def __getitem__(self, idx):
path, cid = self._data[idx]
img = cv2.imread(path, 1)
img = np.pad(img,
((self._pad, self._pad), (self._pad, self._pad), (0, 0)),
mode='constant',
constant_values=0)
img = nd.array(img, dtype='float32') / 255
H, W, C = img.shape
data_shape = (C, H - self._pad * 2, W - self._pad * 2)
if self._fortrain:
augs = image.CreateAugmenter(data_shape=data_shape,
resize=0,
rand_mirror=True,
rand_crop=True,
mean=self._mean,
std=self._std)
else:
augs = image.CreateAugmenter(data_shape=data_shape,
resize=0,
rand_mirror=False,
rand_crop=False,
mean=self._mean,
std=self._std)
for aug in augs:
img = aug(img)
img = nd.transpose(img, (2, 0, 1))
return img, nd.array([cid]).astype("float32")
def transform_train(data, label):
#将图像调整为不同的大小，分别进行数据扩增,因为是使用两个模型进行融合，所以也使用两类数据分别给两个网络进行训练
im1 = image.imresize(data.astype('float32') / 255, 224,
224) #将图像调整为224x224，data的每个像素点除以255,保证在[0-1]
im2 = image.imresize(data.astype('float32') / 255, 299, 299)
#数据增强参数1，给第一个网络
auglist1 = image.CreateAugmenter(data_shape=(3, 224, 224),
resize=0,
rand_crop=False,
rand_resize=False,
rand_mirror=True,
mean=np.array([0.485, 0.456, 0.406]),
std=np.array([0.229, 0.224, 0.225]),
brightness=0,
contrast=0,
saturation=0,
hue=0,
pca_noise=0,
rand_gray=0,
inter_method=2)
# 数据增强参数2，给第二个网络
auglist2 = image.CreateAugmenter(data_shape=(3, 299, 299),
resize=0,
rand_crop=False,
rand_resize=False,
rand_mirror=True,
mean=np.array([0.485, 0.456, 0.406]),
std=np.array([0.229, 0.224, 0.225]),
brightness=0,
contrast=0,
saturation=0,
hue=0,
pca_noise=0,
rand_gray=0,
inter_method=2)
#分别进行增强
for aug in auglist1:
im1 = aug(im1)
for aug in auglist2:
im2 = aug(im2)
# 将数据格式从"高*宽*通道"改为"通道*高*宽"。
im1 = nd.transpose(im1, (2, 0, 1))
im2 = nd.transpose(im2, (2, 0, 1))
#返回给两个数据，分别给两个网络进行训练
return (im1, im2, nd.array([label]).asscalar().astype('float32'))
def transform_test(data, label):
im = data.astype('float32') / 255
auglist = image.CreateAugmenter(data_shape=(3, 224, 224))
for aug in auglist:
im = aug(im)
im = nd.transpose(im, (2, 0, 1))
return im, nd.array([label]).asscalar().astype('float32')
def transform_test(data, label):
im = data.astype('float32') / 255.0
auglist = image.CreateAugmenter(data_shape=(3, 32, 32), mean=np.array([0.4914, 0.4822, 0.4465]),
std=np.array([0.2023, 0.1994, 0.2010]))
for aug in auglist:
im = aug(im)
im = nd.transpose(im, (2, 0, 1))
return im, nd.array([label]).astype('float32')
def transform_test(data, label):
im1 = image.imresize(data.astype('float32') / 255, 224, 224)
im2 = image.imresize(data.astype('float32') / 255, 299, 299)
auglist1 = image.CreateAugmenter(data_shape=(3, 224, 224),
mean=np.array([0.485, 0.456, 0.406]),
std=np.array([0.229, 0.224, 0.225]))
auglist2 = image.CreateAugmenter(data_shape=(3, 299, 299),
mean=np.array([0.485, 0.456, 0.406]),
std=np.array([0.229, 0.224, 0.225]))
for aug in auglist1:
im1 = aug(im1)
for aug in auglist2:
im2 = aug(im2)
# 将数据格式从"高*宽*通道"改为"通道*高*宽"。
im1 = nd.transpose(im1, (2,0,1))
im2 = nd.transpose(im2, (2,0,1))
return (im1,im2, nd.array([label]).asscalar().astype('float32'))
def transform_train(data, label):
im = data.astype('float32') / 255
auglist = image.CreateAugmenter(data_shape=(3, 32, 32))
for aug in auglist:
im = aug(im)
# 将数据格式从"高*宽*通道"改为"通道*高*宽"。
im = nd.transpose(im, (2, 0, 1))
return im, nd.array([label]).asscalar().astype('float32')
def trans_test(data, label):
im = data.astype(np.float32) / 255.
