def get_data(datadir,
size=IMAGE_SIZE,
isTrain=True,
zmin=-1,
zmax=1,
batch=BATCH,
shuffle_read=False):
if isTrain:
augs = [
imgaug.ResizeShortestEdge(int(size * 1.143)),
imgaug.RandomCrop(size),
imgaug.Flip(horiz=True),
]
else:
augs = [
imgaug.ResizeShortestEdge(int(size * 1.143)),
imgaug.CenterCrop(size)
]
def get_images(dir):
files = glob.glob(os.path.join(dir, "*.jpg"))
if shuffle_read:
import random
random.shuffle(files)
else:
files = sorted(files)
image_df = ImageFromFile(files, channel=3, shuffle=isTrain)
image_df = AugmentImageComponent(image_df, augs)
random_df = RandomZData([size, size, 3], zmin, zmax)
return JoinData([random_df, image_df])
names = ['train'] if isTrain else ['test']
df = get_images(*[os.path.join(datadir, n) for n in names])
df = BatchData(df, batch)
return df
def fbresnet_augmentor(isTrain, target_shape):
"""
Augmentor used in fb.resnet.torch, for BGR images in range [0,255].
"""
if isTrain:
augmentors = [
GoogleNetResize(target_shape=target_shape),
imgaug.RandomOrderAug([
imgaug.BrightnessScale((0.6, 1.4), clip=False),
imgaug.Contrast((0.6, 1.4), clip=False),
imgaug.Saturation(0.4, rgb=False),
# rgb-bgr conversion for the constants copied from fb.resnet.torch
imgaug.Lighting(
0.1,
eigval=np.asarray([0.2175, 0.0188, 0.0045][::-1]) * 255.0,
eigvec=np.array([[-0.5675, 0.7192, 0.4009],
[-0.5808, -0.0045, -0.8140],
[-0.5836, -0.6948, 0.4203]],
dtype='float32')[::-1, ::-1])
]),
imgaug.Flip(horiz=True)
]
else:
augmentors = [
imgaug.ResizeShortestEdge(256, cv2.INTER_CUBIC),
imgaug.CenterCrop((224, 224))
]
if target_shape != 224:
augmentors.append(
imgaug.ResizeShortestEdge(target_shape, cv2.INTER_CUBIC))
return augmentors
def _augment(self, img, _):
h, w = img.shape[:2] #获取图像的行数和列数
area = h * w #相当于28*28 要处理的整个区域的大小
for _ in range(10): # 循环十次
targetArea = self.rng.uniform(self.crop_area_fraction,
1.0) * area # 确定目标区域 随机数生成器
aspectR = self.rng.uniform(
self.aspect_ratio_low,
self.aspect_ratio_high) # 随机设置 0.75~1.333的数字
ww = int(np.sqrt(targetArea * aspectR) + 0.5)
hh = int(np.sqrt(targetArea / aspectR) +
0.5) # 上面四步是确定要处理的目标区域的长和宽
if self.rng.uniform() < 0.5:
ww, hh = hh, ww # 如果随机数小于0.5 说明...互换
if hh <= h and ww <= w: # 如果处理的区域小于总长度 正常处理 否则开始下一次循环
# x1，y1为起始点坐标，out是确定的随机区域
x1 = 0 if w == ww else self.rng.randint(0, w - ww)
y1 = 0 if h == hh else self.rng.randint(0, h - hh)
out = img[y1:y1 + hh, x1:x1 + ww]
out = cv2.resize(out, (self.target_shape, self.target_shape),
interpolation=cv2.INTER_CUBIC)
return out
out = imgaug.ResizeShortestEdge(self.target_shape,
interp=cv2.INTER_CUBIC).augment(img)
out = imgaug.CenterCrop(self.target_shape).augment(out)
return out
def fbresnet_augmentor(isTrain):
"""
Augmentor used in fb.resnet.torch, for BGR images in range [0,255].
