def get_data(name, batch):
isTrain = name == 'train'
image_shape = 224
if isTrain:
augmentors = [
# use lighter augs if model is too small
GoogleNetResize(
crop_area_fraction=0.49 if args.width_ratio < 1 else 0.08,
target_shape=image_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),
]),
imgaug.Flip(horiz=True),
]
else:
augmentors = [
imgaug.ResizeShortestEdge(int(image_shape * 256 / 224),
cv2.INTER_CUBIC),
imgaug.CenterCrop((image_shape, image_shape)),
]
return get_imagenet_dataflow(args.data_dir,
name,
batch,
augmentors,
meta_dir=args.meta_dir)
def get_data(train_or_test):
isTrain = train_or_test == 'train'
pp_mean = dataset.SVHNDigit.get_per_pixel_mean()
if isTrain:
d1 = dataset.SVHNDigit('train')
d2 = dataset.SVHNDigit('extra')
ds = RandomMixData([d1, d2])
else:
ds = dataset.SVHNDigit('test')
if isTrain:
augmentors = [
imgaug.CenterPaste((40, 40)),
imgaug.RandomCrop((32, 32)),
#imgaug.Flip(horiz=True),
imgaug.BrightnessAdd(10),
imgaug.Contrast((0.8, 1.2)),
imgaug.GaussianDeform( # this is slow
[(0.2, 0.2), (0.2, 0.8), (0.8, 0.8), (0.8, 0.2)], (32, 32),
0.2, 3),
imgaug.MapImage(lambda x: x - pp_mean),
]
else:
augmentors = [imgaug.MapImage(lambda x: x - pp_mean)]
ds = AugmentImageComponent(ds, augmentors)
ds = BatchData(ds, 128, remainder=not isTrain)
if isTrain:
ds = PrefetchData(ds, 5, 5)
return ds
def get_data(name, batch):
isTrain = name == 'train'
if isTrain:
augmentors = [
GoogleNetResize(crop_area_fraction=0.49),
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 get_imagenet_dataflow(args.data, name, batch, augmentors)
def fbresnet_augmentor(isTrain):
"""
Augmentor used in fb.resnet.torch, for BGR images in range [0,255].
"""
if isTrain:
augmentors = [
GoogleNetResize(),
imgaug.Flip(horiz=True),
imgaug.ToFloat32(),
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])
]),
]
else:
augmentors = [
imgaug.ResizeShortestEdge(256, cv2.INTER_LINEAR),
imgaug.CenterCrop((224, 224)),
imgaug.ToFloat32(),
]
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_config():
#anchors = np.mgrid[0:4,0:4][:,1:,1:].transpose(1,2,0).reshape((-1,2)) / 4.0
# prepare dataset
d1 = dataset.SVHNDigit('train')
d2 = dataset.SVHNDigit('extra')
train = RandomMixData([d1, d2])
test = dataset.SVHNDigit('test')
augmentors = [
imgaug.Resize((40, 40)),
imgaug.BrightnessAdd(30),
imgaug.Contrast((0.5, 1.5)),
imgaug.GaussianDeform( # this is slow
[(0.2, 0.2), (0.2, 0.8), (0.8, 0.8), (0.8, 0.2)], (40, 40), 0.2,
3),
]
train = AugmentImageComponent(train, augmentors)
train = BatchData(train, 128)
nr_proc = 5
train = PrefetchData(train, 5, nr_proc)
step_per_epoch = train.size()
augmentors = [
imgaug.Resize((40, 40)),
]
test = AugmentImageComponent(test, augmentors)
test = BatchData(test, 128, remainder=True)
sess_config = get_default_sess_config(0.8)
lr = tf.train.exponential_decay(learning_rate=1e-3,
global_step=get_global_step_var(),
decay_steps=train.size() * 60,
decay_rate=0.2,
staircase=True,
name='learning_rate')
tf.scalar_summary('learning_rate', lr)
return TrainConfig(
dataset=train,
optimizer=tf.train.AdamOptimizer(lr),
callbacks=Callbacks([
StatPrinter(),
ModelSaver(),
InferenceRunner(dataset_test,
[ScalarStats('cost'),
ClassificationError()])
]),
session_config=sess_config,
model=Model(),
step_per_epoch=step_per_epoch,
max_epoch=350,
)
def get_moco_v1_augmentor():
augmentors = [
TorchvisionCropAndResize(crop_area_fraction=(0.2, 1.)),
RandomGrayScale(0.2),
imgaug.ToFloat32(),
imgaug.RandomOrderAug([
imgaug.BrightnessScale((0.6, 1.4)),
imgaug.Contrast((0.6, 1.4), rgb=False),
imgaug.Saturation(0.4, rgb=False),
# 72 = 180*0.4
imgaug.Hue(range=(-72, 72), rgb=False)
]),
imgaug.ToUint8(),
imgaug.Flip(horiz=True),
]
return augmentors
def get_moco_v1_augmentor():
augmentors = [
imgaug.GoogleNetRandomCropAndResize(crop_area_fraction=(0.2, 1.)),
imgaug.RandomApplyAug(imgaug.Grayscale(rgb=False, keepshape=True), 0.2),
imgaug.ToFloat32(),
imgaug.RandomOrderAug(
[imgaug.BrightnessScale((0.6, 1.4)),
imgaug.Contrast((0.6, 1.4), rgb=False),
imgaug.Saturation(0.4, rgb=False),