auglist = image.CreateAugmenter(data_shape=(3, 32, 32),
mean=mx.nd.array([0.485, 0.456, 0.406]),
std=mx.nd.array([0.229, 0.224, 0.225]))
for aug in auglist:
im = aug(im)
im = nd.transpose(im, (2, 0, 1))
return im, label
def transform_test(data, label):
im = data.astype('float32') / 255
auglist = image.CreateAugmenter(data_shape(3, 32, 32), resize=0,
mean=np.array([0.4914, 0.4822, 0.4465]),
std=np.array([0.2023, 0.1994, 0.2010]))
for aug in auglist:
im = aug(im)
# change data from height*width*channel to channel*height*width
im = nd.transpose(im, (2,0,1))
return (im, nd.array([label]).asscalar().astype('float32'))
def transform_test(data, label):
im = data.astype("float32") / 255
auglist = image.CreateAugmenter(data_shape=(3, 32, 32),
mean=np.array([0., 0., 0.]),
std=np.array([1, 1, 1]))
for aug in auglist:
im = aug(im)
im = nd.transpose(im, (2, 0, 1))
return im, nd.array([label]).asscalar().astype("float32")
def transform_train(data, label):
im = data.astype('float32') / 255
auglist = image.CreateAugmenter(data_shape=(3, 224, 224), resize=256,
rand_crop=True, rand_mirror=True,
mean = np.array([0.485, 0.456, 0.406]),
std = np.array([0.229, 0.224, 0.225]))
for aug in auglist:
im = aug(im)
im = nd.transpose(im, (2,0,1))
return (im, nd.array([label]).asscalar())
def trans_test(data, label):
im = data
auglist = image.CreateAugmenter(data_shape=(3, 224, 224),
resize=256,
mean=True,
std=True)
for aug in auglist:
im = aug(im)
im = nd.transpose(im, (2, 0, 1))
return im, label
def transform_test(data):
im1 = image.imresize(data.astype('float32') / 255, 288, 288)
auglist1 = image.CreateAugmenter(data_shape=(3, 224, 224),
resize=0,
mean=np.array([0.485, 0.456, 0.406]),
std=np.array([0.229, 0.224, 0.225]))
im2 = image.imresize(data.astype('float32') / 255, 363, 363)
auglist2 = image.CreateAugmenter(data_shape=(3, 299, 299),
resize=0,
mean=np.array([0.485, 0.456, 0.406]),
std=np.array([0.229, 0.224, 0.225]))
for aug in auglist1:
im1 = aug(im1)
im1 = nd.transpose(im1, (2, 0, 1))
for aug in auglist2:
im2 = aug(im2)
im2 = nd.transpose(im2, (2, 0, 1))
return im1.expand_dims(axis=0), im2.expand_dims(axis=0)
def transformer(data, label):
im = data.asnumpy()
im = np.pad(im, pad_width=((4, 4), (4, 4), (0, 0)), mode='constant')
im = random_eraser(im)
im = nd.array(im) / 255.