"""
if isTrain:
augmentors = [
GoogleNetResize(),
imgaug.RandomOrderAug([
JohnAug(),
imgaug.BrightnessScale((0.6, 1.4), clip=False),
imgaug.Contrast((0.6, 1.4), clip=False),
imgaug.Saturation(0.4, rgb=False),
# rgb-bgr conversion for the constants copied from fb.resnet.torch
imgaug.Lighting(
0.1,
eigval=np.asarray([0.2175, 0.0188, 0.0045][::-1]) * 255.0,
eigvec=np.array([[-0.5675, 0.7192, 0.4009],
[-0.5808, -0.0045, -0.8140],
[-0.5836, -0.6948, 0.4203]],
dtype='float32')[::-1, ::-1])
]),
imgaug.Flip(horiz=True),
]
else:
round2pow2 = lambda x: 2**(x - 1).bit_length()
augmentors = [
imgaug.ResizeShortestEdge(round2pow2(IMAGE_SIZE), cv2.INTER_CUBIC),
imgaug.CenterCrop((IMAGE_SIZE, IMAGE_SIZE)),
]
return augmentors
def get_tp_augmentor(isTrain):
"""
Augmentor used in fb.resnet.torch, for BGR images in range [0,255].
"""
interpolation = cv2.INTER_CUBIC
# linear seems to have more stable performance.
# but we keep cubic for compatibility with old models
if isTrain:
augmentors = [
imgaug.GoogleNetRandomCropAndResize(interp=interpolation),
# It's OK to remove the following augs if your CPU is not fast enough.
# Removing brightness/contrast/saturation does not have a significant effect on accuracy.
# Removing lighting leads to a tiny drop in accuracy.
# imgaug.RandomOrderAug(
# [imgaug.BrightnessScale((0.6, 1.4), clip=False),
# imgaug.Contrast((0.6, 1.4), rgb=False, clip=False),
# imgaug.Saturation(0.4, rgb=False),
# # rgb-bgr conversion for the constants copied from fb.resnet.torch
# imgaug.Lighting(0.1,
# eigval=np.asarray(
# [0.2175, 0.0188, 0.0045][::-1]) * 255.0,
# eigvec=np.array(
# [[-0.5675, 0.7192, 0.4009],
# [-0.5808, -0.0045, -0.8140],
# [-0.5836, -0.6948, 0.4203]],
# dtype='float32')[::-1, ::-1]
# )]),
imgaug.Flip(horiz=True),
]
else:
augmentors = [
imgaug.ResizeShortestEdge(256, interp=interpolation),
imgaug.CenterCrop((224, 224)),
]
return augmentors
def normal_augmentor(isTrain):
"""
Normal augmentor with random crop and flip only, for BGR images in range [0,255].
"""
if isTrain:
augmentors = [
imgaug.ResizeShortestEdge(256, cv2.INTER_CUBIC),
imgaug.RandomCrop((DEFAULT_IMAGE_SHAPE, DEFAULT_IMAGE_SHAPE)),
imgaug.Flip(horiz=True),
]
else:
augmentors = [
imgaug.ResizeShortestEdge(256, cv2.INTER_CUBIC),
imgaug.CenterCrop((DEFAULT_IMAGE_SHAPE, DEFAULT_IMAGE_SHAPE)),
]
return augmentors
def fbresnet_augmentor(isTrain):
"""
Augmentor used in fb.resnet.torch, for BGR images in range [0,255].
"""
if isTrain:
augmentors = [
GoogleNetResize(),
# It's OK to remove the following augs if your CPU is not fast enough.
# Removing brightness/contrast/saturation does not have a significant effect on accuracy.
# Removing lighting leads to a tiny drop in accuracy.
imgaug.RandomOrderAug(
[imgaug.BrightnessScale((0.6, 1.4), clip=False),
imgaug.Contrast((0.6, 1.4), clip=False),
imgaug.Saturation(0.4, rgb=False),
# rgb-bgr conversion for the constants copied from fb.resnet.torch
imgaug.Lighting(0.1,
eigval=np.asarray(
[0.2175, 0.0188, 0.0045][::-1]) * 255.0,
eigvec=np.array(
[[-0.5675, 0.7192, 0.4009],
[-0.5808, -0.0045, -0.8140],
[-0.5836, -0.6948, 0.4203]],
dtype='float32')[::-1, ::-1]
)]),
imgaug.Flip(horiz=True),
]
else:
augmentors = [
imgaug.ResizeShortestEdge(256, cv2.INTER_CUBIC),
imgaug.CenterCrop((224, 224)),
]
return augmentors
def get_augmentations(is_train):
if is_train:
augmentors = [
GoogleNetResize(crop_area_fraction=0.76,
target_shape=224), # TODO : 76% or 49%?