# 72 = 180*0.4
imgaug.Hue(range=(-72, 72), rgb=False)
]),
imgaug.ToUint8(),
imgaug.Flip(horiz=True),
]
return augmentors
def get_resnet_augmentor():
augmentors = [
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]
)])
]
return augmentors
def get_moco_v2_augmentor():
augmentors = [
TorchvisionCropAndResize(crop_area_fraction=(0.2, 1.)),
imgaug.ToFloat32(),
imgaug.RandomApplyAug(
imgaug.RandomOrderAug([
imgaug.BrightnessScale((0.6, 1.4)),
imgaug.Contrast((0.6, 1.4), rgb=False),
imgaug.Saturation(0.4, rgb=False),
# 18 = 180*0.1
imgaug.Hue(range=(-18, 18), rgb=False)
]),
0.8),
RandomGrayScale(0.2),
imgaug.RandomApplyAug(RandomGaussionBlur([0.1, 2.0], 0.1), 0.5),
imgaug.ToUint8(),
imgaug.Flip(horiz=True),
]
return augmentors
def get_moco_v2_augmentor():
augmentors = [
imgaug.GoogleNetRandomCropAndResize(crop_area_fraction=(0.2, 1.)),
imgaug.ToFloat32(),
imgaug.RandomApplyAug(
imgaug.RandomOrderAug(
[imgaug.BrightnessScale((0.6, 1.4)),
imgaug.Contrast((0.6, 1.4), rgb=False),
imgaug.Saturation(0.4, rgb=False),
# 18 = 180*0.1
imgaug.Hue(range=(-18, 18), rgb=False)
]), 0.8),
imgaug.RandomApplyAug(imgaug.Grayscale(rgb=False, keepshape=True), 0.2),
imgaug.RandomApplyAug(
# 11 = 0.1*224//2
imgaug.GaussianBlur(size_range=(11, 12), sigma_range=[0.1, 2.0]), 0.5),
imgaug.ToUint8(),
imgaug.Flip(horiz=True),
]
return augmentors
def fbresnet_augmentor():
# assme BGR input
augmentors = [
imgaug.GoogleNetRandomCropAndResize(),
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),
]
return augmentors
def get_data(is_train,
batch_size,
data_dir_path,
input_image_size=224,
resize_inv_factor=0.875):
assert (resize_inv_factor > 0.0)
resize_value = int(math.ceil(float(input_image_size) / resize_inv_factor))
if is_train:
augmentors = [
GoogleNetResize(crop_area_fraction=0.08,
target_shape=input_image_size),
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(resize_value, cv2.INTER_CUBIC),
imgaug.ResizeShortestEdge(resize_value, cv2.INTER_LINEAR),
imgaug.CenterCrop((input_image_size, input_image_size))
]
return get_imagenet_dataflow(datadir=data_dir_path,
is_train=is_train,
batch_size=batch_size,
augmentors=augmentors)
def fbresnet_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),
imgaug.ToFloat32(), # avoid frequent casting in each color augmentation
# 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)),
imgaug.Contrast((0.6, 1.4), rgb=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.ToUint8(),
imgaug.Flip(horiz=True),
]
else:
augmentors = [
imgaug.ResizeShortestEdge(256, interp=interpolation),
imgaug.CenterCrop((224, 224)),
]
return augmentors
def get_inat_augmented_data(subset,
options,
lmdb_dir=None,
year='2018',
do_multiprocess=True,
do_validation=False,
is_train=None,
shuffle=None,
n_allow=None):
input_size = options.input_size if options.input_size else 224
isTrain = is_train if is_train is not None else (subset == 'train'
and do_multiprocess)
shuffle = shuffle if shuffle is not None else isTrain
postfix = "" if n_allow is None else "_allow_{}".format(n_allow)
#TODO: Parameterize the cv split to be consider
#Currently hardcoding to 1
cv = 1
# When do_validation is True it will expect *cv_train and *cv_val lmdbs
# Currently the cv_train split is always used
if isTrain:
postfix += '_cv_train_{}'.format(cv)
elif do_validation:
subset = 'train'
postfix += '_cv_val_{}'.format(cv)
if lmdb_dir == None:
lmdb_path = os.path.join(options.data_dir, 'inat_lmdb',
'inat2018_{}{}.lmdb'.format(subset, postfix))
else:
lmdb_path = os.path.join(
options.data_dir, lmdb_dir,
'inat{}_{}{}.lmdb'.format(year, subset, postfix))
ds = LMDBData(lmdb_path, shuffle=False)
if shuffle:
ds = LocallyShuffleData(ds,
1024 * 80) # This is 64G~80G in memory images
ds = PrefetchData(ds, 1024 * 8, 1) # prefetch around 8 G
ds = LMDBDataPoint(ds)
ds = MapDataComponent(ds, lambda x: cv2.imdecode(x, cv2.IMREAD_COLOR),
0) # BGR uint8 data
if isTrain:
class Resize(imgaug.ImageAugmentor):
"""
crop 8%~100% of the original image
See `Going Deeper with Convolutions` by Google.
