def test_channel_droput():
img = np.ones((10, 10, 3), dtype=np.float32)
aug = A.ChannelDropout(channel_drop_range=(1, 1), always_apply=True) # Drop one channel
transformed = aug(image=img)["image"]
assert sum([transformed[:, :, c].max() for c in range(img.shape[2])]) == 2
aug = A.ChannelDropout(channel_drop_range=(2, 2), always_apply=True) # Drop two channels
transformed = aug(image=img)["image"]
assert sum([transformed[:, :, c].max() for c in range(img.shape[2])]) == 1
def get_tfms_albu(p=0.6): return A.Compose([
A.HorizontalFlip(),
A.ShiftScaleRotate(rotate_limit=15),
A.ChannelDropout(p=0.1),
A.RandomRain(p=0.1),
A.GridDistortion(p=0.2)
], p=p)
def get_hard_augmentations(image_size):
return A.Compose([
A.OneOf([
A.ShiftScaleRotate(shift_limit=0.05, scale_limit=0.1,
rotate_limit=45,
border_mode=cv2.BORDER_CONSTANT, value=0),
A.ElasticTransform(alpha_affine=0,
alpha=35,
sigma=5,
border_mode=cv2.BORDER_CONSTANT,
value=0),
A.OpticalDistortion(distort_limit=0.11, shift_limit=0.15,
border_mode=cv2.BORDER_CONSTANT,
value=0),
A.GridDistortion(border_mode=cv2.BORDER_CONSTANT,
value=0),
A.NoOp()
]),
A.OneOf([
A.RandomSizedCrop(min_max_height=(int(image_size[0] * 0.75), image_size[0]),
height=image_size[0],
width=image_size[1], p=0.3),
A.NoOp()
]),
A.ISONoise(p=0.5),
# Brightness/contrast augmentations
A.OneOf([
A.RandomBrightnessContrast(brightness_limit=0.5,
contrast_limit=0.4),
IndependentRandomBrightnessContrast(brightness_limit=0.25,
contrast_limit=0.24),
A.RandomGamma(gamma_limit=(50, 150)),
A.NoOp()
]),
A.OneOf([
A.RGBShift(r_shift_limit=40, b_shift_limit=30, g_shift_limit=30),
A.HueSaturationValue(hue_shift_limit=10,
sat_shift_limit=10),
A.ToGray(p=0.2),
A.NoOp()
]),
A.ChannelDropout(),
A.RandomGridShuffle(p=0.3),
# D4
A.Compose([
A.RandomRotate90(),
A.Transpose()
])
])
def get_middle_man_data_aug(mean, standard_deviation, height, width):
return [
A.Normalize(mean=mean, std=standard_deviation, always_apply=True, p=1.0),
A.PadIfNeeded(min_height=height+8, min_width=width+8),
A.RandomCrop(height, width, always_apply=True, p=1.0),
A.HorizontalFlip(p=0.5),
A.GridDistortion (num_steps=5, distort_limit=0.3, interpolation=1, border_mode=4, value=None, mask_value=None, always_apply=False, p=0.5),
A.ChannelDropout(channel_drop_range=(1, 1), fill_value=0, always_apply=False, p=0.5),
A.Cutout(num_holes=1, max_h_size=8, max_w_size=8, fill_value=mean, always_apply=False, p=0.5),
ToTensor()
]
def get_train_transform(self):
if self.fullsizeimage:
resize_height = 640
resize_width = 1280
else:
resize_height = 512
resize_width = 512
return A.Compose(
[
A.OneOf(
[
A.RandomBrightnessContrast(p=0.5),
A.RGBShift(p=0.5),
A.HueSaturationValue(p=0.5),
A.ToGray(p=0.5),
A.ChannelDropout(p=0.5),
A.ChannelShuffle(p=0.5),
],
p=0.5,
),
A.OneOf(
[
A.Blur(p=0.5),
A.GaussNoise(p=0.5),
A.IAASharpen(p=0.5),
],
p=0.5,
),
A.OneOf(
[
A.Rotate(limit=20, p=0.5),
A.HorizontalFlip(p=0.5),
A.VerticalFlip(p=0.5),
],
p=0.5,
),
A.Resize(height=resize_height, width=resize_width, p=1.0),
ToTensorV2(p=1.0),
],
bbox_params=A.BboxParams(
format="pascal_voc",
min_area=0,
min_visibility=0,
label_fields=["labels"],
),
)
def __init__(self):
self.policy = A.Compose([
A.OneOf([
A.Rotate(180),
A.Flip(),
], p=0.3),
A.ShiftScaleRotate(shift_limit=0.2, scale_limit=0.5, rotate_limit=0, p=0.2),
A.OneOf([
A.CoarseDropout(max_holes=16, max_height=16, max_width=16, p=0.3),
A.GridDropout(ratio=0.3, p=0.3),
]),
A.OneOf([
A.ElasticTransform(sigma=10, alpha_affine=25, p=0.3),
A.RandomFog(fog_coef_lower=0.2, fog_coef_upper=0.7, p=0.2),
], p=0.2),
A.OneOf([
A.IAAAdditiveGaussianNoise(),
A.GaussNoise(),
A.ISONoise()
], p=0.2),
A.OneOf([
A.MotionBlur(p=.3),
A.MedianBlur(blur_limit=5, p=0.3),
