,
x_offset=expand[0],
y_offset=expand[1])
else:
img, bbox = img, label
# random cropping
h, w, _ = img.shape
bbox, crop = experimental.bbox.random_crop_with_constraints(
bbox, (w, h))
x0, y0, w, h = crop
img = mx.image.fixed_crop(img, x0, y0, w, h)
# resize with random interpolation
h, w, _ = img.shape
interp = np.random.randint(0, 5)
img = timage.imresize(img, self._width, self._height, interp=interp)
bbox = tbbox.resize(bbox, (w, h), (self._width, self._height))
# random horizontal flip
h, w, _ = img.shape
img, flips = timage.random_flip(img, px=0.5)
bbox = tbbox.flip(bbox, (w, h), flip_x=flips[0])
if (self._bilateral_kernel_size is not None) and (
self._sigma_vals is not None) or self._grayscale:
img = img.asnumpy()
if (self._bilateral_kernel_size is not None) and (self._sigma_vals
is not None):
img = cv2.bilateralFilter(img, self._bilateral_kernel_size,
self._sigma_vals, self._sigma_vals)
if self._grayscale:
img = np.dot(img[..., :3], [0.299, 0.587, 0.114])
img = np.repeat(img[:, :, None], 3, axis=2)
img = nd.array(img)
# to tensor
img = mx.nd.image.to_tensor(img)
img = mx.nd.image.normalize(img, mean=self._mean, std=self._std)
if self._anchors is None:
return img, bbox.astype(img.dtype)
# generate training target so cpu workers can help reduce the workload on gpu
gt_bboxes = mx.nd.array(bbox[np.newaxis, :, :4])
gt_ids = mx.nd.array(bbox[np.newaxis, :, 4:5])
cls_targets, box_targets, _ = self._target_generator(
self._anchors, None, gt_bboxes, gt_ids)
return img, cls_targets[0], box_targets[0]
def __call__(self, src, label):
"""Apply transform to training image/label."""
# random color jittering
img = experimental.image.random_color_distort(src)
# random expansion with prob 0.5
if np.random.uniform(0, 1) > 0.5:
img, expand = timage.random_expand(
img, fill=[m * 255 for m in self._mean])
bbox = tbbox.translate(label,
x_offset=expand[0],
y_offset=expand[1])
else:
img, bbox = img, label
# random cropping
h, w, _ = img.shape
bbox, crop = experimental.bbox.random_crop_with_constraints(
bbox, (w, h))
x0, y0, w, h = crop
img = mx.image.fixed_crop(img, x0, y0, w, h)
# resize with random interpolation
h, w, _ = img.shape
interp = np.random.randint(0, 5)
img = timage.imresize(img, self._width, self._height, interp=interp)
bbox = tbbox.resize(bbox, (w, h), (self._width, self._height))
# random horizontal flip
h, w, _ = img.shape
img, flips = timage.random_flip(img, px=0.5)
bbox = tbbox.flip(bbox, (w, h), flip_x=flips[0])
# to tensor
img = mx.nd.image.to_tensor(img)
img = mx.nd.image.normalize(img, mean=self._mean, std=self._std)
if self._target_generator is None:
return img, bbox.astype(img.dtype)
# generate training target so cpu workers can help reduce the workload on gpu
gt_bboxes = mx.nd.array(bbox[np.newaxis, :, :4])
gt_ids = mx.nd.array(bbox[np.newaxis, :, 4:5]) # make the one-hot here
if self._mixup:
gt_mixratio = mx.nd.array(bbox[np.newaxis, :, -1:])
else:
gt_mixratio = None
objectness, center_targets, scale_targets, weights, class_targets = self._target_generator(
self._fake_x, self._feat_maps, self._anchors, self._offsets,
gt_bboxes, gt_ids, gt_mixratio)
return (img, objectness[0], center_targets[0], scale_targets[0],
weights[0], class_targets[0], gt_bboxes[0])
def __call__(self, src, label):
"""Apply transform to training image/label."""