imgaug.RandomOrderAug([
imgaug.BrightnessScale((0.6, 1.4), clip=True),
imgaug.Contrast((0.6, 1.4), clip=True),
imgaug.Saturation(0.4, rgb=False),
# rgb-bgr conversion for the constants copied from fb.resnet.torch
imgaug.Lighting(
0.1,
eigval=np.asarray([0.2175, 0.0188, 0.0045][::-1]) * 255.0,
eigvec=np.array([[-0.5675, 0.7192, 0.4009],
[-0.5808, -0.0045, -0.8140],
[-0.5836, -0.6948, 0.4203]],
dtype='float32')[::-1, ::-1])
]),
imgaug.Flip(horiz=True),
]
else:
augmentors = [
imgaug.ResizeShortestEdge(256, cv2.INTER_CUBIC),
imgaug.CenterCrop((224, 224)),
]
return augmentors
def fbresnet_augmentor(isTrain):
"""
Augmentor used in fb.resnet.torch, for BGR images in range [0,255].
"""
if isTrain:
"""
Sec 5.1:
We use scale and aspect ratio data augmentation [35] as
in [12]. The network input image is a 224×224 pixel random
crop from an augmented image or its horizontal flip.
"""
augmentors = [
GoogleNetResize(),
imgaug.RandomOrderAug([
imgaug.BrightnessScale((0.6, 1.4), clip=False),
imgaug.Contrast((0.6, 1.4), clip=False),
imgaug.Saturation(0.4, rgb=False),
# rgb-bgr conversion for the constants copied from fb.resnet.torch
imgaug.Lighting(
0.1,
eigval=np.asarray([0.2175, 0.0188, 0.0045][::-1]) * 255.0,
eigvec=np.array([[-0.5675, 0.7192, 0.4009],
[-0.5808, -0.0045, -0.8140],
[-0.5836, -0.6948, 0.4203]],
dtype='float32')[::-1, ::-1])
]),
imgaug.Flip(horiz=True),
]
else:
augmentors = [
imgaug.ResizeShortestEdge(256, cv2.INTER_CUBIC),
imgaug.CenterCrop((224, 224)),
]
return augmentors
def fbresnet_augmentor(isTrain):
"""
Augmentor used in fb.resnet.torch, for BGR images in range [0,255].
"""
if isTrain:
augmentors = [
GoogleNetResize(),
imgaug.RandomOrderAug([
imgaug.ColorSpace(mode=cv2.COLOR_BGR2RGB),
imgaug.BrightnessScale((0.6, 1.4), clip=False),
imgaug.Contrast((0.6, 1.4), clip=False),
imgaug.Saturation(0.4, rgb=False),
# rgb-bgr conversion for the constants copied from fb.resnet.torch
# imgaug.Lighting(0.1,
# eigval=np.asarray(
# [0.2175, 0.0188, 0.0045][::-1]) * 255.0,
# eigvec=np.array(
# [[-0.5675, 0.7192, 0.4009],
# [-0.5808, -0.0045, -0.8140],
# [-0.5836, -0.6948, 0.4203]],
# dtype='float32')[::-1, ::-1]
# )
]),
imgaug.Flip(horiz=True),
]
else:
augmentors = [
imgaug.ColorSpace(mode=cv2.COLOR_BGR2RGB),
imgaug.ResizeShortestEdge(256, cv2.INTER_CUBIC),
imgaug.CenterCrop((224, 224)),
]
return augmentors
def fbresnet_augmentor(isTrain):
"""
Augmentor used in fb.resnet.torch, for BGR images in range [0,255]. # 残差网络增强图像
"""
if isTrain: # 如果训练数据的话
augmentors = [
GoogleNetResize(), # 定义好了crop_area_fraction等参数
imgaug.