A.Blur(blur_limit=5, p=0.3),
A.GaussianBlur(p=0.3)
], p=0.2),
A.OneOf([
A.ChannelShuffle(p=.3),
A.HueSaturationValue(p=0.3),
A.ToGray(p=0.3),
A.ChannelDropout(p=0.3),
A.InvertImg(p=0.1)
], p=0.2),
A.OneOf([
A.OpticalDistortion(p=0.3),
A.GridDistortion(p=.2),
A.IAAPiecewiseAffine(p=0.3),
], p=0.2),
A.OneOf([
A.CLAHE(clip_limit=2),
A.IAASharpen(),
A.IAAEmboss(),
], p=0.2),
A.RandomBrightnessContrast(brightness_limit=0.3, contrast_limit=0.3, p=0.3),
A.Solarize(p=0.2),
])
def trainset_albumentations(train_dataset_x, train_y):
ch_means = (0.48043839, 0.44820218, 0.39760034)
ch_std = (0.27698959, 0.26908774, 0.28216029)
train_albumentations_transform = A.Compose([
A.Rotate((-20.0, 20.0)),
# A.CourseDropout(0.2),
A.HorizontalFlip(),
A.ChannelDropout(channel_drop_range=(1, 1)),
A.RandomBrightness(0.2),
A.Normalize(
mean=[0.49139968, 0.48215841, 0.44653091],
std=[0.24703223, 0.24348513, 0.26158784],
),
A.Cutout(num_holes=1, max_h_size=10, max_w_size=10, always_apply=True),
ToTensor()
])
return AlbumentationsDataset(
file_paths=train_dataset_x,
labels=train_y,
transform=train_albumentations_transform,
)
def get_training_augmentation(size):
train_transform = [
A.LongestMaxSize(max_size=size, always_apply=True),
A.PadIfNeeded(min_height=size,
min_width=size,
always_apply=True,
border_mode=0),
# A.RandomCrop(height=size, width=size, always_apply=True),
# A.VerticalFlip(p=0.5),
# A.HorizontalFlip(p=0.5),
# A.RandomRotate90(p=0.5),
A.ShiftScaleRotate(scale_limit=0.2,
rotate_limit=0,
shift_limit=0.2,
p=0.1,
border_mode=0),
A.IAAPerspective(p=0.1),
A.CoarseDropout(p=0.1),
A.ChannelDropout(p=0.1),
A.RGBShift(p=0.1),
A.OneOf(
[A.OpticalDistortion(p=0.5),
A.GridDistortion(p=0.5)],
p=0.1,
),
A.OneOf(
[
A.CLAHE(p=0.5),
A.RandomBrightness(p=0.5),
A.RandomGamma(p=0.5),
],
p=0.5,
),
A.OneOf(
[
A.GaussianBlur(p=0.1),
A.IAASharpen(p=0.5),
A.Blur(blur_limit=5, p=0.5),
A.MotionBlur(blur_limit=5, p=0.5),
],
p=0.5,
),
A.OneOf(
[
A.RandomContrast(p=0.5),
A.HueSaturationValue(p=0.5),
],
p=0.1,
),
A.Lambda(mask=round_clip_0_1),
A.Cutout(num_holes=8,
max_h_size=20,
max_w_size=20,
fill_value=0,
always_apply=False,
p=0.2),
A.CoarseDropout(max_holes=8,
max_height=20,
max_width=20,
min_holes=None,
min_height=None,
min_width=None,
fill_value=0,
always_apply=False,
p=0.2),
# A.GlassBlur(sigma=0.7, max_delta=4, iterations=2, always_apply=False, mode='fast', p=0.2)
]
return A.Compose(train_transform)
def get_transform_imagenet(use_albu_aug):
if use_albu_aug:
train_transform = al.Compose([
# al.Flip(p=0.5),
al.Resize(256, 256, interpolation=2),
al.RandomResizedCrop(224,
224,
scale=(0.08, 1.0),
ratio=(3. / 4., 4. / 3.),
interpolation=2),
al.HorizontalFlip(),
al.OneOf(
[
al.OneOf(
[
al.ShiftScaleRotate(
border_mode=cv2.BORDER_CONSTANT,
rotate_limit=30), # , p=0.05),
al.OpticalDistortion(
border_mode=cv2.BORDER_CONSTANT,
distort_limit=5.0,
shift_limit=0.1),
# , p=0.05),
al.GridDistortion(border_mode=cv2.BORDER_CONSTANT
), # , p=0.05),
al.ElasticTransform(
border_mode=cv2.BORDER_CONSTANT,
alpha_affine=15), # , p=0.05),
],
p=0.1),
al.OneOf(
[
al.RandomGamma(), # p=0.05),
al.HueSaturationValue(), # p=0.05),
al.RGBShift(), # p=0.05),
al.CLAHE(), # p=0.05),
al.ChannelShuffle(), # p=0.05),
al.InvertImg(), # p=0.05),
],
p=0.1),
al.OneOf(
[
al.RandomSnow(), # p=0.05),
al.RandomRain(), # p=0.05),
al.RandomFog(), # p=0.05),
al.RandomSunFlare(num_flare_circles_lower=1,
num_flare_circles_upper=2,
src_radius=110),
# p=0.05, ),
al.RandomShadow(), # p=0.05),
],
p=0.1),
al.RandomBrightnessContrast(p=0.1),
al.OneOf(
[
al.GaussNoise(), # p=0.05),
al.ISONoise(), # p=0.05),
al.MultiplicativeNoise(), # p=0.05),
],
p=0.1),
al.OneOf(
[
al.ToGray(), # p=0.05),
