def get_data(name, batch):
isTrain = name == 'train'
if isTrain:
augmentors = [
# use lighter augs if model is too small
imgaug.GoogleNetRandomCropAndResize(crop_area_fraction=(0.49 if args.ratio < 1 else 0.08, 1.)),
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 = [
imgaug.GoogleNetRandomCropAndResize(),
# 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 fbresnet_augmentor(isTrain):
"""
Augmentor used in fb.resnet.torch, for BGR images in range [0,255].
"""
if isTrain:
augmentors = [
imgaug.GoogleNetRandomCropAndResize(),
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 train_dataloader(self):
ds_train = MultiLabelDataset(folder=self.hparams.data,
is_train='train',
fname='covid_train_v5.csv',
types=self.hparams.types,
pathology=self.hparams.pathology,
resize=int(self.hparams.shape),
balancing=None)
ds_train.reset_state()
ag_train = [
# imgaug.Albumentations(
# AB.SmallestMaxSize(self.hparams.shape, p=1.0)),
imgaug.ColorSpace(mode=cv2.COLOR_GRAY2RGB),
# imgaug.Affine(shear=10),
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.Albumentations(AB.CLAHE(tile_grid_size=(32, 32), p=0.5)),
imgaug.RandomOrderAug([
imgaug.Affine(shear=10,
border=cv2.BORDER_CONSTANT,
interp=cv2.INTER_AREA),
imgaug.Affine(translate_frac=(0.01, 0.02),
border=cv2.BORDER_CONSTANT,
interp=cv2.INTER_AREA),
imgaug.Affine(scale=(0.5, 1.0),
border=cv2.BORDER_CONSTANT,
interp=cv2.INTER_AREA),
]),
imgaug.RotationAndCropValid(max_deg=10, interp=cv2.INTER_AREA),
imgaug.GoogleNetRandomCropAndResize(
crop_area_fraction=(0.8, 1.0),
aspect_ratio_range=(0.8, 1.2),
interp=cv2.INTER_AREA,
target_shape=self.hparams.shape),
imgaug.ColorSpace(mode=cv2.COLOR_RGB2GRAY),
imgaug.ToFloat32(),
]
ds_train = AugmentImageComponent(ds_train, ag_train, 0)
# Label smoothing
ag_label = [
imgaug.BrightnessScale((0.8, 1.2), clip=False),
]
# ds_train = AugmentImageComponent(ds_train, ag_label, 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(np.transpose(dp[0], (0, 3, 1, 2))),
torch.tensor(dp[1]).float()
])
return ds_train
def get_basic_augmentor(isTrain):
interpolation = cv2.INTER_LINEAR
if isTrain:
augmentors = [
imgaug.GoogleNetRandomCropAndResize(),
imgaug.Flip(horiz=True),
]
else:
augmentors = [
imgaug.ResizeShortestEdge(256, interp=interpolation),
imgaug.CenterCrop((224, 224)),
]
return augmentors
def resizeOnly_augmentor():
# assme BGR input
augmentors = [
imgaug.GoogleNetRandomCropAndResize(),
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_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 = [
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 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 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_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, ))
def get_input_visualwakewords():
# VISUALWAKEWORDS_CONFIG = {
# "train_instances": "/home/el/Datasets/COCO14/visualwakewords/instances_visualwakewords_train2014.json",
# "train_images": "/home/el/Datasets/COCO14/train2014",
# "val_instances": "/home/el/Datasets/COCO14/visualwakewords/instances_visualwakewords_val2014.json",
# "val_images": "/home/el/Datasets/COCO14/val2014",
# "filter_list": "/home/el/Datasets/COCO14/visualwakewords/mscoco_minival_ids.txt"
# }
target_side_size = 128
train = VisualWakeWords(VISUALWAKEWORDS_CONFIG["train_instances"],
VISUALWAKEWORDS_CONFIG["train_images"],
shuffle=True)
test = VisualWakeWords(VISUALWAKEWORDS_CONFIG["val_instances"],
VISUALWAKEWORDS_CONFIG["val_images"],
filter_list=VISUALWAKEWORDS_CONFIG["filter_list"],
shuffle=False)
train_augmentors = imgaug.AugmentorList([
imgaug.GoogleNetRandomCropAndResize(crop_area_fraction=(0.25, 1.),
target_shape=target_side_size,
interp=cv2.INTER_CUBIC),
imgaug.Flip(horiz=True),
])
autoaugment_policy = ImageNetPolicy()
test_augmentors = imgaug.AugmentorList([
imgaug.ResizeShortestEdge(int(target_side_size * 1.2),
interp=cv2.INTER_CUBIC),
imgaug.CenterCrop((target_side_size, target_side_size)),
])
def preprocess(train):
def apply(x):
image, label = x
onehot = np.zeros(2)
onehot[label] = 1.0
augmentors = train_augmentors if train else test_augmentors
image = augmentors.augment(image)
if train:
image = np.array(autoaugment_policy(Image.fromarray(image)))
return image, onehot
# mean = [0.4767, 0.4488, 0.4074]
# std = [0.2363, 0.2313, 0.2330]
# return (image / 255.0 - mean) / std, onehot
return apply
parallel = min(18,
multiprocessing.cpu_count() // 2) # assuming hyperthreading
train = MapData(train, preprocess(train=True))
train = PrefetchDataZMQ(train, parallel)
test = MultiThreadMapData(test,
parallel,
preprocess(train=False),
strict=True)
test = PrefetchDataZMQ(test, 1)
return train, test, ((target_side_size, target_side_size, 3), (2, ))