"""
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, (input_size, input_size),
interpolation=cv2.INTER_CUBIC)
return out
out = cv2.resize(img, (input_size, input_size),
interpolation=cv2.INTER_CUBIC)
return out
augmentors = [
Resize(),
imgaug.RandomOrderAug([
imgaug.Brightness(30, clip=False),
imgaug.Contrast((0.8, 1.2), clip=False),
imgaug.Saturation(0.4),
# 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),
imgaug.ToUint8()
]
else:
augmentors = [
imgaug.ResizeShortestEdge(256),
imgaug.CenterCrop((input_size, input_size)),
imgaug.ToUint8()
]
ds = AugmentImageComponent(ds, augmentors, copy=False)
if do_multiprocess:
ds = PrefetchDataZMQ(ds, min(24, multiprocessing.cpu_count()))
ds = BatchData(ds,
options.batch_size // options.nr_gpu,
remainder=not isTrain)
return ds
def get_tiny_imagenet_augmented_data(subset, options,
do_multiprocess=True, is_train=None, shuffle=None):
isTrain = is_train if is_train is not None else (subset == 'train' and do_multiprocess)
shuffle = shuffle if shuffle is not None else isTrain
lmdb_path = os.path.join(options.data_dir,
'tiny_imagenet_lmdb', 'tiny_imagenet_{}.lmdb'.format(subset))
# since tiny imagenet is small (200MB zipped) we can shuffle all directly.
# we skipped the LocallyShuffleData and PrefetchData routine.
ds = LMDBData(lmdb_path, shuffle=shuffle)
ds = LMDBDataPoint(ds)
ds = MapDataComponent(ds, lambda x: cv2.imdecode(x, cv2.IMREAD_COLOR), 0)
img_size = 64
if isTrain:
class Resize(imgaug.ImageAugmentor):
"""
crop 8%~100% of the original image
See `Going Deeper with Convolutions` by Google.
"""
def _augment(self, img, _):
h, w = img.shape[:2]
area = h * w
for _ in range(10):
targetArea = self.rng.uniform(0.3, 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, (img_size, img_size), interpolation=cv2.INTER_CUBIC)
return out
out = cv2.resize(img, (img_size, img_size), interpolation=cv2.INTER_CUBIC)
return out
augmentors = [
Resize(),
imgaug.RandomOrderAug(
[imgaug.Brightness(30, clip=False),
imgaug.Contrast((0.8, 1.2), clip=False),
imgaug.Saturation(0.4),
# 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),
imgaug.ToUint8()
]
else:
augmentors = [
imgaug.ResizeShortestEdge(72),
imgaug.CenterCrop((img_size, img_size)),
imgaug.ToUint8()
]
ds = AugmentImageComponent(ds, augmentors, copy=False)
ds = BatchData(ds, options.batch_size // options.nr_gpu, remainder=not isTrain)
if do_multiprocess:
ds = PrefetchData(ds, nr_prefetch=4, nr_proc=4)
return ds
def get_input_imagenet():
train = dataset.ILSVRC12("/datasets/ImageNet/ILSVRC/Data/CLS-LOC",
"train",
dir_structure="train",
shuffle=True)
test = dataset.ILSVRC12("/datasets/ImageNet/ILSVRC/Data/CLS-LOC",
"val",
dir_structure="train",
shuffle=False)
# Copied from tensorpack examples:
# https://github.com/tensorpack/tensorpack/blob/master/examples/ImageNetModels/imagenet_utils.py
train_augmentors = imgaug.AugmentorList([
imgaug.GoogleNetRandomCropAndResize(interp=cv2.INTER_CUBIC),
# 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),
])
test_augmentors = imgaug.AugmentorList([
imgaug.ResizeShortestEdge(256, interp=cv2.INTER_CUBIC),
imgaug.CenterCrop((224, 224)),
])
def preprocess(augmentors):
def apply(x):
image, label = x
onehot = np.zeros(1000)
onehot[label] = 1.0
image = augmentors.augment(image)
return image, onehot
return apply
parallel = min(40,
multiprocessing.cpu_count() // 2) # assuming hyperthreading
train = MapData(train, preprocess(train_augmentors))
train = PrefetchDataZMQ(train, parallel)
test = MultiThreadMapData(test,
parallel,
preprocess(test_augmentors),
strict=True)
test = PrefetchDataZMQ(test, 1)
return train, test, ((224, 224, 3), (1000, ))