RandomOrderAug( # GPU不行的话就把这部分删除Remove these augs if your CPU is not fast enough #imgaug是一个图像增强库
[
imgaug.BrightnessScale((0.6, 1.4), clip=False),
imgaug.Contrast((0.6, 1.4), clip=False),
imgaug.Saturation(0.4, rgb=False),
# rgb-bgr conversion for the constants copied from fb.resnet.torch
imgaug.Lighting(
0.1,
eigval=np.asarray([0.2175, 0.0188, 0.0045][::-1]) *
255.0,
eigvec=np.array([[-0.5675, 0.7192, 0.4009],
[-0.5808, -0.0045, -0.8140],
[-0.5836, -0.6948, 0.4203]],
dtype='float32')[::-1, ::-1])
]),
imgaug.Flip(horiz=True),
]
else: # 如果不是训练数据的话
augmentors = [
imgaug.ResizeShortestEdge(
256, cv2.INTER_CUBIC), # 在保持纵横比的同时，将最短边的大小调整为某个数字。
imgaug.CenterCrop((224, 224)), # 在中间裁剪图像
]
return augmentors
def fbresnet_augmentor(isTrain):
"""
Augmentor used in fb.resnet.torch, for BGR images.
"""
if isTrain:
augmentors = [
GoogleNetResize(),
imgaug.RandomOrderAug([
imgaug.Brightness(30, clip=False),
imgaug.Contrast((0.8, 1.2), clip=False),
imgaug.Saturation(0.4, rgb=False),
# rgb-bgr conversion
imgaug.Lighting(0.1,
eigval=[0.2175, 0.0188, 0.0045][::-1],
eigvec=np.array([[-0.5675, 0.7192, 0.4009],
[-0.5808, -0.0045, -0.8140],
[-0.5836, -0.6948, 0.4203]],
dtype='float32')[::-1, ::-1])
]),
imgaug.Clip(),
imgaug.Flip(horiz=True),
]
else:
augmentors = [
imgaug.ResizeShortestEdge(256, cv2.INTER_CUBIC),
imgaug.CenterCrop((224, 224)),
]
return augmentors
def get_ilsvrc_data_alexnet(is_train, image_size, batchsize, directory):
if is_train:
if not directory.startswith('/'):
ds = ILSVRCTTenthTrain(directory)
else:
ds = ILSVRC12(directory, 'train')
augs = [
imgaug.RandomApplyAug(imgaug.RandomResize((0.9, 1.2), (0.9, 1.2)),
0.7),
imgaug.RandomApplyAug(imgaug.RotationAndCropValid(15), 0.7),
imgaug.RandomApplyAug(
imgaug.RandomChooseAug([
imgaug.SaltPepperNoise(white_prob=0.01, black_prob=0.01),
imgaug.RandomOrderAug([
imgaug.BrightnessScale((0.8, 1.2), clip=False),
imgaug.Contrast((0.8, 1.2), clip=False),
# imgaug.Saturation(0.4, rgb=True),
]),
]),
0.7),
imgaug.Flip(horiz=True),
imgaug.ResizeShortestEdge(256, cv2.INTER_CUBIC),
imgaug.RandomCrop((224, 224)),
]
ds = AugmentImageComponent(ds, augs)
ds = PrefetchData(ds, 1000, multiprocessing.cpu_count())
ds = BatchData(ds, batchsize)
ds = PrefetchData(ds, 10, 4)
else:
if not directory.startswith('/'):
ds = ILSVRCTenthValid(directory)
else:
ds = ILSVRC12(directory, 'val')
ds = AugmentImageComponent(ds, [
imgaug.ResizeShortestEdge(224, cv2.INTER_CUBIC),
imgaug.CenterCrop((224, 224)),
])
ds = PrefetchData(ds, 100, multiprocessing.cpu_count())
ds = BatchData(ds, batchsize)
return ds
def get_data(datadir, size=IMAGESIZE, isTrain=True, zmin=-1, zmax=1):
if isTrain:
augs = [
imgaug.ResizeShortestEdge(int(size * 1.143)),
imgaug.RandomCrop(size),
imgaug.Flip(horiz=True),
]
else:
augs = [
imgaug.ResizeShortestEdge(int(size * 1.143)),
imgaug.CenterCrop(size)
]
def get_images(dir):
files = sorted(glob.glob(os.path.join(dir, "*.jpg")))
df = ImageFromFile(files, channel=3, shuffle=isTrain)
random_df = RandomZData([size, size, 3], zmin, zmax)
return JoinData([random_df, AugmentImageComponent(df, augs)])
names = ['train'] if isTrain else ['test']
df = get_images(*[os.path.join(datadir, n) for n in names])
df = BatchData(df, BATCH if isTrain else TEST_BATCH)
return df
def fbresnet_augmentor_fast(isTrain):
"""
Augmentor used in fb.resnet.torch, for BGR images in range [0,255].