# random color jittering
img = experimental.image.random_color_distort(src)
# random expansion with prob 0.5
if np.random.uniform(0, 1) > 0.5:
img, expand = timage.random_expand(
img, fill=[m * 255 for m in self._mean])
bbox = tbbox.translate(label,
x_offset=expand[0],
y_offset=expand[1])
else:
img, bbox = img, label
# random cropping
h, w, _ = img.shape
bbox, crop = experimental.bbox.random_crop_with_constraints(
bbox, (w, h))
x0, y0, w, h = crop
img = mx.image.fixed_crop(img, x0, y0, w, h)
# resize with random interpolation
h, w, _ = img.shape
interp = np.random.randint(0, 5)
img = timage.imresize(img, self._width, self._height, interp=interp)
bbox = tbbox.resize(bbox, (w, h), (self._width, self._height))
# random horizontal flip
h, w, _ = img.shape
img, flips = timage.random_flip(img, px=0.5)
bbox = tbbox.flip(bbox, (w, h), flip_x=flips[0])
# to tensor
img = mx.nd.image.to_tensor(img)
img = mx.nd.image.normalize(img, mean=self._mean, std=self._std)
if self._anchors is None:
return img, bbox.astype(img.dtype)
# generate training target so cpu workers can help reduce the workload on gpu
gt_bboxes = mx.nd.array(bbox[np.newaxis, :, :4])
gt_ids = mx.nd.array(bbox[np.newaxis, :, 4:5])
anchor_cls_targets, anchor_box_targets, _ = self._target_generator(
self._anchors, None, gt_bboxes, gt_ids)
anchor_cls_targets = mx.nd.where(anchor_cls_targets > 0,
mx.nd.ones_like(anchor_cls_targets),
anchor_cls_targets)
# positive anchor is 1, negative anchor is 0 and ignored is -1.
return img, anchor_cls_targets[0], anchor_box_targets[
0], bbox[:, :5].astype(img.dtype)
def __call__(self, src, label):
"""Apply transform to training image/label."""
# random color jittering
img = experimental.image.random_color_distort(src)
# random expansion with prob 0.5
if np.random.uniform(0, 1) > 0.5:
img, expand = timage.random_expand(img, fill=[m * 255 for m in self._mean])
bbox = tbbox.translate(label, x_offset=expand[0], y_offset=expand[1])
else:
img, bbox = img, label
# random cropping
h, w, _ = img.shape
bbox, crop = experimental.bbox.random_crop_with_constraints(bbox, (w, h))
x0, y0, w, h = crop
img = mx.image.fixed_crop(img, x0, y0, w, h)
# resize with random interpolation
h, w, _ = img.shape
interp = np.random.randint(0, 5)
img = timage.imresize(img, self._width, self._height, interp=interp)
bbox = tbbox.resize(bbox, (w, h), (self._width, self._height))
# random horizontal flip
h, w, _ = img.shape
img, flips = timage.random_flip(img, px=0.5)
bbox = tbbox.flip(bbox, (w, h), flip_x=flips[0])
# to tensor
img = mx.nd.image.to_tensor(img)
img = mx.nd.image.normalize(img, mean=self._mean, std=self._std)
if self._anchors is None:
return img, bbox.astype(img.dtype)
#如果有anchors的输入，则执行下面的运算。计算以前的格式是：
#gt_bboxes里面是一个图像上的box位置，实际像素点位 当前图片里面的框框个数x4
#gt_ids里面是对应图像上的label，当前图片里面的框框个数x1
#下面为batch_size腾出空间
gt_bboxes = mx.nd.array(bbox[np.newaxis, :, :4])
gt_ids = mx.nd.array(bbox[np.newaxis, :, 4:5])
cls_targets, box_targets, _ = self._target_generator(
self._anchors, None, gt_bboxes, gt_ids)
return img, cls_targets[0], box_targets[0]
def new_trainloader_call(self, src, label):
'''
define a new call for trainloader by changing the data augmentation
'''
# random color jittering
img = experimental.image.random_color_distort(src)
# random expansion with prob 0.5
if np.random.uniform(0, 1) > 0.5:
img, expand = timage.random_expand(img,
fill=[m * 255 for m in self._mean])
bbox = tbbox.translate(label, x_offset=expand[0], y_offset=expand[1])
else:
img, bbox = img, label
# random cropping
h, w, _ = img.shape
bbox, crop = experimental.bbox.random_crop_with_constraints(bbox, (w, h))
x0, y0, w, h = crop
img = mx.image.fixed_crop(img, x0, y0, w, h)
# resize with random interpolation
h, w, _ = img.shape
interp = np.random.randint(0, 5)
img = timage.imresize(img, self._width, self._height, interp=interp)
bbox = tbbox.resize(bbox, (w, h), (self._width, self._height))
# to tensor
img = mx.nd.image.to_tensor(img)
img = mx.nd.image.normalize(img, mean=self._mean, std=self._std)
if self._anchors is None:
return img, bbox.astype(img.dtype)
# generate training target so cpu workers can help reduce the workload on gpu
gt_bboxes = mx.nd.array(bbox[np.newaxis, :, :4])
gt_ids = mx.nd.array(bbox[np.newaxis, :, 4:5])
cls_targets, box_targets, _ = self._target_generator(
self._anchors, None, gt_bboxes, gt_ids)
return img, cls_targets[0], box_targets[0]
def __call__(self, src, label):
"""Apply transform to training image/label."""
