Augmentations

torch_em/transform/raw.py

@@ -92,7 +92,7 @@ class RandomContrast():
     """
     Adjust contrast by scaling image to `mean + alpha * (image - mean)`.
     """
-    def __init__(self, alpha=(0.05, 4), mean=0.5, clip_kwargs={'a_min': 0, 'a_max': 1}):
+    def __init__(self, alpha=(0.5, 2), mean=0.5, clip_kwargs={'a_min': 0, 'a_max': 1}):
         self.alpha = alpha
         self.mean = mean
         self.clip_kwargs = clip_kwargs
@@ -109,33 +109,61 @@ class AdditiveGaussianNoise():
     """
     Add random Gaussian noise to image.
     """
-    def __init__(self, scale=(0.0, 0.75)):
+    def __init__(self, scale=(0.0, 0.3), clip_kwargs={'a_min': 0, 'a_max': 1}):
         self.scale = scale
+        self.clip_kwargs = clip_kwargs
 
     def __call__(self, img):
         std = np.random.uniform(self.scale[0], self.scale[1])
         gaussian_noise = np.random.normal(0, std, size=img.shape)
+        if self.clip_kwargs:
+            return np.clip(img + gaussian_noise, 0, 1)
         return img + gaussian_noise
 
 
 class AdditivePoissonNoise():
     """
     Add random Poisson noise to image.
     """
-    def __init__(self, lam=(0.0, 0.75)):
+    # TODO: not sure if Poisson noise like this does make sense
+    # for data that is already normalized
+    def __init__(self, lam=(0.0, 0.1), clip_kwargs={'a_min': 0, 'a_max': 1}):
         self.lam = lam
+        self.clip_kwargs = clip_kwargs
 
     def __call__(self, img):
         lam = np.random.uniform(self.lam[0], self.lam[1])
-        poisson_noise = np.random.poisson(lam, size=img.shape)
+        poisson_noise = np.random.poisson(lam, size=img.shape) / lam
+        if self.clip_kwargs:
+            return np.clip(img + poisson_noise, 0, 1)
         return img + poisson_noise
 
 
+class PoissonNoise():
+    """
+    Add random data-dependent Poisson noise to image.
+    """
+    def __init__(self, multiplier=(5.0, 10.0), clip_kwargs={'a_min': 0, 'a_max': 1}):
+        self.multiplier = multiplier
+        self.clip_kwargs = clip_kwargs
+
+    def __call__(self, img):
+        multiplier = np.random.uniform(self.multiplier[0], self.multiplier[1])
+        offset = img.min()
+        poisson_noise = np.random.poisson((img - offset) * multiplier)
+        if isinstance(img, torch.Tensor):
+            poisson_noise = torch.Tensor(poisson_noise)
+        poisson_noise = poisson_noise / multiplier + offset
+        if self.clip_kwargs:
+            return np.clip(poisson_noise, **self.clip_kwargs)
+        return poisson_noise
+
+
 class GaussianBlur():
     """
     Blur the image.
     """
-    def __init__(self, kernel_size=(2, 24), sigma=(0, 5)):
+    def __init__(self, kernel_size=(2, 12), sigma=(0, 2.5)):
         self.kernel_size = kernel_size
         self.sigma = sigma
 
@@ -176,13 +204,13 @@ def get_raw_transform(normalizer=standardize, augmentation1=None, augmentation2=
 # The default values are made for an image with pixel values in
 # range [0, 1]. That the image is in this range is ensured by an
 # initial normalizations step.
-def get_default_mean_teacher_augmentations(p=0.5):
+def get_default_mean_teacher_augmentations(p=0.3):
     norm = normalize
     aug1 = transforms.Compose([
         normalize,
         transforms.RandomApply([GaussianBlur()], p=p),
-        transforms.RandomApply([AdditiveGaussianNoise()], p=p),
-        transforms.RandomApply([AdditivePoissonNoise()], p=p)
+        transforms.RandomApply([PoissonNoise()], p=p/2),
+        transforms.RandomApply([AdditiveGaussianNoise()], p=p/2),
     ])
     aug2 = transforms.RandomApply([RandomContrast()], p=p)
     return get_raw_transform(