symlog transform of the target

sCOCA_ML/prepare_data/prepare_gravpot_data.py

@@ -1,9 +1,14 @@
+from .utils_functions import *
+
 def prepare_data(batch,
-                 scale_phi_ini:float   = 1000.0,
+                 scale_phi_ini:float        = 1000.0,
-                 scale_delta_ini:float =   12.0,
+                 scale_delta_ini:float      =   12.0,
-                 scale_target:float    =  600.0,
+                 scale_target:float         =   50.0,
+                 lin_threshold_target:float =     2.,
                  ):
     
+    target_transform = CustomSymlogTransform(threshold=lin_threshold_target)
+    
     # delta_ini = batch['input'][:, [0], :, :, :]
     phi_ini = batch['input'][:, [1], :, :, :]
     D1 = batch['style'][:, [0], None, None, None]
@@ -17,6 +22,7 @@ def prepare_data(batch,
 
     _target = (gravpot/D1 - phi_ini)/D1
     _target /= scale_target
+    _target = target_transform(_target)
 
     _style = batch['style']
     

sCOCA_ML/prepare_data/utils_functions.py

@@ -0,0 +1,41 @@
+import torch
+
+class Transformation:
+    def __init__(self, transform=None, inverse_transform=None):
+        self.transform = transform
+        self.inverse_transform = inverse_transform
+
+    def __call__(self, sample):
+        if self.transform:
+            return self.transform(sample)
+        return sample
+    
+    def inverse(self, sample):
+        if self.inverse_transform:
+            return self.inverse_transform(sample)
+        return sample
+    
+
+def custom_symlog(x, threshold=1.):
+    """
+    Custom symmetric logarithm function.
+    Returns log(1 + |x| / threshold) * sign(x) * threshold.
+    """
+    return torch.sign(x) * torch.log1p(torch.abs(x) / threshold) * threshold
+
+def custom_inverse_symlog(x, threshold=1.):
+    """
+    Custom inverse symmetric logarithm function.
+    Returns sign(x) * (exp(|x| / threshold) - 1) * threshold.
+    """
+    return torch.sign(x) * (torch.exp(torch.abs(x) / threshold) - 1) * threshold
+
+
+def CustomSymlogTransform(threshold=1.):
+    """
+    Custom transformation object for symmetric logarithm.
+    """
+    return Transformation(
+        transform=lambda x: custom_symlog(x, threshold),
+        inverse_transform=lambda x: custom_inverse_symlog(x, threshold)
+    )