al.ToSepia(), # p=0.05),
al.Solarize(), # p=0.05),
al.Equalize(), # p=0.05),
al.Posterize(), # p=0.05),
al.FancyPCA(), # p=0.05),
],
p=0.1),
al.OneOf(
[
# al.MotionBlur(blur_limit=1),
al.Blur(blur_limit=[3, 5]),
al.MedianBlur(blur_limit=[3, 5]),
al.GaussianBlur(blur_limit=[3, 5]),
],
p=0.1),
al.OneOf(
[
al.CoarseDropout(), # p=0.05),
al.Cutout(), # p=0.05),
al.GridDropout(), # p=0.05),
al.ChannelDropout(), # p=0.05),
al.RandomGridShuffle(), # p=0.05),
],
p=0.1),
al.OneOf(
[
al.Downscale(), # p=0.1),
al.ImageCompression(quality_lower=60), # , p=0.1),
],
p=0.1),
],
p=0.5),
al.Normalize(),
ToTensorV2()
])
else:
train_transform = transforms.Compose([
transforms.Resize(256),
transforms.RandomResizedCrop(224),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
])
test_transform = transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
])
if use_albu_aug:
train_transform = MultiDataTransformAlbu(train_transform)
else:
train_transform = MultiDataTransform(train_transform)
return train_transform, test_transform
def __call__(self, data):
rgb, thermal, depth, audio, label, id = data
rgb = rgb.astype(np.float32)
height, width, _ = rgb.shape
albumentations_transform_pixel = {
'Blur': albumentations.Blur(),
#'CLAHE':albumentations.CLAHE(),
'ChannelDropout': albumentations.ChannelDropout(),
'ChannelShuffle': albumentations.ChannelShuffle(),
'CoarseDropout': albumentations.CoarseDropout(),
#'Equalize':albumentations.Equalize(),
#'FancyPCA':albumentations.FancyPCA(),
'GaussNoise': albumentations.GaussNoise(),
'GaussianBlur': albumentations.GaussianBlur(),
#'GlassBlur':albumentations.GlassBlur(),
'HueSaturationValue': albumentations.HueSaturationValue(),
'IAAAdditiveGaussianNoise':
albumentations.IAAAdditiveGaussianNoise(),
#'ISONoise':albumentations.ISONoise(),
'RGBShift': albumentations.RGBShift(),
'RandomBrightnessContrast':
albumentations.RandomBrightnessContrast(),
'RandomFog': albumentations.RandomFog(),
#'RandomGamma':albumentations.RandomGamma(),
'RandomRain': albumentations.RandomRain(),
'RandomShadow': albumentations.RandomShadow(),
'RandomSnow': albumentations.RandomSnow(),
'RandomSunFlare': albumentations.RandomSunFlare(),
'Solarize': albumentations.Solarize(),
}
albumentations_transform_bbox = {
#'HorizontalFlip':albumentations.HorizontalFlip(),
#'VerticalFlip':albumentations.VerticalFlip(),
#'CenterCrop':albumentations.CenterCrop(height=height-10, width=width-10, p=0.5),
#'RandomCropNearBBox':albumentations.RandomCropNearBBox(p=0.5),
#'Crop':albumentations.Crop(x_min=10, y_min =10, y_max=height-10, x_max=width-10, p=0.5),
#'ElasticTransform':albumentations.ElasticTransform(),
#'ShiftScaleRotate':albumentations.ShiftScaleRotate(),
}
transform = np.random.choice(
['None'] + list(albumentations_transform_pixel.keys()) +
list(albumentations_transform_bbox.keys()))
if transform in albumentations_transform_pixel:
aug = albumentations.Compose(
[albumentations_transform_pixel[transform]],
bbox_params={
'format': 'pascal_voc',
'label_fields': ['labels']
})
try:
annots = np.array(annots).astype(np.float32)
aug_result = aug(image=rgb,
bboxes=annots[:, :4],
labels=annots[:, 4])
rgb = aug_result['image']
annots = np.hstack([
aug_result['bboxes'],
np.array(aug_result['labels']).reshape(-1, 1)
])
except Exception as e:
print(
f"transform={transform} aug_result['bboxes']={aug_result['bboxes']} aug_result['labels']={aug_result['labels']}"
)
raise Exception(e)
elif transform in albumentations_transform_bbox:
aug = albumentations.Compose(
[albumentations_transform_bbox[transform]],
bbox_params={
'format': 'pascal_voc',
'label_fields': ['labels']
})
try:
annots = np.array(annots).astype(np.float32)
aug_result = aug(image=rgb,
bboxes=annots[:, :4],
