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_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(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 train_dataloader(self):
"""Summary
Returns:
TYPE: Description
"""
ds_train = CustomDataSet(folder=self.hparams.data,
train_or_valid='train',
size=np.inf,
hparams=self.hparams)
ds_train.reset_state()
ag_train = [
imgaug.Affine(shear=10, interp=cv2.INTER_NEAREST),
imgaug.Affine(translate_frac=(0.01, 0.02),
interp=cv2.INTER_NEAREST),
imgaug.Affine(scale=(0.25, 1.0), interp=cv2.INTER_NEAREST),
imgaug.RotationAndCropValid(max_deg=10, interp=cv2.INTER_NEAREST),
imgaug.GoogleNetRandomCropAndResize(
crop_area_fraction=(0.8, 1.0),
aspect_ratio_range=(0.8, 1.2),
interp=cv2.INTER_NEAREST,
target_shape=self.hparams.shape),
imgaug.Resize(self.hparams.shape, interp=cv2.INTER_NEAREST),
imgaug.Flip(horiz=True, vert=False, prob=0.5),
imgaug.Flip(horiz=False, vert=True, prob=0.5),
imgaug.Transpose(prob=0.5),
imgaug.Albumentations(AB.RandomRotate90(p=1)),
imgaug.ToFloat32(),
]
ds_train = AugmentImageComponent(
ds_train,
[
# imgaug.Float32(),
# imgaug.RandomChooseAug([
# imgaug.Albumentations(AB.IAAAdditiveGaussianNoise(p=0.25)),
# imgaug.Albumentations(AB.GaussNoise(p=0.25)),
# ]),
# imgaug.ToUint8(),
imgaug.RandomChooseAug([
imgaug.Albumentations(AB.Blur(blur_limit=4, p=0.25)),
imgaug.Albumentations(AB.MotionBlur(blur_limit=4, p=0.25)),
imgaug.Albumentations(AB.MedianBlur(blur_limit=4, p=0.25)),
]),
imgaug.RandomChooseAug([
# imgaug.Albumentations(AB.IAASharpen(p=0.5)),
# imgaug.Albumentations(AB.IAAEmboss(p=0.5)),
imgaug.Albumentations(AB.RandomBrightnessContrast(p=0.5)),
]),
imgaug.ToUint8(),
imgaug.Albumentations(AB.CLAHE(tile_grid_size=(32, 32),
p=0.5)),
],
0)
ds_train = AugmentImageComponents(ds_train, ag_train, [0, 1])
ds_train = BatchData(ds_train, self.hparams.batch, remainder=True)
if self.hparams.debug:
ds_train = FixedSizeData(ds_train, 2)
ds_train = MultiProcessRunner(ds_train, num_proc=4, num_prefetch=16)
ds_train = PrintData(ds_train)
ds_train = MapData(
ds_train, lambda dp: [
torch.tensor(dp[0][:, np.newaxis, :, :]).float(),
torch.tensor(dp[1][:, np.newaxis, :, :]).float(),
])
return ds_train
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