# random color jittering
img = experimental.image.random_color_distort(src)
# img, bbox = img,label
# random expansion with prob 0.5
if np.random.uniform(0, 1) > 0.5:
img, expand = timage.random_expand(
img, fill=[m * 255 for m in self._mean])
bbox = tbbox.translate(label,
x_offset=expand[0],
y_offset=expand[1])
else:
img, bbox = img, label
# random cropping
h, w, _ = img.shape
bbox, crop = experimental.bbox.random_crop_with_constraints(
bbox, (w, h))
x0, y0, w, h = crop
img = mx.image.fixed_crop(img, x0, y0, w, h)
# resize with random interpolation
h, w, _ = img.shape
interp = np.random.randint(0, 5)
img = timage.imresize(img, self._width, self._height, interp=interp)
bbox = tbbox.resize(bbox, (w, h), (self._width, self._height))
# random horizontal flip
# h, w, _ = img.shape
# img, flips = timage.random_flip(img, px=0.5)
# bbox = tbbox.flip(bbox, (w, h), flip_x=flips[0])
# rabdom rotation
h, w, _ = img.shape
clockwise_rotation_num = np.random.randint(0, 4)
if clockwise_rotation_num == 0:
pass
elif clockwise_rotation_num == 1:
###顺时针90度
img = nd.transpose(img, [1, 0, 2])
img = img[:, ::-1, :]
bbox = np.array([
h - bbox[:, 3], bbox[:, 0], h - bbox[:, 1], bbox[:, 2],
bbox[:, 4], bbox[:, 5]
]).T
bbox[:, 5] = (bbox[:, 5] + 1) % 4
elif clockwise_rotation_num == 2:
##顺时针180度
img = img[::-1, ::-1, :]
bbox = np.array([
w - bbox[:, 2], h - bbox[:, 3], w - bbox[:, 0], h - bbox[:, 1],
bbox[:, 4], bbox[:, 5]
]).T
bbox[:, 5] = (bbox[:, 5] + 2) % 4
else:
# 顺时针270度
img = nd.transpose(img, [1, 0, 2])
img = img[::-1, :, :]
bbox = np.array([
bbox[:, 1], w - bbox[:, 2], bbox[:, 3], w - bbox[:, 0],
bbox[:, 4], bbox[:, 5]
]).T
bbox[:, 5] = (bbox[:, 5] + 3) % 4
# to tensor
img = mx.nd.image.to_tensor(img)
img = mx.nd.image.normalize(img, mean=self._mean, std=self._std)
if self._anchors is None:
return img, bbox.astype(img.dtype)
# generate training target so cpu workers can help reduce the workload on gpu
gt_bboxes = mx.nd.array(bbox[np.newaxis, :, :4])
gt_ids = mx.nd.array(bbox[np.newaxis, :, 4:5])
gt_ori = mx.nd.array(bbox[np.newaxis, :, 5:6])
cls_targets, box_targets, _ = self._target_generator(
self._anchors, None, gt_bboxes, gt_ids)
ori_targets, box_targets, _ = self._target_generator(
self._anchors, None, gt_bboxes, gt_ori)
return img, cls_targets[0], ori_targets[0], box_targets[0]