"""
if isTrain:
augmentors = [
GoogleNetResize(),
]
else:
augmentors = [
imgaug.ResizeShortestEdge(256, cv2.INTER_CUBIC),
imgaug.CenterCrop((224, 224)),
]
return augmentors
def _augment(self, img, _):
h, w = img.shape[:2]
area = h * w
for _ in range(10):
targetArea = self.rng.uniform(self.crop_area_fraction, 1.0) * area
aspectR = self.rng.uniform(self.aspect_ratio_low, self.aspect_ratio_high)
ww = int(np.sqrt(targetArea * aspectR) + 0.5)
hh = int(np.sqrt(targetArea / aspectR) + 0.5)
if self.rng.uniform() < 0.5:
ww, hh = hh, ww
if hh <= h and ww <= w:
x1 = 0 if w == ww else self.rng.randint(0, w - ww)
y1 = 0 if h == hh else self.rng.randint(0, h - hh)
out = img[y1:y1 + hh, x1:x1 + ww]
out = cv2.resize(out, (self.target_shape, self.target_shape), interpolation=cv2.INTER_CUBIC)
return out
out = imgaug.ResizeShortestEdge(self.target_shape, interp=cv2.INTER_CUBIC).augment(img)
out = imgaug.CenterCrop(self.target_shape).augment(out)
return out
def _augment(self, img, _):
h, w = img.shape[:2]
area = h * w
for _ in range(10):
targetArea = self.rng.uniform(0.08, 1.0) * area
aspectR = self.rng.uniform(0.75, 1.333)
ww = int(np.sqrt(targetArea * aspectR))
hh = int(np.sqrt(targetArea / aspectR))
if self.rng.uniform() < 0.5:
ww, hh = hh, ww
if hh <= h and ww <= w:
x1 = 0 if w == ww else self.rng.randint(0, w - ww)
y1 = 0 if h == hh else self.rng.randint(0, h - hh)
out = img[y1:y1 + hh, x1:x1 + ww]
out = cv2.resize(out, (224, 224), interpolation=cv2.INTER_CUBIC)
return out
out = imgaug.ResizeShortestEdge(224, interp=cv2.INTER_CUBIC).augment(img)
out = imgaug.CenterCrop(224).augment(out)
return out
def fbresnet_augmentor(isTrain):
"""
Augmentor used in fb.resnet.torch, for BGR images in range [0,255].
"""
interpolation = cv2.INTER_LINEAR
if isTrain:
"""
Sec 5.1:
We use scale and aspect ratio data augmentation [35] as
in [12]. The network input image is a 224×224 pixel random
crop from an augmented image or its horizontal flip.
"""