auglist = image.CreateAugmenter(data_shape=(3, 32, 32), rand_crop=True, rand_mirror=True,
mean=mx.nd.array([0.4914, 0.4824, 0.4467]),
std=mx.nd.array([0.2471, 0.2435, 0.2616]))
for aug in auglist:
im = aug(im)
im = nd.transpose(im, (2, 0, 1))
return im, label
def transform_train(data, label):
im = image.imresize(data.astype('float32') / 255, 96, 96)
auglist = image.CreateAugmenter(data_shape=(3, 96, 96), resize=0,
rand_crop=False, rand_resize=False, rand_mirror=True,
mean=None, std=None,
brightness=0, contrast=0,
saturation=0, hue=0,
pca_noise=0, rand_gray=0, inter_method=2)
for aug in auglist:
im = aug(im)
# 将数据格式从"高*宽*通道"改为"通道*高*宽"。
im = nd.transpose(im, (2,0,1))
return (im, nd.array([label]).asscalar().astype('float32'))
def transform_train(data, label):
im1 = image.imresize(data.astype('float32') / 255, 224, 224)
im2 = image.imresize(data.astype('float32') / 255, 299, 299)
auglist1 = image.CreateAugmenter(data_shape=(3, 224, 224),
resize=0,
rand_crop=False,
rand_resize=False,
rand_mirror=True,
mean=np.array([0.485, 0.456, 0.406]),
std=np.array([0.229, 0.224, 0.225]),
brightness=0,
contrast=0,
saturation=0,
hue=0,
pca_noise=0,
rand_gray=0,
inter_method=2)
auglist2 = image.CreateAugmenter(data_shape=(3, 299, 299),
resize=0,
rand_crop=False,
rand_resize=False,
rand_mirror=True,
mean=np.array([0.485, 0.456, 0.406]),
std=np.array([0.229, 0.224, 0.225]),
brightness=0,
contrast=0,
saturation=0,
hue=0,
pca_noise=0,
rand_gray=0,
inter_method=2)
for aug in auglist1:
im1 = aug(im1)
for aug in auglist2:
im2 = aug(im2)
# 将数据格式从"高*宽*通道"改为"通道*高*宽"。
im1 = nd.transpose(im1, (2, 0, 1))
im2 = nd.transpose(im2, (2, 0, 1))
return (im1, im2, nd.array([label]).asscalar().astype('float32'))
def transform_train(data, label):
im = data.astype('float32') / 255
auglist = image.CreateAugmenter(data_shape=(3, 299, 299), resize=320,
rand_crop=True, rand_mirror=True,
mean = np.array([0.485, 0.456, 0.406]),
std = np.array([0.229, 0.224, 0.225]),
brightness=0.125, contrast=0.125,
saturation = 0.125, pca_noise = 0.05, inter_method = 10
)
for aug in auglist:
im = aug(im)
im = nd.transpose(im, (2,0,1))
return (im, nd.array([label]).asscalar())
def transform_train(data, label):
im = data.asnumpy()
# im = np.pad(im, ((4, 4), (4, 4), (0, 0)), mode='constant', constant_values=0)
im = np.pad(im, ((4, 4), (4, 4), (0, 0)), mode='reflect')
im = nd.array(im, dtype='float32') / 255.0
auglist = image.CreateAugmenter(data_shape=(3, 32, 32), resize=0, rand_mirror=True,
rand_crop=True,
mean=np.array([0.4914, 0.4822, 0.4465]),
std=np.array([0.2023, 0.1994, 0.2010]))
for aug in auglist:
im = aug(im)
im = nd.transpose(im, (2, 0, 1)) # channel x height x weight
return im, nd.array([label]).astype('float32')
def transform_train(data, label):
im = data.astype('float32') / 255
aug_list = image.CreateAugmenter(data_shape=(3, 32, 32), resize=0,
rand_crop=False, rand_resize=False, rand_mirror=True,
mean=np.array([0.4914, 0.4822, 0.4465]),
std=np.array([0.2023, 0.1994, 0.2010]),
brightness=0, contrast=0,
saturation=0, hue=0,
pca_noise=0, rand_gray=0, inter_method=2)
for aug in aug_list:
im = aug(im)
# 将数据格式从"高*宽*通道"改为"通道*高*宽"
im = nd.transpose(im, (2, 0, 1))
return im, nd.array([label]).asscalar().astype('float32')
def transform_train(data, label):
im = data.astype('float32') / 255
auglist = image.CreateAugmenter(data_shape(3, 32, 32), resize=0,
rand_crop=True, rand_resize=True, rand_mirror=True,