labels=annots[:, 4])
rgb = aug_result['image']
label = np.hstack([
aug_result['bboxes'],
np.array(aug_result['labels']).reshape(-1, 1)
])
except Exception as e:
print(
f"transform={transform} aug_result['bboxes']={aug_result['bboxes']} aug_result['labels']={aug_result['labels']}"
)
raise Exception(e)
return rgb, thermal, depth, audio, label, id
A.ShiftScaleRotate(
rotate_limit=18, p=1, border_mode=cv.BORDER_CONSTANT),
A.IAAAffine(shear=10, p=1, mode="constant"),
#A.Perspective(scale=(0.05, 0.15), keep_size=True, pad_mode=0, pad_val=0,
# mask_pad_val=0, fit_output=False, interpolation=1, always_apply=False, p=1),
],
p=1.0,
),
A.OneOf(
[
A.FancyPCA(alpha=0.1, always_apply=False, p=1),
A.Blur(p=1),
A.ToGray(p=0.8),
A.ColorJitter(
brightness=0.1, contrast=0.1, saturation=0.1, hue=0.1, p=1),
A.ChannelDropout((1, 1), fill_value=0, always_apply=False, p=1),
],
p=0.3,
),
A.OneOf(
[ #A.GaussNoise (var_limit=(10.0, 50.0), mean=0, per_channel=True, always_apply=False, p=0.5),
A.Equalize(mode='cv',
by_channels=True,
mask=None,
mask_params=(),
always_apply=False,
p=0.8),
A.MotionBlur(blur_limit=4, p=1),
],
p=0.1,
)
#return augmented["image"], augmented["mask"]
return input_tensor,label_tensor
def checkdataLoader(self):
print("dummy")
if __name__=="__main__":
transform=aldu.Compose([
###pixel_lvl_transform
aldu.CLAHE(p=1),
#aldu.RandomBrightnessContrast(),
aldu.ChannelDropout(),
#aldu.ISONoise() ,
#aldu.Downscale(),
#aldu.MultiplicativeNoise(),
###spatial lvl transform
aldu.RandomCrop(256,256),
aldu.HorizontalFlip(),
aldu.Rotate(),
aldu.MaskDropout(),
aldu.ElasticTransform(),
aldu.GridDistortion(),
### normalize
#aldu.Normalize()
])
AB.OneOf([
AB.CLAHE(clip_limit=2),
AB.IAASharpen(),
AB.IAAEmboss(),
AB.RandomBrightnessContrast(),
], p=0.2
),
# Change Gamma and Saturation
AB.HueSaturationValue(p=0.2),
AB.RandomGamma(p=0.2),
# Random Color Channel manipulation
AB.OneOf([
AB.ChannelShuffle(p=0.5),
AB.ChannelDropout(channel_drop_range=(1, 1), fill_value=0, p=0.5),
AB.ChannelDropout(channel_drop_range=(1, 1), fill_value=0, p=0.5),
AB.ChannelDropout(channel_drop_range=(1, 1), fill_value=0, p=0.5),
AB.ChannelDropout(channel_drop_range=(1, 1), fill_value=0, p=0.5),
AB.ChannelDropout(channel_drop_range=(1, 1), fill_value=0, p=0.5),
AB.ChannelDropout(channel_drop_range=(2, 2), fill_value=0, p=0.5),
AB.ChannelDropout(channel_drop_range=(2, 2), fill_value=0, p=0.5),
AB.ChannelDropout(channel_drop_range=(2, 2), fill_value=0, p=0.5),
], p=0.2
),
# Normalize image using values from ImageNet
AB.Normalize(
mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225],
)
albu.RandomRotate90(p=0.5),
albu.Flip(p=0.5),
albu.Transpose(p=0.5),
albu.ShiftScaleRotate(shift_limit=0.0, scale_limit=0.2, rotate_limit=30, p=0.4),
albu.core.composition.PerChannel(
albu.OneOf([
albu.RandomBrightnessContrast(brightness_limit=0.2, contrast_limit=0.2, p=0.25),
# albu.HueSaturationValue(hue_shift_limit=0, sat_shift_limit=0.2, val_shift_limit=0.2, p=0.25),
albu.Blur(blur_limit=3, p=.1)
]), p=1.0),
albu.Cutout(max_h_size=70),
# albu.ChannelDropout(p=0.05)
]),
"w_dropchannel": albu.Compose([
albu.ChannelDropout(channel_drop_range=(1, 1), p=0.7),
albu.RandomResizedCrop(1024, 1024, (0.7, 1)),
albu.MaskDropout(max_objects=14, p=0.9),
albu.RandomRotate90(p=0.5),
albu.Flip(p=0.5),
albu.Transpose(p=0.5),
albu.ShiftScaleRotate(shift_limit=0.2, scale_limit=0.2, rotate_limit=30, p=0.4),
albu.core.composition.PerChannel(
albu.OneOf([
albu.RandomBrightnessContrast(brightness_limit=0.2, contrast_limit=0.2, p=0.25),
albu.HueSaturationValue(hue_shift_limit=0, sat_shift_limit=0.2, val_shift_limit=0.2, p=0.25),
albu.Blur(blur_limit=3, p=.1)
]), p=1.0),
albu.Cutout(max_h_size=100),
