tracked down the losses bug

losses.py

@@ -17,7 +17,7 @@ from utils import PURE_RGB
 #     return smoothness_loss
 
 
-def simple_preservation_loss(inputs, outputs, target_inputs=None, target_outputs=None):
+def preservation_loss(inputs, outputs, target_inputs=None, target_outputs=None):
     # Distance Preservation Component (or scaled euclidean if given targets)
     # Encourages the model to keep relative distances from the RGB space in the transformed space
     if target_inputs is None:
@@ -40,49 +40,19 @@ def simple_preservation_loss(inputs, outputs, target_inputs=None, target_outputs
     # print(rgb_norm)
 
     # Calculate 1D Space Norm (modulo 1 to account for circularity)
-    transformed_norm_a = torch.triu(
+    transformed_norm = circle_norm(outputs, target_outputs)
-        torch.norm((outputs[:, None] - target_outputs[None, :]) % 1, dim=-1)
-    )
-    transformed_norm_b = torch.triu(
-        torch.norm((1 + outputs[:, None] - target_outputs[None, :]) % 1, dim=-1)
-    )
-    transformed_norm = torch.minimum(transformed_norm_a, transformed_norm_b)
 
     diff = torch.pow(rgb_norm - transformed_norm, 2)
+    N = len(outputs)
+    return torch.sum(diff) / (N * (N - 1)) / 2
 
-    return torch.mean(diff)
-
-
-def preservation_loss(inputs, outputs, target_inputs=None, target_outputs=None):
-    # Distance Preservation Component (or scaled euclidean if given targets)
-    # Encourages the model to keep relative distances from the RGB space in the transformed space
-    if target_inputs is None:
-        target_inputs = inputs
-    else:
-        assert target_outputs is not None
-    if target_outputs is None:
-        target_outputs = outputs
-
-    # Calculate RGB Norm
-    max_rgb_distance = torch.sqrt(torch.tensor(2 + 1))  # scale to [0, 1]
-    # max_rgb_distance = 1
-    rgb_norm = (
-        torch.triu(torch.norm(inputs[:, None, :] - target_inputs[None, :, :], dim=-1))
-        / max_rgb_distance
-    )
-    # connect (0, 0, 0) and (1, 1, 1): max_rgb_distance in the RGB space
-    rgb_norm = rgb_norm % 1
-    # print(rgb_norm)
-
-    # Calculate 1D Space Norm (modulo 1 to account for circularity)
-    transformed_norm = torch.triu(
-        torch.norm((outputs[:, None] - target_outputs[None, :]) % 1, dim=-1)
-    )
-
-    diff = torch.abs(rgb_norm - transformed_norm)
-    # print(diff)
 
-    return torch.mean(diff)
+def circle_norm(vector, other_vector):
+    # Assumes vectors are of shape (N,1)
+    loss_a = torch.triu(torch.abs((vector - other_vector.T)))
+    loss_b = torch.triu(1 - torch.abs((vector - other_vector.T)))
+    loss = torch.minimum(loss_a, loss_b)
+    return loss
 
 
 def separation_loss(red, green, blue):

model.py

@@ -3,12 +3,9 @@ import torch
 import torch.nn as nn
 from torch.optim.lr_scheduler import ReduceLROnPlateau
 
-from losses import (  # noqa: F401
+from losses import preservation_loss
-    calculate_separation_loss,
+
-    preservation_loss,
+# from utils import PURE_RGB
-    simple_preservation_loss,
-)
-from utils import PURE_RGB
 
 
 class ColorTransformerModel(L.LightningModule):
@@ -52,24 +49,22 @@ class ColorTransformerModel(L.LightningModule):
     def training_step(self, batch, batch_idx):
         inputs, labels = batch  # x are the RGB inputs, labels are the strings
         outputs = self.forward(inputs)
-        rgb_tensor = PURE_RGB.to(self.device)
+        # rgb_tensor = PURE_RGB.to(self.device)
-        p_loss = simple_preservation_loss(
+        p_loss = preservation_loss(
             inputs,
             outputs,
-            target_inputs=rgb_tensor,
+            # target_inputs=rgb_tensor,
-            target_outputs=self.forward(rgb_tensor),
+            # target_outputs=self.forward(rgb_tensor),
         )
         alpha = self.hparams.alpha
-        # loss = p_loss
 
-        # distance = torch.minimum(
+        # N = len(outputs)
-        #     torch.norm(outputs - labels), torch.norm(1 + outputs - labels)
+        # distance = circle_norm(outputs, labels) / (N*(N-1)/2)
-        # ).mean()
         distance = torch.norm(outputs - labels).mean()
 
         # Backprop with this:
         loss = (1 - alpha) * p_loss + alpha * distance
-        # p_loss is unsupervised
+        # p_loss is unsupervised (preserve relative distances - either in-batch or to-target)
         # distance is supervised.
         self.log("hp_metric", loss)