mean=np.array([0.4914, 0.4822, 0.4465]),
std=np.array([0.2023, 0.1994, 0.2010]),
brightness=0, contrast=0,
saturation=0, hue=0,
pca_noise=0, rand_gray=0, inter_method=2)
for aug in auglist:
im = aug(im)
# change data from height*width*channel to channel*height*width
im = nd.transpose(im, (2,0,1))
return (im, nd.array([label]).asscalar().astype('float32'))
def transform_train(data, label):
im = data.astype("float32") / 255
auglist = image.CreateAugmenter(data_shape=(3, 32, 32),
resize=0,
rand_crop=True, rand_resize=True, rand_gray=0.5, rand_mirror=True,
mean=np.array([0., 0., 0.]), std=np.array([1, 1, 1]),
brightness=0.5, contrast=0.5, saturation=0.5,
hue=0.5, pca_noise=0, inter_method=2)
for aug in auglist:
im = aug(im)
# 将数据格式从"高*宽*通道"改为"通道*高*宽"。
im = nd.transpose(im, (2, 0, 1))
return im, nd.array([label]).asscalar().astype("float32")
def transform_train(data, label):
im = data.asnumpy()
im = np.pad(im, ((4, 4), (4, 4), (0, 0)), mode='constant')
# 前半部分为np augment. 后半部分为nd augment，注意要求转换为float型
im = nd.array(im, dtype="float32") / 255
auglist = image.CreateAugmenter(data_shape=(3, 32, 32),
resize=0,
rand_mirror=True,
rand_crop=True,
mean=nd.array([0.4914, 0.4822, 0.4465]),
std=nd.array([0.2023, 0.1994, 0.2010]))
for aug in auglist:
im = aug(im)
im = im.transpose((2, 0, 1))
return im, label.astype('float32')
def transform_train_DA2(data, label):
im = data.astype(np.float32) / 255
auglist = [image.RandomSizedCropAug(size=(32, 32), min_area=0.49, ratio=(0.5, 2))]
_aug = image.CreateAugmenter(data_shape=(3, 32, 32), resize=0,
rand_crop=False, rand_resize=False, rand_mirror=True,
mean=np.array([0.4914, 0.4822, 0.4465]),
std=np.array([0.2023, 0.1994, 0.2010]),
brightness=0.3, contrast=0.3, saturation=0.3, hue=0.3,
pca_noise=0.01, rand_gray=0, inter_method=2)
auglist.append(image.RandomOrderAug(_aug))
for aug in auglist:
im = aug(im)
im = nd.transpose(im, (2, 0, 1))
return (im, nd.array([label]).asscalar().astype('float32'))
def augment(data):
aug_list = image.CreateAugmenter(data_shape=(3, 32, 32),
rand_crop=True,
rand_mirror=True)
data = nd.pad(data,
pad_width=(0, 0, 0, 0, 2, 2, 2, 2),
mode='constant',
constant_value=0)
data = nd.transpose(data, (0, 2, 3, 1))
temp = []
for d in data:
for aug in aug_list:
d = aug(d)
temp.append(d)
data = nd.stack(*temp)
data = nd.transpose(data, (0, 3, 1, 2))
return data
def transformer(data, label):
jitter_param = 0.4
lighting_param = 0.1
im = data
auglist = image.CreateAugmenter(data_shape=(3, 224, 224),
rand_crop=True,
rand_resize=True,
rand_mirror=True,
brightness=jitter_param,
saturation=jitter_param,
contrast=jitter_param,
pca_noise=lighting_param,
mean=True,
std=True)
for aug in auglist:
im = aug(im)
im = nd.transpose(im, (2, 0, 1))
return im, label
def __init__(self, batch_size, img_h, img_w, ctx):
self.h = img_h
self.w = img_w
self.bs = batch_size
self.ctx = ctx
self.BIL = PILIMG.BILINEAR
self.all_img = []
base_dir = os.path.dirname(os.path.abspath(__file__))
#png_img_path = os.path.join(base_dir, 'syn_images')
png_img_path = '/media/nolan/9fc64877-3935-46df-9ad0-c601733f5888/insulator_raw_images'
for img in os.listdir(png_img_path):
img_path = os.path.join(png_img_path, img)
self.all_img.append(img_path)
self.pil_image_enhance = PILImageEnhance(G=0.1, noise_var=2)
self.augs = image.CreateAugmenter(data_shape=(3, img_h, img_w),
inter_method=10, brightness=0.1, contrast=0.3, saturation=0.3, hue=1.0, pca_noise=0.1)
print('Loading background!')