# albu.ChannelDropout(p=0.05)
def get_hard_augmentations_v2(image_size):
return A.Compose([
A.OneOf([
A.ShiftScaleRotate(shift_limit=0.05,
scale_limit=0.1,
rotate_limit=45,
border_mode=cv2.BORDER_CONSTANT,
value=0),
A.ElasticTransform(alpha_affine=0,
alpha=35,
sigma=5,
border_mode=cv2.BORDER_CONSTANT,
value=0),
A.OpticalDistortion(distort_limit=0.11,
shift_limit=0.15,
border_mode=cv2.BORDER_CONSTANT,
value=0),
A.GridDistortion(border_mode=cv2.BORDER_CONSTANT, value=0),
A.NoOp()
]),
A.OneOf([
ZeroTopAndBottom(p=0.3),
A.RandomSizedCrop(min_max_height=(int(image_size[0] * 0.75),
image_size[0]),
height=image_size[0],
width=image_size[1],
p=0.3),
A.NoOp()
]),
A.ISONoise(p=0.5),
A.JpegCompression(p=0.3, quality_lower=75),
# Brightness/contrast augmentations
A.OneOf([
A.RandomBrightnessContrast(brightness_limit=0.5,
contrast_limit=0.4),
IndependentRandomBrightnessContrast(brightness_limit=0.25,
contrast_limit=0.24),
A.RandomGamma(gamma_limit=(50, 150)),
A.NoOp()
]),
A.OneOf([
FancyPCA(alpha_std=6),
A.RGBShift(r_shift_limit=40, b_shift_limit=30, g_shift_limit=30),
A.HueSaturationValue(hue_shift_limit=10, sat_shift_limit=10),
A.ToGray(p=0.2),
A.NoOp()
]),
# Intentionally destroy image quality and assign 0 class in this case
# A.Compose([
# BrightnessContrastDestroy(p=0.1),
# A.OneOf([
# MakeTooBlurry(),
# MakeTooBlurryMedian(),
# A.NoOp()
# ], p=0.1),
# ]),
# Add preprocessing method as an augmentation
ChannelIndependentCLAHE(p=0.5),
A.OneOf([
A.ChannelDropout(p=0.2),
A.CoarseDropout(p=0.1,
max_holes=2,
max_width=256,
max_height=256,
min_height=16,
min_width=16),
A.NoOp()
]),
A.RandomGridShuffle(p=0.3),
DiagnosisNoise(p=0.2),
# D4
A.Compose([A.RandomRotate90(), A.Transpose()])
])
def get_train_transforms_mmdetection(input_size,
use_crop=False,
use_no_color_aug=False,
use_center_crop=False,
center_crop_ratio=0.9,
use_gray=False):
if isinstance(input_size, int):
input_size = (input_size[0], input_size[1])
return al.Compose([
al.RandomResizedCrop(height=input_size[0],
width=input_size[1],
scale=(0.4, 1.0),
interpolation=0,
p=0.5),
al.Resize(input_size[0], input_size[1], p=1.0),
al.HorizontalFlip(p=0.5),
al.OneOf([
al.ShiftScaleRotate(border_mode=0,
shift_limit=(-0.2, 0.2),
scale_limit=(-0.2, 0.2),
rotate_limit=(-20, 20)),
al.OpticalDistortion(border_mode=0,
distort_limit=[-0.5, 0.5],
shift_limit=[-0.5, 0.5]),
al.GridDistortion(
num_steps=5, distort_limit=[-0., 0.3], border_mode=0),
al.ElasticTransform(border_mode=0),
al.IAAPerspective(),
al.RandomGridShuffle()
],
p=0.1),
al.Rotate(limit=(-25, 25), border_mode=0, p=0.1),
al.OneOf([
al.RandomBrightnessContrast(brightness_limit=(-0.2, 0.2),
contrast_limit=(-0.2, 0.2)),
al.HueSaturationValue(hue_shift_limit=(-20, 20),
sat_shift_limit=(-30, 30),
val_shift_limit=(-20, 20)),
al.RandomGamma(gamma_limit=(30, 150)),
al.RGBShift(),
al.CLAHE(clip_limit=(1, 15)),
al.ChannelShuffle(),
al.InvertImg(),
],
p=0.1),
al.RandomSnow(p=0.05),
al.RandomRain(p=0.05),
al.RandomFog(p=0.05),
al.RandomSunFlare(num_flare_circles_lower=1,
num_flare_circles_upper=2,
src_radius=110,
p=0.05),
al.RandomShadow(p=0.05),
al.GaussNoise(var_limit=(10, 20), p=0.05),
al.ISONoise(color_shift=(0, 15), p=0.05),
al.MultiplicativeNoise(p=0.05),
al.OneOf([
al.ToGray(p=1. if use_gray else 0.05),
al.ToSepia(p=0.05),
al.Solarize(p=0.05),
al.Equalize(p=0.05),
al.Posterize(p=0.05),
al.FancyPCA(p=0.05),
],
p=0.05),
al.OneOf([
al.MotionBlur(blur_limit=(3, 7)),
al.Blur(blur_limit=(3, 7)),
al.MedianBlur(blur_limit=3),
al.GaussianBlur(blur_limit=3),
],
p=0.05),
al.CoarseDropout(p=0.05),
al.Cutout(num_holes=30,
max_h_size=37,
max_w_size=37,
fill_value=0,
p=0.05),
al.GridDropout(p=0.05),
al.ChannelDropout(p=0.05),
