From ab581084a838825a059217363e60ae13d0947e13 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Anna=20B=C3=B8gevang=20Ekner?= <s193396@dtu.dk>
Date: Tue, 11 Feb 2025 11:01:28 +0100
Subject: [PATCH] refactored augmentations to use monai instead
---
qim3d/ml/_augmentations.py | 89 ++++++++++++++++++++------------------
1 file changed, 46 insertions(+), 43 deletions(-)
diff --git a/qim3d/ml/_augmentations.py b/qim3d/ml/_augmentations.py
index e12b1a51..ede3796c 100644
--- a/qim3d/ml/_augmentations.py
+++ b/qim3d/ml/_augmentations.py
@@ -1,8 +1,8 @@
-"""Class for choosing the level of data augmentations with albumentations"""
+"""Class for choosing the level of data augmentations with MONAI"""
class Augmentation:
"""
- Class for defining image augmentation transformations using the Albumentations library.
+ Class for defining image augmentation transformations using the MONAI library.
Args:
resize (str, optional): Specifies how the images should be reshaped to the appropriate size.
@@ -11,6 +11,7 @@ class Augmentation:
transform_test (str, optional): level of transformation for the test set.
mean (float, optional): The mean value for normalizing pixel intensities.
std (float, optional): The standard deviation value for normalizing pixel intensities.
+ is_3d (bool, optional): Specifies if the images are 3D or 2D. Default is True.
Raises:
ValueError: If the ´resize´ is neither 'crop', 'resize' or 'padding'.
@@ -25,7 +26,8 @@ class Augmentation:
transform_validation: str | None = None,
transform_test: str | None = None,
mean: float = 0.5,
- std: float = 0.5
+ std: float = 0.5,
+ is_3d: bool = True,
):
if resize not in ['crop', 'reshape', 'padding']:
@@ -37,70 +39,71 @@ class Augmentation:
self.transform_train = transform_train
self.transform_validation = transform_validation
self.transform_test = transform_test
+ self.is_3d = is_3d
- def augment(self, im_h: int, im_w: int, level: bool | None = None):
+ def augment(self, im_h: int, im_w: int, im_d: int | None = None, level: str | None = None):
"""
- Returns an albumentations.core.composition.Compose class depending on the augmentation level.
- A baseline augmentation is implemented regardless of the level, and a set of augmentations are added depending of the level.
- The A.Resize() function is used if the user has specified a 'resize' int or tuple at the creation of the Augmentation class.
+ Creates an augmentation pipeline based on the specified level.
Args:
- im_h (int): image height for resize.
- im_w (int): image width for resize.
- level (str, optional): level of augmentation.
+ im_h (int): Height of the image.
+ im_w (int): Width of the image.
+ im_d (int, optional): Depth of the image (for 3D).
+ level (str, optional): Level of augmentation. One of [None, 'light', 'moderate', 'heavy'].
Raises:
ValueError: If `level` is neither None, light, moderate nor heavy.
"""
- import albumentations as A
- from albumentations.pytorch import ToTensorV2
+ from monai.transforms import (
+ Compose, RandRotate90, RandFlip, RandAffine, ToTensor, \
+ RandGaussianSmooth, NormalizeIntensity, Resize, CenterSpatialCrop, SpatialPad
+ )
# Check if one of standard augmentation levels
if level not in [None,'light','moderate','heavy']:
raise ValueError(f"Invalid transformation level: {level}. Please choose one of the following levels: None, 'light', 'moderate', 'heavy'.")
- # Baseline
- baseline_aug = [
- A.Normalize(mean = (self.mean),std = (self.std)),
- ToTensorV2()
- ]
+ # Baseline augmentations
+ baseline_aug = [ToTensor()]
+
+ # For 2D, add normalization to the baseline augmentations
+ # TODO: Figure out how to properly do this in 3D (normalization should be done channel-wise)
+ if not self.is_3d:
+ baseline_aug.append(NormalizeIntensity(subtrahend=self.mean, divisor=self.std))
+
+ # Resize augmentations
if self.resize == 'crop':
- resize_aug = [
- A.CenterCrop(im_h,im_w)
- ]
+ resize_aug = [CenterSpatialCrop((im_d, im_h, im_w))] if self.is_3d else [CenterSpatialCrop((im_h, im_w))]
+
elif self.resize == 'reshape':
- resize_aug =[
- A.Resize(im_h,im_w)
- ]
- elif self.resize == 'padding':
- resize_aug = [
- A.PadIfNeeded(im_h,im_w,border_mode = 0) # OpenCV border mode
- ]
+ resize_aug = [Resize((im_d, im_h, im_w))] if self.is_3d else [Resize((im_h, im_w))]
+ elif self.resize == 'padding':
+ resize_aug = [SpatialPad((im_d, im_h, im_w))] if self.is_3d else [SpatialPad((im_h, im_w))]
+
# Level of augmentation
if level == None:
level_aug = []
+
elif level == 'light':
- level_aug = [
- A.RandomRotate90()
- ]
+ level_aug = [RandRotate90(prob=1, spatial_axes=(0, 1, 2))] if self.is_3d else [RandRotate90(prob=1)]
+
elif level == 'moderate':
level_aug = [
- A.RandomRotate90(),
- A.HorizontalFlip(p = 0.3),
- A.VerticalFlip(p = 0.3),
- A.GlassBlur(sigma = 0.7, p = 0.1),
- A.Affine(scale = [0.9,1.1], translate_percent = (0.1,0.1))
+ RandRotate90(prob=1, spatial_axes=(0, 1, 2)) if self.is_3d else RandRotate90(prob=1),
+ RandFlip(prob=0.3, spatial_axis=0),
+ RandFlip(prob=0.3, spatial_axis=1),
+ RandGaussianSmooth(sigma_x=(0.7, 0.7), prob=0.1),
+ RandAffine(prob=0.5, translate_range=(0.1, 0.1), scale_range=(0.9, 1.1)),
]
+
elif level == 'heavy':
level_aug = [
- A.RandomRotate90(),
- A.HorizontalFlip(p = 0.7),
- A.VerticalFlip(p = 0.7),
- A.GlassBlur(sigma = 1.2, iterations = 2, p = 0.3),
- A.Affine(scale = [0.8,1.4], translate_percent = (0.2,0.2), shear = (-15,15))
+ RandRotate90(prob=1, spatial_axes=(0, 1, 2)) if self.is_3d else RandRotate90(prob=1),
+ RandFlip(prob=0.7, spatial_axis=0),
+ RandFlip(prob=0.7, spatial_axis=1),
+ RandGaussianSmooth(sigma_x=(1.2, 1.2), prob=0.3),
+ RandAffine(prob=0.5, translate_range=(0.2, 0.2), scale_range=(0.8, 1.4), shear_range=(-15, 15))
]
- augment = A.Compose(level_aug + resize_aug + baseline_aug)
-
- return augment
\ No newline at end of file
+ return Compose(baseline_aug + resize_aug + level_aug)
\ No newline at end of file
--
GitLab