def fbresnet_augmentor(isTrain):
"""
Augmentor used in fb.resnet.torch, for BGR images in range [0,255].
"""
if isTrain:
augmentors = [
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), 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_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 train_dataloader(self):
# REQUIRED
ds_train = PretrainedSNEMI(folder=self.hparams.data_path,
train_or_valid='train',
size=40000,
# resize=int(self.hparams.shape),
debug=DEBUG
)
ag_train = [
imgaug.RotationAndCropValid(max_deg=180, interp=cv2.INTER_NEAREST),
imgaug.Flip(horiz=True, vert=False),
imgaug.Flip(horiz=False, vert=True),
imgaug.Transpose(),
imgaug.GoogleNetRandomCropAndResize(crop_area_fraction=(0.2, 0.5),
aspect_ratio_range=(0.5, 2.0),
interp=cv2.INTER_NEAREST,
target_shape=self.hparams.shape),
imgaug.ToFloat32(),
]
ds_train.reset_state()
# ds_train = AugmentImageComponent(ds_train, ag_train, 0)
ds_train = AugmentImageComponent(ds_train, [imgaug.Albumentations(AB.RandomBrightnessContrast())], (0))
ds_train = AugmentImageComponents(ds_train, ag_train, (0, 1))
ds_train = MapData(ds_train, lambda dp: [dp[0], 255.0*(dp[1]>0)*(1-skimage.segmentation.find_boundaries(dp[1], mode='inner'))])
ds_train = BatchData(ds_train, self.hparams.batch)
ds_train = PrintData(ds_train)
ds_train = MapData(ds_train, lambda dp: [torch.tensor(dp[0][:,np.newaxis,:,:]), torch.tensor(dp[1][:,np.newaxis,:,:]).float()])
ds_train = MultiProcessRunner(ds_train, num_proc=32, num_prefetch=8)
return ds_train