al.Downscale(scale_min=0.5, scale_max=0.9, p=0.1),
al.ImageCompression(quality_lower=60, p=0.2),
al.Normalize(),
ToTensorV2()
])
def train_function(gpu, world_size, node_rank, gpus, fold_number, group_name):
import torch.multiprocessing
torch.multiprocessing.set_sharing_strategy('file_system')
torch.manual_seed(25)
np.random.seed(25)
rank = node_rank * gpus + gpu
dist.init_process_group(
backend='nccl',
init_method='env://',
world_size=world_size,
rank=rank
)
device = torch.device("cuda:{}".format(gpu) if torch.cuda.is_available() else "cpu")
batch_size = 64
width_size = 416
init_lr = 1e-4
end_lr = 1e-6
n_epochs = 20
emb_size = 512
margin = 0.5
dropout = 0.0
iters_to_accumulate = 1
if rank == 0:
wandb.init(project='shopee_effnet0', group=group_name, job_type=str(fold_number))
checkpoints_dir_name = 'effnet0_{}_{}_{}'.format(width_size, dropout, group_name)
os.makedirs(checkpoints_dir_name, exist_ok=True)
wandb.config.model_name = checkpoints_dir_name
wandb.config.batch_size = batch_size
wandb.config.width_size = width_size
wandb.config.init_lr = init_lr
wandb.config.n_epochs = n_epochs
wandb.config.emb_size = emb_size
wandb.config.dropout = dropout
wandb.config.iters_to_accumulate = iters_to_accumulate
wandb.config.optimizer = 'adam'
wandb.config.scheduler = 'ShopeeScheduler'
df = pd.read_csv('../../dataset/reliable_validation_tm.csv')
train_df = df[df['fold_group'] != fold_number]
train_transforms = alb.Compose([
alb.RandomResizedCrop(width_size, width_size),
alb.ShiftScaleRotate(shift_limit=0.1, rotate_limit=30),
alb.HorizontalFlip(),
alb.OneOf([
alb.Sequential([
alb.HueSaturationValue(hue_shift_limit=50),
alb.RandomBrightnessContrast(),
]),
alb.FancyPCA(),
alb.ChannelDropout(),
alb.ChannelShuffle(),
alb.RGBShift()
]),
alb.CoarseDropout(max_height=int(width_size*0.1), max_width=int(width_size*0.1)),
alb.OneOf([
alb.ElasticTransform(),
alb.OpticalDistortion(),
alb.GridDistortion()
]),
alb.Resize(width_size, width_size),
alb.Normalize(mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225)),
ToTensorV2()
])
train_set = ImageDataset(train_df, train_df, '../../dataset/train_images', train_transforms)
sampler = DistributedSampler(train_set, num_replicas=world_size, rank=rank, shuffle=True)
train_dataloader = DataLoader(train_set, batch_size=batch_size // world_size, shuffle=False, num_workers=4,
sampler=sampler)
# valid_df = df[df['fold_strat'] == fold_number]
valid_transforms = alb.Compose([
alb.Resize(width_size, width_size),
alb.Normalize(mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225)),
ToTensorV2()
])
# valid_set = ImageDataset(train_df, valid_df, '../../dataset/train_images', valid_transforms)
# valid_dataloader = DataLoader(valid_set, batch_size=batch_size // world_size, shuffle=False, num_workers=4)
test_df = df[df['fold_group'] == fold_number]
test_set = ImageDataset(test_df, test_df, '../../dataset/train_images', valid_transforms)
test_dataloader = DataLoader(test_set, batch_size=batch_size // world_size, shuffle=False, num_workers=4)
model = EfficientNetArcFace(emb_size, train_df['label_group'].nunique(), device, dropout=dropout,
backbone='tf_efficientnet_b0_ns', pretrained=True, margin=margin, is_amp=True)
model = SyncBatchNorm.convert_sync_batchnorm(model)
model.to(device)
model = DistributedDataParallel(model, device_ids=[gpu])
scaler = GradScaler()
criterion = CrossEntropyLoss()
# criterion = LabelSmoothLoss(smoothing=0.1)
optimizer = optim.Adam(model.parameters(), lr=init_lr)
# scheduler = CosineAnnealingLR(optimizer, T_max=n_epochs, eta_min=end_lr,
# last_epoch=-1)
# scheduler = CosineAnnealingWarmRestarts(optimizer, T_0=2000, T_mult=1,
# eta_min=end_lr, last_epoch=-1)
scheduler = ShopeeScheduler(optimizer, lr_start=init_lr,