augmentors = [
imgaug.GoogleNetRandomCropAndResize(interp=interpolation),
# It's OK to remove the following augs if your CPU is not fast enough.
# Removing brightness/contrast/saturation does not have a significant effect on accuracy.
# Removing lighting leads to a tiny drop in accuracy.
imgaug.RandomOrderAug(
[imgaug.BrightnessScale((0.6, 1.4), clip=False),
imgaug.Contrast((0.6, 1.4), rgb=False, clip=False),
imgaug.Saturation(0.4, rgb=False),
# rgb-bgr conversion for the constants copied from fb.resnet.torch
imgaug.Lighting(0.1,
eigval=np.asarray(
[0.2175, 0.0188, 0.0045][::-1]) * 255.0,
eigvec=np.array(
[[-0.5675, 0.7192, 0.4009],
[-0.5808, -0.0045, -0.8140],
[-0.5836, -0.6948, 0.4203]],
dtype='float32')[::-1, ::-1]
)]),
imgaug.Flip(horiz=True),
]
else:
augmentors = [
imgaug.ResizeShortestEdge(256, interp=interpolation),
imgaug.CenterCrop((224, 224)),
]
return augmentors
def prediction_incorrect(logits,
label,
topk=1,
name='incorrect_vector'):
with tf.name_scope('prediction_incorrect'):
x = tf.logical_not(tf.nn.in_top_k(logits, label, topk))
return tf.cast(x, tf.float32, name=name)
wrong = prediction_incorrect(logits, label, 1, name='wrong-top1')
add_moving_summary(tf.reduce_mean(wrong, name='train-error-top1'))
wrong = prediction_incorrect(logits, label, 5, name='wrong-top5')
add_moving_summary(tf.reduce_mean(wrong, name='train-error-top5'))
return loss
if __name__ == '__main__':
import argparse
from tensorpack.dataflow import TestDataSpeed
parser = argparse.ArgumentParser()
parser.add_argument('--data', required=True)
parser.add_argument('--batch', type=int, default=32)
args = parser.parse_args()
augs = fbresnet_augmentor(False)
augs = [imgaug.ResizeShortestEdge(256), imgaug.CenterCrop(224)]
df = get_imagenet_dataflow(args.data, 'train', args.batch, augs)
TestDataSpeed(df).start()
def fbresnet_augmentor(isTrain, crop_method, color_augmentation):
"""
Augmentor used in fb.resnet.torch, for BGR images in range [0,255].
"""
execution_lst = []
if isTrain:
augmentors = [
# 1. crop_method
# a) GoogleNetResize
GoogleNetResize(),
# b) ShortestEdgeResize
imgaug.ResizeShortestEdge(256),
# c) GlobalWarp
imgaug.Resize(226), # NOTE: for CAM generation
imgaug.RandomCrop((224, 224)),
# d) CAMCrop
# (when CAMCrop is set, the output from the original DataFlow has already been cropped)
# 2. color_augmentation
imgaug.RandomOrderAug([
imgaug.BrightnessScale((0.6, 1.4), clip=False),
imgaug.Contrast((0.6, 1.4), clip=False),
imgaug.Saturation(0.4, rgb=False),
# rgb-bgr conversion for the constants copied from fb.resnet.torch
imgaug.Lighting(
0.1,
eigval=np.asarray([0.2175, 0.0188, 0.0045][::-1]) * 255.0,
eigvec=np.array([[-0.5675, 0.7192, 0.4009],
[-0.5808, -0.0045, -0.8140],
[-0.5836, -0.6948, 0.4203]],
dtype='float32')[::-1, ::-1])
]),
imgaug.Flip(horiz=True),
]
#
if crop_method == 'GoogleNetResize':
print(
'--> perform GoogleNetResize cropping method during the training pipeline'
)
execution_lst.extend([0])
elif crop_method == 'ShortestEdgeResize':
print(
'--> perform ShortestEdgeResize cropping method during the training pipeline'
)
execution_lst.extend([1, 3])
elif crop_method == 'GlobalWarp':
print(
'--> perform GlobalWarp cropping method during the training pipeline'
)
execution_lst.extend([2, 3])
elif crop_method == 'CAMCrop':
# enable CAMCrop @ 20171124
print(
'*** Perform CAMCrop to better the training dynamics and the results ***'
)
if color_augmentation:
print(
'--> perform color augmentation during the training pipeline')
execution_lst.extend([4])
else:
print(
'--> discard the color jittering process during the training pipeline'
)
# perform mirror reflection augmentation anyway
execution_lst.extend([5])
else:
augmentors = [
imgaug.ResizeShortestEdge(256, cv2.INTER_CUBIC),
imgaug.CenterCrop((224, 224)),
imgaug.RandomCrop((224, 224)),
]
if crop_method == 'RandomCrop':
execution_lst.extend([0, 2])
elif crop_method == 'CenterCrop':
execution_lst.extend([0, 1])
return [
item_ for id_, item_ in enumerate(augmentors) if id_ in execution_lst
]