def pre_processing(im_path,
data_shape,
resize=128,
mean_rgb=(0, 0, 0),
std_rgb=(1, 1, 1)):
try:
with open(im_path, 'rb') as f:
im = image.imdecode(f.read())
augmenter = image.CreateAugmenter(data_shape,
resize=resize,
mean=np.array(mean_rgb),
std=np.array(std_rgb))
for aug_func in augmenter:
im = aug_func(im)
im = im.transpose((2, 0, 1)).expand_dims(axis=0)
return im
except Exception as e:
print(e)
return None
def get_image_augs(shape, scope="train"):
augs_params = train_augs_params if scope == 'train' else test_augs_params
augs_params['data_shape'] = (0, shape, shape)
augs = image.CreateAugmenter(**augs_params)
return augs
type=str,
help='symbol path')
sgroup.add_argument('--output-blobs',
default='',
type=str,
help='output blocks name')
igroup = parser.add_argument_group()
igroup.add_argument('--arch', default='resnet101_v2', help='architecture')
igroup.add_argument('--num_classes',
default=1000,
type=int,
help='num classes: WebFace: 10575, VGG-Face2: 9131')
args = parser.parse_args()
args.im_exts = [ext.strip() for ext in args.im_exts.split(',')]
ctx = mx.gpu(args.gpu) if args.gpu >= 0 else mx.cpu()
data_shape = tuple(map(int, args.data_shape.split(',')))
mean_rgb = tuple(map(float, args.mean.split(',')))
std_rgb = tuple(map(float, args.std.split(',')))
augmenter = image.CreateAugmenter(data_shape,
resize=args.resize,
mean=np.array(mean_rgb),
std=np.array(std_rgb))
extracting_features()
train_pic, train_label, test_pic, test_label = load_cifar(data_dir)
# ------------------------------------ #
train_pic = np.transpose(train_pic.astype('float32') / 255, (0, 2, 3, 1))
test_pic = np.transpose(test_pic.astype('float32') / 255, (0, 2, 3, 1))
mean = [0.4914, 0.4822, 0.4465]
std = [0.2470, 0.2435, 0.2616]
for i in range(3):
train_pic[:, :, :, i] = (train_pic[:, :, :, i] - mean[i]) / std[i]
test_pic[:, :, :, i] = (test_pic[:, :, :, i] - mean[i]) / std[i]
train_pic = np.transpose(train_pic, (0, 3, 1, 2))
test_pic = np.transpose(test_pic, (0, 3, 1, 2))
aug_train = image.CreateAugmenter(data_shape=(3, 32, 32),
rand_crop=True,
rand_mirror=True)
# ------------------------------------ #
batch_size = 128
train_data = gluon.data.DataLoader(gluon.data.ArrayDataset(
train_pic, train_label.astype('float32')),
batch_size,
shuffle=True)
test_data = gluon.data.DataLoader(gluon.data.ArrayDataset(
test_pic, test_label.astype('float32')),
batch_size,
shuffle=False)
# ------------------------------------ #
help='hyper param of mix-up')
parser.add_argument('--save-period',
type=int,
default=10,
help='period in epoch of model saving.')
_args = parser.parse_args()
return _args
os.environ[
'MXNET_GLUON_REPO'] = 'https://apache-mxnet.s3.cn-north-1.amazonaws.com.cn/'
os.environ['MXNET_ENABLE_GPU_P2P'] = '0'
train_auglist = image.CreateAugmenter(data_shape=(3, 32, 32),
rand_crop=True,
rand_mirror=True,
mean=nd.array([0.485, 0.456, 0.406]),
std=nd.array([0.229, 0.224, 0.225]))
val_auglist = image.CreateAugmenter(data_shape=(3, 32, 32),
mean=nd.array([0.485, 0.456, 0.406]),
std=nd.array([0.229, 0.224, 0.225]))
args = parse_args()
use_mix_up = True if args.mix_up else False
def transform_train(data, label):
im = data.astype('float32') / 255
im = image.copyMakeBorder(im, 4, 4, 4, 4)
for aug in train_auglist:
im = aug(im)
im = nd.transpose(im, (2, 0, 1))