lr_max=init_lr*batch_size, lr_min=end_lr)
for epoch in range(n_epochs):
train_loss, train_duration, train_f1 = train_one_epoch(
model, train_dataloader, optimizer, criterion, device, scaler,
scheduler=None, iters_to_accumulate=iters_to_accumulate)
scheduler.step()
if rank == 0:
# valid_loss, valid_duration, valid_f1 = evaluate(model, valid_dataloader, criterion, device)
embeddings = get_embeddings(model, test_dataloader, device)
embeddings_f1 = validate_embeddings_f1(embeddings, test_df)
wandb.log({'train_loss': train_loss, 'train_f1': train_f1,
'embeddings_f1': embeddings_f1, 'epoch': epoch})
filename = '{}_foldnum{}_epoch{}_train_loss{}_f1{}'.format(
checkpoints_dir_name, fold_number+1, epoch+1,
round(train_loss, 3), round(embeddings_f1, 3))
torch.save(model.module.state_dict(), os.path.join(checkpoints_dir_name, '{}.pth'.format(filename)))
# np.savez_compressed(os.path.join(checkpoints_dir_name, '{}.npz'.format(filename)), embeddings=embeddings)
print('FOLD NUMBER %d\tEPOCH %d:\t'
'TRAIN [duration %.3f sec, loss: %.3f, avg f1: %.3f]\t'
'VALID EMBEDDINGS [avg f1: %.3f]\tCurrent time %s' %
(fold_number + 1, epoch + 1, train_duration,
train_loss, train_f1, embeddings_f1,
str(datetime.now(timezone('Europe/Moscow')))))
if rank == 0:
wandb.finish()
def get_train_transforms_atopy(input_size,
use_crop=False,
use_no_color_aug=False):
if use_crop:
resize = [
al.Resize(int(input_size * 1.2), int(input_size * 1.2)),
al.RandomSizedCrop(min_max_height=(int(input_size * 0.6),
int(input_size * 1.2)),
height=input_size,
width=input_size)
]
else:
resize = [al.Resize(input_size, input_size)]
return al.Compose(resize + [
al.Flip(p=0.5),
al.OneOf([
al.RandomRotate90(),
al.Rotate(limit=180),
], p=0.5),
al.OneOf([
al.ShiftScaleRotate(),
al.OpticalDistortion(),
al.GridDistortion(),
al.ElasticTransform(),
],
p=0.3),
al.RandomGridShuffle(p=0.05),
al.OneOf([
al.RandomGamma(),
al.HueSaturationValue(),
al.RGBShift(),
al.CLAHE(),
al.ChannelShuffle(),
al.InvertImg(),
],
p=0.1),
al.RandomSnow(p=0.05),
al.RandomRain(p=0.05),
al.RandomFog(p=0.05),
al.RandomSunFlare(p=0.05),
al.RandomShadow(p=0.05),
al.RandomBrightnessContrast(p=0.05),
al.GaussNoise(p=0.2),
al.ISONoise(p=0.2),
al.MultiplicativeNoise(p=0.2),
al.ToGray(p=0.05),
al.ToSepia(p=0.05),
al.Solarize(p=0.05),
al.Equalize(p=0.05),
al.Posterize(p=0.05),
al.FancyPCA(p=0.05),
al.OneOf([
al.MotionBlur(blur_limit=3),
al.Blur(blur_limit=3),
al.MedianBlur(blur_limit=3),
al.GaussianBlur(blur_limit=3),
],
p=0.05),
al.CoarseDropout(p=0.05),
al.Cutout(p=0.05),
al.GridDropout(p=0.05),
al.ChannelDropout(p=0.05),
al.Downscale(p=0.1),
al.ImageCompression(quality_lower=60, p=0.2),
al.Normalize(),
ToTensorV2()
])
def train_process(data_path, config):
def _worker_init_fn_():
import random
import numpy as np
import torch
random_seed = config.random_seed
torch.manual_seed(random_seed)
np.random.seed(random_seed)
random.seed(random_seed)
if torch.cuda.is_available():
torch.cuda.manual_seed(random_seed)
input_size = (config.img_height, config.img_width)
PAD_VALUE = (0, 0, 0)
IGNORE_INDEX = 255
transforms = [
abm.RandomResizedCrop(
scale=(0.7, 1),
ratio=(1.5, 2),
height=config.img_height,
width=config.img_width,
interpolation=cv2.INTER_NEAREST,
always_apply=True,
),
abm.OneOf([abm.IAAAdditiveGaussianNoise(),
abm.GaussNoise()], p=0.5),
abm.OneOf(
[
abm.MedianBlur(blur_limit=3),
abm.GaussianBlur(blur_limit=3),
abm.MotionBlur(blur_limit=3),
],
p=0.5,
),
abm.OneOf([
abm.ShiftScaleRotate(
rotate_limit=7,
interpolation=cv2.INTER_NEAREST,
border_mode=cv2.BORDER_CONSTANT,
value=PAD_VALUE,
mask_value=IGNORE_INDEX,
p=1.0,
),
abm.ElasticTransform(
interpolation=cv2.INTER_NEAREST,
border_mode=cv2.BORDER_CONSTANT,
alpha_affine=30,
value=PAD_VALUE,
mask_value=IGNORE_INDEX,
p=1.0,
),
abm.Perspective(
scale=(0.05),
interpolation=cv2.INTER_NEAREST,
pad_mode=cv2.BORDER_CONSTANT,
pad_val=PAD_VALUE,
mask_pad_val=IGNORE_INDEX,
keep_size=True,
fit_output=True,
p=1.0,
),
]),
abm.RandomGamma(gamma_limit=(80, 120), p=0.5),
abm.RandomBrightnessContrast(brightness_limit=(-0.5, 0.5),
contrast_limit=(-0.5, 0.5),
p=0.5),
abm.HueSaturationValue(hue_shift_limit=20,
sat_shift_limit=30,
val_shift_limit=20,
p=0.5),
abm.RandomShadow(p=0.5),
abm.ChannelShuffle(p=0.5),
abm.ChannelDropout(p=0.5),
abm.HorizontalFlip(p=0.5),
abm.ImageCompression(quality_lower=50, p=0.5),
abm.Cutout(num_holes=100, max_w_size=8, max_h_size=8, p=0.5),
]
data_transform = DataTransformBase(transforms=transforms,
input_size=input_size,
normalize=True)
train_dataset = EgoRailDataset(data_path=data_path,
phase="train",
transform=data_transform)
val_dataset = EgoRailDataset(data_path=data_path,
phase="val",
transform=data_transform)
# train_dataset.weighted_class()
weighted_values = [8.90560578, 1.53155476]
train_data_loader = DataLoader(
train_dataset,
batch_size=config.batch_size,
shuffle=True,
num_workers=config.num_workers,
drop_last=True,
worker_init_fn=_worker_init_fn_(),
)
val_data_loader = DataLoader(
val_dataset,
batch_size=config.batch_size,
shuffle=False,
num_workers=config.num_workers,
drop_last=True,
)
data_loaders_dict = {"train": train_data_loader, "val": val_data_loader}
model = BiSeNetV2(n_classes=config.num_classes)
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
criterion = OHEMCELoss(thresh=config.ohem_ce_loss_thresh,
weighted_values=weighted_values)
base_lr_rate = config.lr_rate / (config.batch_size *
config.batch_multiplier)
base_weight_decay = config.weight_decay * (config.batch_size *
config.batch_multiplier)
def _lambda_epoch(epoch):
import math
max_epoch = config.num_epochs
return math.pow((1 - epoch * 1.0 / max_epoch), 0.9)
optimizer = torch.optim.SGD(
model.parameters(),
lr=base_lr_rate,
momentum=config.momentum,
weight_decay=base_weight_decay,
)
scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=_lambda_epoch)
trainer = BiSeNetV2Trainer(
model=model,
criterion=criterion,
metric_func=None,
optimizer=optimizer,
data_loaders_dict=data_loaders_dict,
config=config,
scheduler=scheduler,
device=device,
)
if config.snapshot and os.path.isfile(config.snapshot):
trainer.resume_checkpoint(config.snapshot)
with torch.autograd.set_detect_anomaly(True):
trainer.train()
transforms.append(A.ShiftScaleRotate(p=p))
if p := trans_cfg.get('elastictransform', False):
transforms.append(A.ElasticTransform(p=p))
if p := trans_cfg.get('griddistortion', False):
transforms.append(A.GridDistortion(p=p))
if p := trans_cfg.get('hflip', False):
transforms.append(A.HorizontalFlip(p=p))
if p := trans_cfg.get('vflip', False):
transforms.append(A.VerticalFlip(p=p))
if p := trans_cfg.get('brightnesscontrast', False):
transforms.append(A.RandomBrightnessContrast(p=p))
if p := trans_cfg.get('griddropout', False):
transforms.append(
A.GridDropout(fill_value=0, mask_fill_value=0, p=p))
if p := trans_cfg.get('channeldropout', False):
transforms.append(A.ChannelDropout(channel_drop_range=(1, 1), p=p))
if p := trans_cfg.get('blur', False):
transforms.append(
A.OneOf([
A.MedianBlur(blur_limit=5, p=p),
A.Blur(blur_limit=5, p=p)
]))
if p := trans_cfg.get('noise', False):
transforms.append(
A.OneOf([A.GaussNoise(p=p),
A.MultiplicativeNoise(p=p)]))
if p := trans_cfg.get('hsv', False):
transforms.append(A.HueSaturationValue(p=p))
# Do these last so are less likely to need to reflect etc. during skew/rotation
if trans_cfg.get('centrecrop', False):
